{"id":7385,"date":"2026-10-06T11:28:53","date_gmt":"2026-10-06T03:28:53","guid":{"rendered":"https:\/\/nullthought.net\/?p=7385"},"modified":"2026-10-06T11:28:56","modified_gmt":"2026-10-06T03:28:56","slug":"%e5%85%89%e5%ad%90ai%e8%8a%af%e7%89%87%ef%bc%9a%e4%bb%8e%e5%85%89%e4%ba%92%e8%bf%9e%e3%80%81%e5%85%89%e8%ae%a1%e7%ae%97%e5%88%b0%e5%8e%9f%e4%bd%8d%e5%ad%a6%e4%b9%a0%ef%bc%8c%e4%b8%80","status":"publish","type":"post","link":"https:\/\/nullthought.net\/?p=7385","title":{"rendered":"\u5149\u5b50AI\u82af\u7247\uff1a\u4ece\u5149\u4e92\u8fde\u3001\u5149\u8ba1\u7b97\u5230\u201c\u539f\u4f4d\u5b66\u4e60\u201d\uff0c\u4e00\u573a\u6b63\u5728\u53d1\u751f\u7684AI\u8ba1\u7b97\u67b6\u6784\u53d8\u9769"},"content":{"rendered":"\n

\u622a\u81f3 2026 \u5e74 10 \u6708 6 \u65e5<\/strong>\uff0c\u5149\u5b50 AI \u82af\u7247\u6b63\u5728\u7ecf\u5386\u4e00\u4e2a\u975e\u5e38\u5173\u952e\u7684\u9636\u6bb5\u53d8\u5316\u3002<\/p>\n\n\n\n

\u8fc7\u53bb\uff0c\u4eba\u4eec\u8ba8\u8bba\u5149\u8ba1\u7b97\uff0c\u5f80\u5f80\u96c6\u4e2d\u5728\u4e00\u4e2a\u95ee\u9898\u4e0a\uff1a\u5149\u80fd\u4e0d\u80fd\u6bd4\u7535\u5b50\u66f4\u5feb\u3001\u66f4\u7701\u7535\u5730\u5b8c\u6210\u77e9\u9635\u4e58\u6cd5\uff1f<\/strong><\/p>\n\n\n\n

\u73b0\u5728\uff0c\u95ee\u9898\u5df2\u7ecf\u660e\u663e\u6539\u53d8\u3002<\/p>\n\n\n\n

2025\u20142026 \u5e74\uff0c\u4e00\u7cfb\u5217\u79d1\u7814\u6210\u679c\u5f00\u59cb\u8bc1\u660e\uff0c\u5149\u5b50\u82af\u7247\u4e0d\u4ec5\u80fd\u591f\u8fdb\u884c\u9ad8\u901f\u77e9\u9635\u8ba1\u7b97\uff0c\u8fd8\u9010\u6b65\u80fd\u591f\u8fd0\u884c ResNet\u3001BERT\u3001LLM\u3001\u6269\u6563\u6a21\u578b\u7b49\u771f\u6b63\u7684 AI \u5de5\u4f5c\u8d1f\u8f7d\uff1b\u80fd\u591f\u5728\u82af\u7247\u4e0a\u8fdb\u884c\u53cd\u5411\u4f20\u64ad\u8bad\u7ec3\uff1b\u80fd\u591f\u76f4\u63a5\u4f9d\u636e\u771f\u5b9e\u5149\u5b50\u5668\u4ef6\u7684\u7269\u7406\u72b6\u6001\u8fdb\u884c\u201c\u539f\u4f4d\u5b66\u4e60\u201d\uff1b\u751a\u81f3\u5f00\u59cb\u62e5\u6709\u77ed\u671f\u548c\u957f\u671f\u8bb0\u5fc6\u3002<\/p>\n\n\n\n

\u4e0e\u6b64\u540c\u65f6\uff0c\u4ea7\u4e1a\u754c\u7684\u53d1\u5c55\u65b9\u5411\u66f4\u52a0\u660e\u786e\uff1a\u9996\u5148\u5927\u89c4\u6a21\u5546\u4e1a\u5316\u7684\u5f88\u53ef\u80fd\u4e0d\u662f\u5b8c\u5168\u53d6\u4ee3 GPU \u7684\u201c\u5149\u5b50 GPU\u201d\uff0c\u800c\u662f\u5149\u4e92\u8fde\u3001Optical I\/O\u3001CPO\uff0c\u4ee5\u53ca\u7535\u5b50\u4e0e\u5149\u5b50\u7ec4\u6210\u7684\u5f02\u6784 AI \u7cfb\u7edf\u3002<\/strong><\/p>\n\n\n\n

2026 \u5e74 9 \u6708\u300aNature Photonics\u300b\u53d1\u8868\u7684\u6700\u65b0 Perspective \u4e5f\u6307\u51fa\uff0c\u5149\u5b50 AI \u52a0\u901f\u5668\u6b63\u5728\u80fd\u6548\u3001\u541e\u5410\u91cf\u3001\u8ba1\u7b97\u5bc6\u5ea6\u548c\u5ef6\u8fdf\u7b49\u6307\u6807\u4e0a\u4e0d\u65ad\u7f29\u5c0f\u4e0e\u5148\u8fdb\u6570\u5b57 AI \u82af\u7247\u7684\u5dee\u8ddd\uff0c\u800c\u4e91\u7aef\u901a\u7528\u5149\u5b50\u52a0\u901f\u5668\u548c\u8fb9\u7f18\u4e13\u7528\u5149\u8ba1\u7b97\u7cfb\u7edf\u53ef\u80fd\u5f62\u6210\u4e0d\u540c\u7684\u53d1\u5c55\u8def\u7ebf\u3002<\/p>\n\n\n\n

\"\"
\u5149\u5b50\u4eba\u5de5\u667a\u80fd\u82af\u7247\u5168\u666f\u56fe<\/figcaption><\/figure>\n\n\n\n

\u4e00\u3001\u9996\u5148\u8981\u5206\u6e05\uff1a\u5149\u5b50 AI \u82af\u7247\u5176\u5b9e\u6709\u4e09\u6761\u8def\u7ebf<\/h2>\n\n\n\n

\u4eca\u5929\u6240\u8c13\u201c\u5149\u5b50 AI\u201d\uff0c\u5b9e\u9645\u4e0a\u81f3\u5c11\u5305\u542b\u4e09\u4e2a\u4e0d\u540c\u5c42\u6b21\u3002<\/p>\n\n\n\n

\u7b2c\u4e00\u7c7b\u662f Photonic Computing\u2014\u2014\u76f4\u63a5\u7528\u5149\u8ba1\u7b97<\/strong>\u3002<\/p>\n\n\n\n

\u795e\u7ecf\u7f51\u7edc\u6700\u6838\u5fc3\u7684\u5de5\u4f5c\u4e4b\u4e00\u662f\u77e9\u9635\u2014\u5411\u91cf\u4e58\u6cd5\u3002\u4f20\u7edf GPU \u9700\u8981\u5927\u91cf\u6676\u4f53\u7ba1\u4e0d\u65ad\u6267\u884c\u4e58\u6cd5\u548c\u52a0\u6cd5\uff0c\u800c\u5149\u5b50\u82af\u7247\u53ef\u4ee5\u5229\u7528\u5e72\u6d89\u3001\u884d\u5c04\u3001\u76f8\u4f4d\u3001\u632f\u5e45\u3001\u6ce2\u957f\u7b49\u7269\u7406\u5c5e\u6027\uff0c\u5728\u5149\u4f20\u64ad\u8fc7\u7a0b\u4e2d\u76f4\u63a5\u5b8c\u6210\u77e9\u9635\u53d8\u6362\u3002<\/p>\n\n\n\n

Mach-Zehnder Interferometer\u3001\u5fae\u73af\u8c10\u632f\u5668\u3001\u884d\u5c04\u7ed3\u6784\u7b49\uff0c\u90fd\u53ef\u4ee5\u6210\u4e3a\u201c\u5149\u5b66\u4e58\u6cd5\u5668\u201d\u7684\u7ec4\u6210\u90e8\u5206\u3002<\/p>\n\n\n\n

\u7b2c\u4e8c\u7c7b\u662f \u5149\u7535\u6df7\u5408\u8ba1\u7b97<\/strong>\u3002<\/p>\n\n\n\n

\u8fd9\u662f\u76ee\u524d\u66f4\u73b0\u5b9e\u7684\u8def\u7ebf\uff1a<\/p>\n\n\n\n

\u5149\u8d1f\u8d23\u5927\u89c4\u6a21\u7ebf\u6027\u8fd0\u7b97\uff0c\u7535\u5b50\u8d1f\u8d23\u5b58\u50a8\u3001\u63a7\u5236\u3001\u975e\u7ebf\u6027\u8fd0\u7b97\u3001\u6570\u5b57\u63a5\u53e3\u4ee5\u53ca\u7cbe\u786e\u8ba1\u7b97\u3002<\/strong><\/p>\n\n\n\n

Lightmatter\u3001Lightelligence\u3001Lumai\u3001Q.ANT \u7b49\u516c\u53f8\u7684\u4ea7\u54c1\uff0c\u672c\u8d28\u4e0a\u90fd\u5c5e\u4e8e\u4e0d\u540c\u5f62\u5f0f\u7684\u5149\u7535\u5f02\u6784\u8ba1\u7b97\uff0c\u800c\u4e0d\u662f\u5b8c\u5168\u629b\u5f03 CMOS\u3002<\/p>\n\n\n\n

\u7b2c\u4e09\u7c7b\u662f Photonic Interconnect \/ Optical I\/O\u2014\u2014\u5149\u4e92\u8fde<\/strong>\u3002<\/p>\n\n\n\n

\u5b83\u5e76\u4e0d\u8d1f\u8d23\u8fd0\u884c\u795e\u7ecf\u7f51\u7edc\u672c\u8eab\uff0c\u800c\u662f\u89e3\u51b3\uff1a<\/p>\n\n\n\n

GPU \u2194 GPU\u3001GPU \u2194 HBM\/DRAM\u3001chiplet \u2194 chiplet\u3001\u670d\u52a1\u5668 \u2194 \u670d\u52a1\u5668\u4e4b\u95f4\u7684\u6570\u636e\u642c\u8fd0\u3002<\/strong><\/p>\n\n\n\n

\u6070\u6070\u662f\u8fd9\u4e00\u65b9\u5411\uff0c\u76ee\u524d\u8ddd\u79bb\u5927\u89c4\u6a21\u4ea7\u4e1a\u5316\u6700\u8fd1\u3002<\/p>\n\n\n\n

\u56e0\u6b64\uff0c\u73b0\u9636\u6bb5\u53ef\u4ee5\u7b80\u5355\u6982\u62ec\u4e3a\uff1a<\/p>\n\n\n\n

\n

\u77ed\u671f\u770b\u5149\u4e92\u8fde\uff0c\u4e2d\u671f\u770b\u5149\u7535\u6df7\u5408\u8ba1\u7b97\uff0c\u957f\u671f\u624d\u770b\u66f4\u5b8c\u6574\u7684\u5149\u5b50\u8ba1\u7b97\u4f53\u7cfb\u3002<\/strong><\/p>\n<\/blockquote>\n\n\n\n

\n\n\n\n

\u4e8c\u3001\u4e3a\u4ec0\u4e48 AI \u6070\u597d\u628a\u5149\u5b50\u6280\u672f\u63a8\u5230\u4e86\u4e34\u754c\u70b9\uff1f<\/h1>\n\n\n\n

AI \u7cfb\u7edf\u4eca\u5929\u9762\u5bf9\u7684\u5df2\u7ecf\u4e0d\u4ec5\u662f\u201c\u7b97\u529b\u4e0d\u591f\u201d\u3002<\/p>\n\n\n\n

\u5b9e\u9645\u4e0a\u540c\u65f6\u5b58\u5728\u4e09\u5835\u5899\uff1a<\/p>\n\n\n\n

\u8ba1\u7b97\u529f\u8017\u5899\u3001\u5185\u5b58\u5899\u548c I\/O \u5899\u3002<\/strong><\/p>\n\n\n\n

\u5c24\u5176\u662f\u5927\u6a21\u578b\u3002<\/p>\n\n\n\n

GPU \u5185\u90e8 Tensor Core \u7684\u77e9\u9635\u8ba1\u7b97\u5df2\u7ecf\u975e\u5e38\u5feb\uff0c\u4f46\u6a21\u578b\u53c2\u6570\u3001\u6fc0\u6d3b\u503c\u3001KV Cache\u3001\u68af\u5ea6\u7b49\u6570\u636e\u9700\u8981\u4e0d\u65ad\u5728 HBM\u3001GPU \u548c\u5176\u4ed6\u8ba1\u7b97\u8282\u70b9\u4e4b\u95f4\u79fb\u52a8\u3002<\/p>\n\n\n\n

\u8d8a\u6765\u8d8a\u591a\u7684\u529f\u8017\uff0c\u4e0d\u662f\u6d88\u8017\u5728\u201c\u4e58\u6cd5\u201d\u672c\u8eab\uff0c\u800c\u662f\u5728\uff1a<\/p>\n\n\n\n

\u642c\u6570\u636e\u3002<\/strong><\/p>\n\n\n\n

\u5149\u5728\u8fd9\u91cc\u6709\u51e0\u4e2a\u5929\u7136\u4f18\u52bf\u3002<\/p>\n\n\n\n

\u9996\u5148\u662f\u5927\u89c4\u6a21\u5e76\u884c<\/strong>\u3002<\/p>\n\n\n\n

\u4e00\u6839\u5149\u7ea4\u6216\u6ce2\u5bfc\u53ef\u4ee5\u901a\u8fc7 WDM \u540c\u65f6\u4f20\u8f93\u591a\u4e2a\u4e0d\u540c\u6ce2\u957f\u3002\u7a7a\u95f4\u3001\u504f\u632f\u751a\u81f3\u76f8\u4f4d\uff0c\u4e5f\u53ef\u4ee5\u5f62\u6210\u989d\u5916\u7684\u5e76\u884c\u8ba1\u7b97\u7ef4\u5ea6\u3002<\/p>\n\n\n\n

\u5176\u6b21\u662f\u9ad8\u901f\u4e0e\u8ddd\u79bb\u4f18\u52bf<\/strong>\u3002<\/p>\n\n\n\n

\u9ad8\u901f\u7535\u4e92\u8fde\u8ddd\u79bb\u589e\u52a0\u540e\uff0c\u635f\u8017\u3001\u5747\u8861\u548c\u529f\u8017\u8fc5\u901f\u6076\u5316\uff0c\u800c\u5149\u7ea4\u66f4\u52a0\u9002\u5408\u673a\u67b6\u95f4\u4e43\u81f3\u66f4\u957f\u8ddd\u79bb\u7684\u9ad8\u901f\u6570\u636e\u4ea4\u6362\u3002<\/p>\n\n\n\n

\u7b2c\u4e09\u662f\u5149\u5b66\u4e0e\u7ebf\u6027\u4ee3\u6570\u5929\u7136\u5339\u914d<\/strong>\u3002<\/p>\n\n\n\n

\u5149\u7684\u5e72\u6d89\u548c\u884d\u5c04\u8fc7\u7a0b\u672c\u8eab\u5c31\u662f\u6570\u5b66\u610f\u4e49\u4e0a\u7684\u7ebf\u6027\u53d8\u6362\uff0c\u56e0\u6b64\u67d0\u4e9b\u77e9\u9635\u8ba1\u7b97\u4e0d\u4e00\u5b9a\u9700\u8981\u901a\u8fc7\u5927\u91cf\u6570\u5b57\u4e58\u6cd5\u5668\u9010\u9879\u6267\u884c\u3002<\/p>\n\n\n\n

\u4f46\u5149\u4e5f\u5b58\u5728\u660e\u663e\u5f31\u70b9\uff1a<\/p>\n\n\n\n

\u96be\u5b58\u50a8\u3001\u96be\u5b9e\u73b0\u7cbe\u786e\u6570\u5b57\u903b\u8f91\u3001\u975e\u7ebf\u6027\u8fd0\u7b97\u56f0\u96be\u3002<\/strong><\/p>\n\n\n\n

\u6240\u4ee5\u672a\u6765\u771f\u6b63\u6709\u7ade\u4e89\u529b\u7684\u67b6\u6784\uff0c\u5f88\u53ef\u80fd\u4e0d\u662f\u201c\u5168\u5149 GPU\u201d\uff0c\u800c\u662f\uff1a<\/p>\n\n\n\n

\n

\u7535\u5b50\u8d1f\u8d23\u8bb0\u5fc6\u3001\u63a7\u5236\u548c\u7cbe\u786e\u6570\u5b57\u903b\u8f91\uff0c\u5149\u8d1f\u8d23\u5927\u89c4\u6a21\u77e9\u9635\u8fd0\u7b97\u548c\u9ad8\u901f\u6570\u636e\u79fb\u52a8\u3002<\/strong><\/p>\n<\/blockquote>\n\n\n\n

\n\n\n\n

\u4e09\u3001\u79d1\u7814\u91cd\u5fc3\u5df2\u7ecf\u53d8\u4e86\uff1a\u4e0d\u518d\u53ea\u662f\u8bc1\u660e\u201c\u5149\u80fd\u7b97\u201d<\/h1>\n\n\n\n

\u8fc7\u53bb\u5149\u5b50\u795e\u7ecf\u7f51\u7edc\u7814\u7a76\u7684\u4e00\u4e2a\u5178\u578b\u95ee\u9898\u662f\uff1a<\/p>\n\n\n\n

\u5f88\u591a\u8bba\u6587\u6700\u7ec8\u8fd8\u662f\u505c\u7559\u5728 MNIST \u4e4b\u7c7b\u7684\u5c0f\u578b\u5b9e\u9a8c\u4efb\u52a1\u4e0a\u3002<\/p>\n\n\n\n

\u6240\u4ee5\u5373\u4f7f\u67d0\u4e2a\u5b9e\u9a8c\u7cfb\u7edf\u80fd\u591f\u8fbe\u5230\u6781\u9ad8\u7684 TOPS\/W\uff0c\u4e5f\u5f88\u96be\u8bc1\u660e\u5b83\u80fd\u591f\u8fdb\u5165\u771f\u5b9e AI \u5de5\u4f5c\u8d1f\u8f7d\u3002<\/p>\n\n\n\n

\u4f46\u662f\u4ece 2023 \u5e74\u5f00\u59cb\uff0c\u8fd9\u79cd\u60c5\u51b5\u53d1\u751f\u4e86\u660e\u663e\u53d8\u5316\u3002<\/p>\n\n\n\n

\u65f6\u95f4<\/th>\u4ee3\u8868\u6210\u679c<\/th>\u4e3b\u8981\u610f\u4e49<\/th><\/tr>
2023<\/td>ACCEL<\/td>\u5149\u7535\u5168\u6a21\u62df\u89c6\u89c9\u8ba1\u7b97\uff0c\u5c1d\u8bd5\u7ed5\u5f00\u5927\u89c4\u6a21 ADC<\/td><\/tr>
2024<\/td>Taichi<\/td>\u5149\u5b50 chiplet \u6269\u5c55\u5230\u5343\u7c7b\u522b\u8bc6\u522b\u548c AIGC<\/td><\/tr>
2025<\/td>Lightmatter<\/td>\u8fd0\u884c ResNet\u3001BERT\u3001\u6df1\u5ea6\u5f3a\u5316\u5b66\u4e60<\/td><\/tr>
2026<\/td>Nokia Bell Labs<\/td>\u82af\u7247\u5185\u90e8\u5b9e\u73b0\u53cd\u5411\u4f20\u64ad\u8bad\u7ec3<\/td><\/tr>
2026<\/td>Photonic MoE<\/td>\u628a MoE \u7684\u201c\u6269\u5bbd\u201d\u601d\u60f3\u5f15\u5165\u5149\u5b50\u795e\u7ecf\u7f51\u7edc<\/td><\/tr>
2026<\/td>3D PNN<\/td>\u7528\u4e09\u7ef4\u5149\u5b50\u7ed3\u6784\u7a81\u7834\u4e8c\u7ef4\u5e03\u7ebf\u548c\u5bc6\u5ea6\u9650\u5236<\/td><\/tr>
2026<\/td>FLARE<\/td>\u628a\u8ba1\u7b97\u3001\u611f\u77e5\u4ee5\u53ca\u957f\u77ed\u671f\u8bb0\u5fc6\u7ed3\u5408\u8d77\u6765<\/td><\/tr>
2026.10<\/td>INSPIRE<\/td>\u8ba9\u771f\u5b9e\u5149\u5b50\u82af\u7247\u76f4\u63a5\u6d4b\u91cf\u68af\u5ea6\u5e76\u8fdb\u884c\u539f\u4f4d\u5b66\u4e60<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

ACCEL \u5728 2023 \u5e74\u5c55\u793a\u4e86 4.6 POPS \u8ba1\u7b97\u901f\u5ea6\u548c 74.8 POPS\/W \u7cfb\u7edf\u80fd\u6548<\/strong>\uff0c\u5176\u4e2d\u8d85\u8fc7 99% \u7684\u8fd0\u7b97\u7531\u5149\u5b66\u5b8c\u6210\uff0c\u5e76\u901a\u8fc7\u5149\u7535\u5168\u6a21\u62df\u67b6\u6784\u51cf\u5c11\u5927\u89c4\u6a21 ADC \u5e26\u6765\u7684\u5f00\u9500\u3002<\/p>\n\n\n\n

2024 \u5e74\u7684 Taichi \u5219\u8fdb\u4e00\u6b65\u5229\u7528\u5149\u5b50 chiplet \u548c\u884d\u5c04\u2014\u5e72\u6d89\u6df7\u5408\u67b6\u6784\uff0c\u5c06\u89c4\u6a21\u63a8\u8fdb\u5230\u767e\u4e07\u7ea7\u795e\u7ecf\u5143\u80fd\u529b\uff0c\u5e76\u62a5\u544a 160 TOPS\/W<\/strong> \u80fd\u6548\uff0c\u5728\u5b9e\u9a8c\u4e2d\u5b9e\u73b0\u5343\u7c7b\u522b\u7ea7\u8bc6\u522b\u548c AI \u751f\u6210\u5185\u5bb9\u4efb\u52a1\u3002<\/p>\n\n\n\n

\u8fd9\u610f\u5473\u7740\u95ee\u9898\u5df2\u7ecf\u4ece\uff1a<\/p>\n\n\n\n

\u201c\u5149\u80fd\u4e0d\u80fd\u505a\u795e\u7ecf\u7f51\u7edc\uff1f\u201d<\/strong><\/p>\n\n\n\n

\u9010\u6e10\u53d8\u6210\uff1a<\/p>\n\n\n\n

\u201c\u5149\u5b50\u82af\u7247\u80fd\u4e0d\u80fd\u8fd0\u884c\u771f\u5b9e AI\u3001\u80fd\u4e0d\u80fd\u8bad\u7ec3\u3001\u80fd\u4e0d\u80fd\u9002\u5e94\u5236\u9020\u8bef\u5dee\u3001\u80fd\u4e0d\u80fd\u62e5\u6709\u72b6\u6001\u548c\u8bb0\u5fc6\u3001\u6700\u540e\u80fd\u4e0d\u80fd\u89c4\u6a21\u5316\uff1f\u201d<\/strong><\/p>\n\n\n\n

\n\n\n\n

\u56db\u3001Lightmatter\uff1a\u4ece\u201c\u5149\u80fd\u7b97\u201d\u63a8\u8fdb\u5230\u201c\u771f\u6b63 AI \u6a21\u578b\u80fd\u8fd0\u884c\u201d<\/h1>\n\n\n\n

2025 \u5e74\u53d1\u8868\u4e8e\u300aNature\u300b\u7684 Universal photonic artificial intelligence acceleration<\/strong> \u662f\u4e00\u4e2a\u91cd\u8981\u8282\u70b9\u3002<\/p>\n\n\n\n

Lightmatter \u56e2\u961f\u5c55\u793a\u4e86\u4e00\u5957\u5149\u5b50 AI \u5904\u7406\u5668\uff0c\u4e0d\u518d\u53ea\u8fd0\u884c\u7b80\u5355\u5206\u7c7b\u7f51\u7edc\uff0c\u800c\u662f\u8fdb\u4e00\u6b65\u8fd0\u884c\u4e86 ResNet\u3001BERT \u4ee5\u53ca Atari \u6df1\u5ea6\u5f3a\u5316\u5b66\u4e60\u6a21\u578b<\/strong>\u3002<\/p>\n\n\n\n

\u8fd9\u9879\u5de5\u4f5c\u7684\u610f\u4e49\u5e76\u4e0d\u53ea\u662f\u901f\u5ea6\u3002<\/p>\n\n\n\n

\u66f4\u91cd\u8981\u7684\u662f\uff1a<\/p>\n\n\n\n

\u7cbe\u5ea6\u3002<\/strong><\/p>\n\n\n\n

\u5149\u8ba1\u7b97\u5c5e\u4e8e\u6a21\u62df\u8ba1\u7b97\uff0c\u56e0\u6b64\u4f1a\u53d7\u5230\uff1a<\/p>\n\n\n\n

\u76f8\u4f4d\u8bef\u5dee\u3001\u6fc0\u5149\u566a\u58f0\u3001\u70ed\u6f02\u79fb\u3001\u5236\u9020\u8bef\u5dee\u4ee5\u53ca\u5668\u4ef6\u975e\u7406\u60f3\u6027\u7b49\u5f71\u54cd\u3002<\/p>\n\n\n\n

\u6b64\u524d\u5f88\u591a\u5149\u5b66\u8ba1\u7b97\u5668\u7406\u8bba\u6027\u80fd\u5f88\u9ad8\uff0c\u4f46\u4e00\u65e6\u8fd0\u884c\u5927\u578b\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\uff0c\u8bef\u5dee\u53ef\u80fd\u9010\u5c42\u7d2f\u79ef\u3002<\/p>\n\n\n\n

\u56e0\u6b64\uff0c\u8fd9\u9879\u5de5\u4f5c\u7684\u771f\u6b63\u610f\u4e49\u662f\uff1a<\/p>\n\n\n\n

\n

\u5149\u5b50\u5904\u7406\u5668\u5f00\u59cb\u4ece\u201c\u8bc1\u660e\u53ef\u4ee5\u505a\u77e9\u9635\u4e58\u6cd5\u201d\uff0c\u8fdb\u5165\u201c\u80fd\u591f\u6267\u884c\u73b0\u5b9e AI \u6a21\u578b\u201d\u7684\u9636\u6bb5\u3002<\/p>\n<\/blockquote>\n\n\n\n

\u8fd9\u4e5f\u662f\u5149\u8ba1\u7b97\u8fc8\u5411\u5b9e\u7528\u5316\u5fc5\u987b\u8de8\u8fc7\u7684\u4e00\u9053\u95e8\u69db\u3002<\/p>\n\n\n\n

\n\n\n\n

\u4e94\u30012026 \u5e74\u7b2c\u4e00\u6b21\u91cd\u8981\u8de8\u8d8a\uff1a\u8ba9\u5149\u5b50\u82af\u7247\u81ea\u5df1\u5b8c\u6210\u53cd\u5411\u4f20\u64ad<\/h1>\n\n\n\n

\u4e0b\u4e00\u9053\u969c\u788d\u662f\uff1a<\/p>\n\n\n\n

\u8bad\u7ec3\u3002<\/strong><\/p>\n\n\n\n

\u4f20\u7edf\u5149\u5b50\u795e\u7ecf\u7f51\u7edc\u5f80\u5f80\u91c7\u7528\u8fd9\u6837\u7684\u6d41\u7a0b\uff1a<\/p>\n\n\n\n

\u8bad\u7ec3\u5728 GPU \u4e0a\u5b8c\u6210 \u2192 \u5f97\u5230\u6743\u91cd \u2192 \u628a\u6743\u91cd\u914d\u7f6e\u5230\u5149\u5b50\u82af\u7247 \u2192 \u5149\u5b50\u82af\u7247\u53ea\u8d1f\u8d23\u63a8\u7406\u3002<\/p>\n\n\n\n

\u95ee\u9898\u5728\u4e8e\uff0c\u8f6f\u4ef6\u4e2d\u7684\u201c\u7406\u60f3\u5149\u5b50\u82af\u7247\u201d\u548c\u771f\u6b63\u5236\u9020\u51fa\u6765\u7684\u82af\u7247\u5e76\u4e0d\u5b8c\u5168\u76f8\u540c\u3002<\/p>\n\n\n\n

2026 \u5e74 3 \u6708\uff0cNokia Bell Labs \u5728\u300aNature\u300b\u53d1\u8868 Integrated photonic neural network with on-chip backpropagation training<\/strong>\u3002<\/p>\n\n\n\n

\u7814\u7a76\u4eba\u5458\u628a\u68af\u5ea6\u53cd\u5411\u4f20\u64ad\u771f\u6b63\u63a8\u8fdb\u5230\u96c6\u6210\u5149\u5b50\u795e\u7ecf\u7f51\u7edc\u5185\u90e8\uff0c\u5b9e\u73b0\u4e86\u5305\u62ec\u7ebf\u6027\u8ba1\u7b97\u3001\u975e\u7ebf\u6027\u6fc0\u6d3b\u4ee5\u53ca\u68af\u5ea6\u76f8\u5173\u8fc7\u7a0b\u5728\u5185\u7684\u7247\u4e0a\u8bad\u7ec3\u3002<\/p>\n\n\n\n

\u8fd9\u662f\u4e00\u4e2a\u91cd\u8981\u53d8\u5316\uff1a<\/p>\n\n\n\n

\u8fc7\u53bb\u662f\uff1a<\/p>\n\n\n\n

GPU \u8bad\u7ec3 \u2192 \u5149\u5b50\u82af\u7247\u6267\u884c\u3002<\/strong><\/p>\n\n\n\n

\u73b0\u5728\u5f00\u59cb\u53d8\u6210\uff1a<\/p>\n\n\n\n

\u5149\u5b50\u82af\u7247\u81ea\u5df1\u53c2\u4e0e\u8bad\u7ec3\u3002<\/strong><\/p>\n\n\n\n

\u6362\u53e5\u8bdd\u8bf4\uff0c\u5149\u5b50 AI \u82af\u7247\u5f00\u59cb\u4ece\u5355\u7eaf\u7684\u201c\u9ad8\u901f\u63a8\u7406\u534f\u5904\u7406\u5668\u201d\uff0c\u671d\u7740\u771f\u6b63\u7684\u53ef\u8bad\u7ec3\u8ba1\u7b97\u4ecb\u8d28<\/strong>\u6f14\u8fdb\u3002<\/p>\n\n\n\n

\n\n\n\n

\u516d\u30012026 \u5e74 10 \u6708\u7684\u6700\u65b0\u7a81\u7834\uff1a\u771f\u5b9e\u82af\u7247\u76f4\u63a5\u201c\u539f\u4f4d\u5b66\u4e60\u201d<\/h1>\n\n\n\n

\u4f46\u7247\u4e0a\u8bad\u7ec3\u4e4b\u540e\uff0c\u8fd8\u5b58\u5728\u4e00\u4e2a\u66f4\u6df1\u7684\u95ee\u9898\u3002<\/p>\n\n\n\n

\u5373\u4f7f\u8bad\u7ec3\u7b97\u6cd5\u8fd0\u884c\u5728\u5149\u5b50\u7cfb\u7edf\u9644\u8fd1\uff0c\u4eba\u4eec\u5f80\u5f80\u4ecd\u7136\u9700\u8981\u4e00\u4e2a\u76f8\u5bf9\u51c6\u786e\u7684\u6570\u5b57\u7269\u7406\u6a21\u578b<\/strong>\uff1a<\/p>\n\n\n\n

\u8f6f\u4ef6\u5fc5\u987b\u77e5\u9053\u6bcf\u4e2a\u5149\u5b66\u5668\u4ef6\u5927\u7ea6\u662f\u4ec0\u4e48\u72b6\u6001\uff0c\u7136\u540e\u8ba1\u7b97\u68af\u5ea6\u3002<\/p>\n\n\n\n

\u53ef\u662f\u73b0\u5b9e\u82af\u7247\u5b58\u5728\u5927\u91cf\u96be\u4ee5\u5b8c\u5168\u5efa\u6a21\u7684\u56e0\u7d20\uff1a<\/p>\n\n\n\n

\u5236\u9020\u8bef\u5dee\u3001\u70ed\u6f02\u79fb\u3001\u5668\u4ef6\u8001\u5316\u3001\u8026\u5408\u8bef\u5dee\u3001\u5149\u5b66\u6563\u5c04\u7b49\u3002<\/p>\n\n\n\n

\u56e0\u6b64\u771f\u6b63\u7406\u60f3\u7684\u65b9\u5f0f\u662f\uff1a<\/p>\n\n\n\n

\n

\u4e0d\u8981\u8bd5\u56fe\u5b8c\u5168\u6a21\u62df\u771f\u5b9e\u82af\u7247\uff0c\u800c\u662f\u76f4\u63a5\u95ee\u771f\u5b9e\u82af\u7247\uff1a\u201c\u4f60\u7684\u68af\u5ea6\u662f\u591a\u5c11\uff1f\u201d<\/strong><\/p>\n<\/blockquote>\n\n\n\n

2026 \u5e74 10 \u6708 5 \u65e5\uff0c\u300aNature Computational Science\u300b\u53d1\u8868\uff1a<\/p>\n\n\n\n

Photonic neuromorphic learning via generalized in situ physical gradient descent<\/strong><\/p>\n\n\n\n

\u6e05\u534e\u5927\u5b66\u7b49\u56e2\u961f\u63d0\u51fa\u4e86 INSPIRE\u2014\u2014in situ physical gradient descent\uff0c\u5373\u539f\u4f4d\u7269\u7406\u68af\u5ea6\u4e0b\u964d\u3002<\/strong><\/p>\n\n\n\n

\u5176\u5173\u952e\u601d\u60f3\u662f\u5229\u7528\u7247\u4e0a\u5408\u6210\u65f6\u95f4\u53cd\u6f14\u5168\u606f<\/strong>\uff0c\u76f4\u63a5\u6d4b\u91cf\u5149\u5b50\u7cfb\u7edf\u4e2d\u53cc\u5411\u4f20\u64ad\u6a21\u5f0f\u7684\u5b8c\u6574\u590d\u5149\u573a\u3002<\/p>\n\n\n\n

\u6709\u4e86\u771f\u5b9e\u7269\u7406\u7cfb\u7edf\u5185\u90e8\u7684\u5149\u573a\u4fe1\u606f\uff0c\u5c31\u53ef\u4ee5\u76f4\u63a5\u8ba1\u7b97\uff1a<\/p>\n\n\n\n

\u53c2\u6570\u5e94\u8be5\u5f80\u54ea\u4e2a\u65b9\u5411\u8c03\u6574\u3002<\/strong><\/p>\n\n\n\n

\u4e5f\u5c31\u662f\u8bf4\uff0c\u68af\u5ea6\u4e0d\u518d\u4e3b\u8981\u6765\u81ea\u4e00\u4e2a\u6a21\u62df\u6a21\u578b\uff0c\u800c\u662f\u76f4\u63a5\u6765\u81ea\u771f\u5b9e\u82af\u7247\u672c\u8eab<\/strong>\u3002<\/p>\n\n\n\n

\u5b9e\u9a8c\u8bad\u7ec3\u51fa\u7684\u77e9\u9635\u76f8\u5bf9\u8bef\u5dee\u8fbe\u5230\u7ea6 0.26%<\/strong>\u3002\u66f4\u91cd\u8981\u7684\u662f\uff0c\u8be5\u65b9\u6cd5\u5177\u6709\u8f83\u5f3a\u7684 topology-agnostic\u2014\u2014\u62d3\u6251\u65e0\u5173\u6027<\/strong>\uff0c\u5e76\u4e0d\u4ec5\u4ec5\u9002\u7528\u4e8e\u67d0\u4e00\u79cd\u56fa\u5b9a MZI \u7f51\u7edc\u3002<\/p>\n\n\n\n

\u7814\u7a76\u4eba\u5458\u751a\u81f3\u5229\u7528\u6563\u5c04\u4ecb\u8d28\u8bad\u7ec3\u51fa\u4e86\u89c4\u6a21\u5927\u4e8e\u539f\u751f\u53ef\u8c03\u5355\u5143\u6570\u91cf\u7684\u77e9\u9635\u3002<\/p>\n\n\n\n

\u8fdb\u4e00\u6b65\uff0c\u4ed6\u4eec\u53c8\u5c06 INSPIRE \u7528\u4e8e in situ meta-learning\u2014\u2014\u539f\u4f4d\u5143\u5b66\u4e60<\/strong>\uff0c\u62a5\u544a\u5b9e\u73b0 251 \u500d\u6a21\u578b\u538b\u7f29<\/strong>\u4ee5\u53ca 136 \u500d\u4efb\u52a1\u7279\u5b9a\u8bad\u7ec3\u52a0\u901f<\/strong>\u3002<\/p>\n\n\n\n

\u8fd9\u4e00\u6210\u679c\u7684\u610f\u4e49\u5f88\u503c\u5f97\u5f3a\u8c03\u3002<\/p>\n\n\n\n

Nokia Bell Labs \u7684\u7a81\u7834\u4e3b\u8981\u89e3\u51b3\uff1a<\/p>\n\n\n\n

\n

\u201c\u5149\u5b50\u795e\u7ecf\u7f51\u7edc\u80fd\u4e0d\u80fd\u5728\u82af\u7247\u4e0a\u8fdb\u884c\u53cd\u5411\u4f20\u64ad\uff1f\u201d<\/strong><\/p>\n<\/blockquote>\n\n\n\n

INSPIRE \u8fdb\u4e00\u6b65\u95ee\uff1a<\/p>\n\n\n\n

\n

\u201c\u80fd\u4e0d\u80fd\u4e0d\u4f9d\u8d56\u4e00\u4e2a\u5b8c\u7f8e\u7684\u6570\u5b57\u82af\u7247\u6a21\u578b\uff0c\u800c\u8ba9\u771f\u5b9e\u5149\u5b50\u82af\u7247\u76f4\u63a5\u6839\u636e\u81ea\u5df1\u7684\u7269\u7406\u72b6\u6001\u5b66\u4e60\uff1f\u201d<\/strong><\/p>\n<\/blockquote>\n\n\n\n

\u7b54\u6848\u5f00\u59cb\u53d8\u6210\uff1a<\/p>\n\n\n\n

\u53ef\u4ee5\u3002<\/strong><\/p>\n\n\n\n

\u8fd9\u610f\u5473\u7740\u5149\u5b50\u8ba1\u7b97\u7814\u7a76\u5f00\u59cb\u4ece Photonic Computing<\/strong> \u5411 Physical Learning\u2014\u2014\u7269\u7406\u5b66\u4e60<\/strong>\u63a8\u8fdb\u3002<\/p>\n\n\n\n

\u53ef\u4ee5\u628a\u76ee\u524d\u7684\u6f14\u8fdb\u7b80\u5355\u7406\u89e3\u4e3a\uff1a<\/p>\n\n\n\n

\u8f6f\u4ef6\u8bad\u7ec3 \u2192 \u786c\u4ef6\u63a8\u7406<\/strong><\/p>\n\n\n\n

\u2193<\/p>\n\n\n\n

\u5149\u5b50\u82af\u7247\u53c2\u4e0e\u8bad\u7ec3<\/strong><\/p>\n\n\n\n

\u2193<\/p>\n\n\n\n

\u76f4\u63a5\u6d4b\u91cf\u7269\u7406\u68af\u5ea6<\/strong><\/p>\n\n\n\n

\u2193<\/p>\n\n\n\n

\u82af\u7247\u6839\u636e\u81ea\u5df1\u7684\u771f\u5b9e\u7269\u7406\u72b6\u6001\u8fdb\u884c\u539f\u4f4d\u5b66\u4e60<\/strong><\/p>\n\n\n\n

\u8fd9\u662f\u5149\u5b50 AI \u8fd1\u51e0\u5e74\u975e\u5e38\u91cd\u8981\u7684\u4e00\u6761\u6280\u672f\u4e3b\u7ebf\u3002<\/p>\n\n\n\n

\n\n\n\n

\u4e03\u3001\u4e0d\u53ea\u662f\u8bad\u7ec3\uff1a\u5149\u5b50\u795e\u7ecf\u7f51\u7edc\u8fd8\u5728\u89e3\u51b3\u201c\u89c4\u6a21\u201d\u548c\u201c\u8bb0\u5fc6\u201d<\/h1>\n\n\n\n

\u53e6\u4e00\u4e2a\u95ee\u9898\u662f\uff1a<\/p>\n\n\n\n

AI \u6a21\u578b\u8d8a\u6765\u8d8a\u5927\uff0c\u4f46\u5149\u5b50\u7f51\u7edc\u4e0d\u80fd\u7b80\u5355\u65e0\u9650\u589e\u52a0\u6df1\u5ea6\u3002<\/p>\n\n\n\n

\u56e0\u4e3a\u5149\u8def\u8d8a\u6df1\uff1a<\/p>\n\n\n\n

\u635f\u8017\u8d8a\u5927\u3001\u8bef\u5dee\u8d8a\u5bb9\u6613\u7d2f\u79ef\u3001\u63a7\u5236\u590d\u6742\u5ea6\u4e5f\u8d8a\u9ad8\u3002<\/strong><\/p>\n\n\n\n

2026 \u5e74\u7684 Photonic Mixture-of-Experts<\/strong> \u5de5\u4f5c\u63d0\u51fa\u4e86\u4e00\u6761\u5f88\u6709\u610f\u601d\u7684\u8def\u5f84\uff1a<\/p>\n\n\n\n

\u4e0d\u4e00\u5b9a\u7ee7\u7eed\u52a0\u6df1\uff0c\u800c\u662f\u52a0\u5bbd\u3002<\/strong><\/p>\n\n\n\n

\u7814\u7a76\u4eba\u5458\u5728\u5355\u7247\u82af\u7247\u4e2d\u90e8\u7f72\u591a\u4e2a\u5149\u5b66 Expert\uff0c\u6839\u636e\u4efb\u52a1\u52a8\u6001\u8def\u7531\u5230\u4e0d\u540c\u5149\u5b50\u7f51\u7edc\u3002\u5b9e\u9a8c\u82af\u7247\u5305\u542b 3 \u4e2a\u534f\u540c\u5149\u5b66\u4e13\u5bb6\u7f51\u7edc\u548c 18 \u4e2a\u5e76\u884c kernel\uff0c\u5728\u591a\u9886\u57df\u56fe\u50cf\u5206\u7c7b\u4e2d\u5b9e\u73b0\u5e73\u5747 97.1%<\/strong> \u7684\u51c6\u786e\u7387\uff0c\u540c\u65f6\u51cf\u5c11\u7ea6 67% \u6570\u5b57\u53c2\u6570\u5f00\u9500<\/strong>\u3002<\/p>\n\n\n\n

\u8fd9\u5b9e\u9645\u4e0a\u628a\u6570\u5b57\u5927\u6a21\u578b\u4e2d\u7684 MoE \u601d\u60f3\u7b2c\u4e00\u6b21\u6bd4\u8f83\u5b8c\u6574\u5730\u79fb\u690d\u5230\u4e86\u7247\u4e0a\u5149\u8ba1\u7b97\u3002<\/p>\n\n\n\n

\u53e6\u5916\u4e00\u4e2a\u65b9\u5411\u5219\u662f\u4e09\u7ef4\u5149\u5b50\u96c6\u6210\u3002<\/p>\n\n\n\n

\u56e0\u4e3a\u4e8c\u7ef4\u82af\u7247\u4e0a\u7684\u6ce2\u5bfc\u4ea4\u53c9\u3001\u9762\u79ef\u548c I\/O \u6570\u91cf\u90fd\u4f1a\u6210\u4e3a\u74f6\u9888\uff0c\u7814\u7a76\u4eba\u5458\u5f00\u59cb\u5229\u7528 3D Photonic Neural Network<\/strong>\uff0c\u628a\u5149\u8def\u6269\u5c55\u5230\u5782\u76f4\u7ef4\u5ea6\uff0c\u4ece\u800c\u63d0\u5347\u5149\u8ba1\u7b97\u5bc6\u5ea6\u548c\u8fde\u63a5\u80fd\u529b\u3002<\/p>\n\n\n\n

\u672a\u6765\u7684\u5927\u89c4\u6a21\u5149\u8ba1\u7b97\uff0c\u5f88\u53ef\u80fd\u4e0d\u4f1a\u53ea\u662f\uff1a<\/p>\n\n\n\n

\u66f4\u5927\u7684\u4e8c\u7ef4\u5149\u5b66\u77e9\u9635\u3002<\/strong><\/p>\n\n\n\n

\u800c\u4f1a\u9010\u6e10\u8f6c\u5411\uff1a<\/p>\n\n\n\n

\u5149\u5b50 chiplet + 3D photonics + \u591a\u8ba1\u7b97\u6838\u5fc3\u3002<\/strong><\/p>\n\n\n\n

\n\n\n\n

\u516b\u3001\u5149\u5b50\u8ba1\u7b97\u5f00\u59cb\u62e5\u6709\u201c\u8bb0\u5fc6\u201d<\/h1>\n\n\n\n

\u8fd9\u662f\u53e6\u4e00\u4e2a\u975e\u5e38\u5173\u952e\u7684\u53d8\u5316\u3002<\/p>\n\n\n\n

\u5149\u6700\u5927\u7684\u4f18\u70b9\u4e4b\u4e00\u662f\uff1a<\/p>\n\n\n\n

\u8dd1\u5f97\u5feb\u3002<\/strong><\/p>\n\n\n\n

\u4f46\u5b83\u6700\u5927\u7684\u7f3a\u70b9\u4e5f\u662f\uff1a<\/p>\n\n\n\n

\u8dd1\u5f97\u592a\u5feb\uff0c\u5f88\u96be\u7559\u4e0b\u6765\u3002<\/strong><\/p>\n\n\n\n

\u7535\u5b50\u82af\u7247\u53ef\u4ee5\u5229\u7528 SRAM\u3001DRAM\u3001Flash \u7b49\u4fdd\u5b58\u72b6\u6001\uff0c\u800c\u4f20\u7edf\u5149\u5b50\u7cfb\u7edf\u5929\u7136\u7f3a\u4e4f\u7c7b\u4f3c\u80fd\u529b\u3002<\/p>\n\n\n\n

2026 \u5e74 8 \u6708\uff0c\u6e05\u534e\u56e2\u961f\u5728 Nature Sensors<\/em> \u53d1\u8868 Fully in-memory photonic computing<\/strong>\uff0c\u63d0\u51fa FLARE \u67b6\u6784\u3002<\/p>\n\n\n\n

\u7814\u7a76\u56e2\u961f\u628a\u5149\u3001\u7535\u4ee5\u53ca\u5149\u7535\u673a\u5236\u7ed3\u5408\uff0c\u5728\u4e00\u9897\u5355\u7247\u82af\u7247\u4e2d\u5b9e\u73b0 7,378 \u4e2a\u5177\u6709\u8bb0\u5fc6\u80fd\u529b\u7684\u5149\u5b50\u795e\u7ecf\u5143<\/strong>\u3002<\/p>\n\n\n\n

\u8fd9\u4e9b\u5149\u5b50\u795e\u7ecf\u5143\u540c\u65f6\u5177\u6709\uff1a<\/p>\n\n\n\n

GHz \u7ea7\u77ed\u671f\u52a8\u6001<\/strong><\/p>\n\n\n\n

\u4ee5\u53ca<\/p>\n\n\n\n

\u7ea6 7.45 \u79d2\u957f\u671f\u8bb0\u5fc6\u3002<\/strong><\/p>\n\n\n\n

\u7814\u7a76\u56e2\u961f\u8fdb\u4e00\u6b65\u7ec4\u6210\u5341\u6838\u3001\u591a\u5c42\u7cfb\u7edf\uff0c\u5e76\u7528\u4e8e\u81ea\u4e3b\u7ade\u901f\u65e0\u4eba\u673a\u5bfc\u822a\u548c\u73af\u5883\u9002\u5e94\u3002<\/p>\n\n\n\n

\u5b83\u7684\u91cd\u8981\u610f\u4e49\u751a\u81f3\u53ef\u80fd\u8d85\u8fc7\u5355\u7eaf\u63d0\u9ad8 TOPS\u3002<\/p>\n\n\n\n

\u56e0\u4e3a\u672a\u6765 AI \u82af\u7247\u771f\u6b63\u8981\u89e3\u51b3\u7684\u4e0d\u53ea\u662f\uff1a<\/p>\n\n\n\n

\u201c\u600e\u4e48\u7b97\u5f97\u5feb\uff1f\u201d<\/strong><\/p>\n\n\n\n

\u800c\u662f\uff1a<\/p>\n\n\n\n

\u201c\u6570\u636e\u4e3a\u4ec0\u4e48\u8981\u4e0d\u65ad\u4ece\u5b58\u50a8\u5668\u642c\u5230\u8ba1\u7b97\u5355\u5143\uff0c\u518d\u4ece\u8ba1\u7b97\u5355\u5143\u642c\u56de\u5b58\u50a8\u5668\uff1f\u201d<\/strong><\/p>\n\n\n\n

\u5982\u679c\u5149\u8ba1\u7b97\u80fd\u591f\u628a\uff1a<\/p>\n\n\n\n

Sensor + Memory + Compute<\/strong><\/p>\n\n\n\n

\u8fdb\u4e00\u6b65\u878d\u5408\uff0c\u5c31\u6709\u53ef\u80fd\u51fa\u73b0\u771f\u6b63\u610f\u4e49\u4e0a\u7684\uff1a<\/p>\n\n\n\n

\u5149\u5b50\u611f\u5b58\u7b97\u4e00\u4f53\u3002<\/strong><\/p>\n\n\n\n

\n\n\n\n

\u4e5d\u3001\u79d1\u7814\u5df2\u7ecf\u8d70\u5f97\u5f88\u8fdc\uff0c\u4f46\u4ea7\u4e1a\u9996\u5148\u7206\u53d1\u7684\u662f\u201c\u5149\u4e92\u8fde\u201d<\/h1>\n\n\n\n

\u8fd9\u91cc\u5c31\u51fa\u73b0\u4e86\u4e00\u4e2a\u5f88\u6709\u610f\u601d\u7684\u79d1\u7814\u2014\u4ea7\u4e1a\u5dee\u5f02\u3002<\/p>\n\n\n\n

\u5b9e\u9a8c\u5ba4\u6700\u5438\u5f15\u773c\u7403\u7684\u662f\uff1a<\/p>\n\n\n\n

\u201c\u5149\u80fd\u4e0d\u80fd\u76f4\u63a5\u505a AI \u8ba1\u7b97\uff1f\u201d<\/strong><\/p>\n\n\n\n

\u4ea7\u4e1a\u754c\u76ee\u524d\u6700\u613f\u610f\u82b1\u94b1\u89e3\u51b3\u7684\u5374\u662f\uff1a<\/p>\n\n\n\n

\u201c\u600e\u6837\u8ba9\u6210\u5343\u4e0a\u4e07\u9897 GPU \u66f4\u9ad8\u6548\u5730\u4e92\u76f8\u901a\u4fe1\uff1f\u201d<\/strong><\/p>\n\n\n\n

2026 \u5e74\u6700\u5178\u578b\u7684\u4f8b\u5b50\u6765\u81ea NVIDIA\u3002<\/p>\n\n\n\n

NVIDIA Vera Rubin \u5e73\u53f0\u5df2\u7ecf\u628a Spectrum-X Ethernet Photonics<\/strong> \u7eb3\u5165\u6b63\u5f0f\u67b6\u6784\u3002<\/p>\n\n\n\n

\u5176\u4ea4\u6362\u7cfb\u7edf\u91c7\u7528 CPO\uff0c\u628a\u5149\u5b66\u5f15\u64ce\u4e0e\u4ea4\u6362 ASIC \u66f4\u7d27\u5bc6\u5730\u5c01\u88c5\u5728\u4e00\u8d77\uff0c\u5e76\u4f7f\u7528 200 Gb\/s SerDes<\/strong>\u3002<\/p>\n\n\n\n

NVIDIA \u5728 2026 \u5e74 5 \u6708\u5ba3\u5e03 Spectrum-X Ethernet Photonics \u5df2\u8fdb\u5165\u751f\u4ea7\uff0c\u5e76\u5ba3\u79f0\u5176\u76f8\u6bd4\u4f20\u7edf\u53ef\u63d2\u62d4\u5149\u6536\u53d1\u5668\u7f51\u7edc\u53ef\u4ee5\u8fbe\u5230\u7ea6 5 \u500d\u7f51\u7edc\u529f\u8017\u6548\u7387<\/strong>\u3002<\/p>\n\n\n\n

\u8fd9\u662f\u4e00\u4e2a\u5f88\u91cd\u8981\u7684\u4ea7\u4e1a\u4fe1\u53f7\uff1a<\/p>\n\n\n\n

\n

Photonics \u5df2\u7ecf\u5f00\u59cb\u8fdb\u5165 AI Factory \u7684\u6838\u5fc3\u57fa\u7840\u8bbe\u65bd\u3002<\/strong><\/p>\n<\/blockquote>\n\n\n\n

\n\n\n\n

\u5341\u3001Marvell \u6536\u8d2d Celestial AI\uff1a\u5149\u4e92\u8fde\u5f00\u59cb\u6210\u4e3a\u6838\u5fc3\u534a\u5bfc\u4f53\u8d44\u4ea7<\/h1>\n\n\n\n

2026 \u5e74 2 \u6708\uff0cMarvell \u5b8c\u6210\u5bf9 Celestial AI \u7684\u6536\u8d2d\u3002<\/p>\n\n\n\n

Celestial AI \u7684\u6838\u5fc3\u6280\u672f\u4e0d\u662f\u201c\u7528\u5149\u4ee3\u66ff GPU \u7b97\u77e9\u9635\u201d\uff0c\u800c\u662f\uff1a<\/p>\n\n\n\n

Photonic Fabric\u3002<\/strong><\/p>\n\n\n\n

\u5b83\u7684\u76ee\u6807\u662f\u5229\u7528\u5149\u8fde\u63a5\u5927\u89c4\u6a21 AI \u52a0\u901f\u5668\uff0c\u89e3\u51b3 scale-up \u7f51\u7edc\u4e2d\u7684\u5e26\u5bbd\u548c\u529f\u8017\u95ee\u9898\u3002<\/p>\n\n\n\n

\u8fd9\u4e2a\u4e8b\u4ef6\u975e\u5e38\u5177\u6709\u4ee3\u8868\u6027\u3002<\/p>\n\n\n\n

\u56e0\u4e3a\u5b83\u8bf4\u660e\u4f20\u7edf\u534a\u5bfc\u4f53\u5de8\u5934\u5df2\u7ecf\u5f00\u59cb\u628a\u5149\u4e92\u8fde\u89c6\u4e3a\uff1a<\/p>\n\n\n\n

\u4e0b\u4e00\u4ee3 AI \u82af\u7247\u7cfb\u7edf\u4e2d\u7684\u6838\u5fc3 IP\u3002<\/strong><\/p>\n\n\n\n

\u800c\u4e0d\u518d\u53ea\u662f\u4e00\u4e2a\u5916\u56f4\u901a\u4fe1\u5668\u4ef6\u3002<\/p>\n\n\n\n

\n\n\n\n

\u5341\u4e00\u3001Lightmatter \u7684\u8f6c\u5411\u5c24\u5176\u503c\u5f97\u6ce8\u610f<\/h1>\n\n\n\n

Lightmatter \u6700\u521d\u56e0\u4e3a\u5149\u8ba1\u7b97\u82af\u7247\u53d7\u5230\u5173\u6ce8\u3002<\/p>\n\n\n\n

\u4f46\u4eca\u5929\u516c\u53f8\u7684\u5546\u4e1a\u91cd\u5fc3\u660e\u663e\u8d8a\u6765\u8d8a\u9760\u8fd1\uff1a<\/p>\n\n\n\n

Passage \u5149\u4e92\u8fde\u5e73\u53f0\u3002<\/strong><\/p>\n\n\n\n

2026 \u5e74\u63a8\u51fa\u7684 Passage L20 \u63d0\u4f9b\uff1a<\/p>\n\n\n\n

\u5355\u65b9\u5411 6.4 Tbps\uff0c\u4e5f\u5c31\u662f\u53d1\u9001\u548c\u63a5\u6536\u5404 6.4 Tbps\u3002<\/strong><\/p>\n\n\n\n

\u540c\u65f6\u91c7\u7528\u53cc\u5411\u6ce2\u957f\u590d\u7528\uff0c\u8ba9\u540c\u4e00\u6839\u5149\u7ea4\u627f\u62c5\u4e24\u4e2a\u65b9\u5411\u7684\u6570\u636e\u4f20\u8f93\u3002<\/p>\n\n\n\n

\u516c\u53f8\u8fd8\u5728\u63a8\u52a8 1.6 Tbps\/fiber \u7b49\u66f4\u9ad8\u5bc6\u5ea6\u5149\u8fde\u63a5\u3002<\/p>\n\n\n\n

\u8fd9\u6761\u4ea7\u54c1\u8def\u7ebf\u975e\u5e38\u5177\u6709\u542f\u793a\u6027\uff1a<\/p>\n\n\n\n

\u5148\u628a\u5149\u7528\u4e8e\u82af\u7247\u4e4b\u95f4\uff0c\u518d\u9010\u6e10\u628a\u5149\u63a8\u8fdb\u82af\u7247\u5185\u90e8\u3002<\/strong><\/p>\n\n\n\n

\u4e5f\u5c31\u662f\u8bf4\uff1a<\/p>\n\n\n\n

Interconnect \u2192 Optical I\/O \u2192 Photonic Interposer \u2192 Photonic Compute<\/strong><\/p>\n\n\n\n

\u53ef\u80fd\u6bd4\u76f4\u63a5\u5236\u9020\u4e00\u4e2a\u201c\u7eaf\u5149 GPU\u201d\u66f4\u52a0\u73b0\u5b9e\u3002<\/p>\n\n\n\n

\n\n\n\n

\u5341\u4e8c\u3001Ayar Labs\uff1a\u4e13\u6ce8 Optical I\/O<\/h1>\n\n\n\n

Ayar Labs \u7684\u601d\u8def\u66f4\u52a0\u660e\u786e\uff1a<\/p>\n\n\n\n

\u5b83\u57fa\u672c\u4e0d\u53bb\u91cd\u65b0\u8bbe\u8ba1\u4e00\u9897 GPU\uff0c\u800c\u662f\u4e13\u6ce8\u4e8e\uff1a<\/p>\n\n\n\n

\u5982\u4f55\u628a GPU\/XPU \u7684\u7535 I\/O \u53d8\u6210\u5149 I\/O\u3002<\/strong><\/p>\n\n\n\n

2026 \u5e74 9 \u6708\uff0c\u516c\u53f8\u5ba3\u5e03\u65b0\u589e 1.5 \u4ebf\u7f8e\u5143\u878d\u8d44\uff0c\u4f7f 2026 \u5e74\u4e00\u7ea7\u8d44\u672c\u878d\u8d44\u603b\u989d\u8fbe\u5230 6.5 \u4ebf\u7f8e\u5143<\/strong>\u3002<\/p>\n\n\n\n

AMD\u3001Intel\u3001MediaTek\u3001NVIDIA \u7b49\u90fd\u5728\u5176\u6218\u7565\u6295\u8d44\u8005\u548c\u4ea7\u4e1a\u751f\u6001\u4e2d\u3002\u8d44\u91d1\u4e3b\u8981\u7528\u4e8e\u4ea7\u54c1\u5f00\u53d1\u3001\u9a8c\u8bc1\u3001\u5236\u9020\u751f\u6001\u5efa\u8bbe\u4ee5\u53ca\u9ad8\u4ea7\u91cf\u751f\u4ea7\u51c6\u5907\u3002<\/p>\n\n\n\n

\u8fd9\u8bf4\u660e\u4ea7\u4e1a\u7ade\u4e89\u5df2\u7ecf\u4ece\uff1a<\/p>\n\n\n\n

\u201c\u5149\u82af\u7247\u80fd\u4e0d\u80fd\u505a\u51fa\u6765\uff1f\u201d<\/strong><\/p>\n\n\n\n

\u9010\u6e10\u8fdb\u5165\uff1a<\/p>\n\n\n\n

\u201c\u80fd\u4e0d\u80fd\u5236\u9020\u3001\u5c01\u88c5\u3001\u6d4b\u8bd5\u5e76\u5927\u6279\u91cf\u4ea4\u4ed8\uff1f\u201d<\/strong><\/p>\n\n\n\n

\n\n\n\n

\u5341\u4e09\u3001\u771f\u6b63\u7684\u201c\u5149 AI \u8ba1\u7b97\u4ea7\u54c1\u201d\u4e5f\u5f00\u59cb\u51fa\u73b0<\/h1>\n\n\n\n

\u4ea7\u4e1a\u53e6\u4e00\u7aef\uff0c\u771f\u6b63\u7528\u5149\u53c2\u4e0e AI \u6570\u5b66\u8fd0\u7b97\u7684\u7cfb\u7edf\u4e5f\u5f00\u59cb\u8d70\u51fa\u5b9e\u9a8c\u5ba4\u3002<\/p>\n\n\n\n

Lightelligence<\/h3>\n\n\n\n

Lightelligence \u7684 PACE 2<\/strong> \u5df2\u7ecf\u505a\u6210 PCIe Gen4 x16 \u52a0\u901f\u5361\u3002<\/p>\n\n\n\n

\u5176\u6838\u5fc3\u5305\u62ec\uff1a<\/p>\n\n\n\n

128\u00d7128 \u53ef\u914d\u7f6e\u5149\u5b66\u77e9\u9635<\/strong><\/p>\n\n\n\n

8-bit \u5149\u8ba1\u7b97\u8f93\u51fa\u7cbe\u5ea6<\/strong><\/p>\n\n\n\n

\u8d85\u8fc7 40,000 \u4e2a\u5149\u5b50\u5668\u4ef6<\/strong><\/p>\n\n\n\n

\u4ee5\u53ca<\/p>\n\n\n\n

3D TSV \u5148\u8fdb\u5c01\u88c5\u3002<\/strong><\/p>\n\n\n\n

\u8fd9\u6761\u8def\u7ebf\u975e\u5e38\u91cd\u8981\uff1a<\/p>\n\n\n\n

\u4e0d\u662f\u8981\u6c42\u7528\u6237\u5efa\u7acb\u4e00\u53f0\u5168\u65b0\u7684\u201c\u5149\u8ba1\u7b97\u673a\u201d\uff0c\u800c\u662f\u8ba9\u5149\u8ba1\u7b97\u4f5c\u4e3a\u666e\u901a\u670d\u52a1\u5668\u4e2d\u7684\u4e00\u4e2a\u5f02\u6784\u52a0\u901f\u5668\u5b58\u5728\u3002<\/p>\n\n\n\n

\n\n\n\n

Lumai<\/h3>\n\n\n\n

2026 \u5e74 4 \u6708\uff0c\u725b\u6d25\u5927\u5b66\u5b75\u5316\u516c\u53f8 Lumai \u63a8\u51fa Iris<\/strong> \u5149\u8ba1\u7b97\u63a8\u7406\u670d\u52a1\u5668\u7cfb\u5217\u3002<\/p>\n\n\n\n

\u5176\u4e2d Iris Nova \u5df2\u7ecf\u80fd\u591f\u8fd0\u884c Llama 8B \u548c Llama 70B<\/strong>\u3002<\/p>\n\n\n\n

\u5176\u7cfb\u7edf\u4ecd\u7136\u662f\u6df7\u5408\u67b6\u6784\uff1a<\/p>\n\n\n\n

\u6570\u5b57\u5904\u7406\u5668\u8d1f\u8d23\u8f6f\u4ef6\u4e0e\u63a7\u5236\uff0cOptical Tensor Engine \u8d1f\u8d23\u6838\u5fc3\u6570\u5b66\u8ba1\u7b97\u3002<\/strong><\/p>\n\n\n\n

Lumai \u5ba3\u79f0\u7279\u5b9a\u67b6\u6784\u4e0b\u80fd\u591f\u5b9e\u73b0\u6700\u9ad8\u7ea6 90% \u7684\u80fd\u8017\u964d\u4f4e<\/strong>\uff0c\u4e0d\u8fc7\u8fd9\u4e00\u6307\u6807\u76ee\u524d\u4ecd\u4e3b\u8981\u6765\u81ea\u4f01\u4e1a\u81ea\u8eab\u62ab\u9732\uff0c\u9700\u8981\u66f4\u591a\u72ec\u7acb benchmark \u9a8c\u8bc1\u3002<\/p>\n\n\n\n

\u5c24\u5176\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff1a<\/p>\n\n\n\n

Lumai \u628a\u91cd\u70b9\u653e\u5230\u4e86 LLM Prefill<\/strong>\u3002<\/p>\n\n\n\n

\u8fd9\u662f\u975e\u5e38\u5408\u7406\u7684\u3002<\/p>\n\n\n\n

\u56e0\u4e3a prefill \u9636\u6bb5\u5177\u6709\u5927\u91cf\u5bc6\u96c6\u77e9\u9635\u8fd0\u7b97\uff0c\u975e\u5e38\u9002\u5408\u5149\u5b66\u5e76\u884c\u8ba1\u7b97\uff1b\u800c decode \u9636\u6bb5\u5219\u66f4\u591a\u53d7\u5230 KV Cache \u548c\u5185\u5b58\u5e26\u5bbd\u5236\u7ea6\u3002<\/p>\n\n\n\n

\n\n\n\n

Q.ANT<\/h3>\n\n\n\n

Q.ANT \u5219\u91c7\u7528\u53e6\u4e00\u79cd\u6750\u6599\u4f53\u7cfb\uff1a<\/p>\n\n\n\n

Thin-Film Lithium Niobate on Insulator\u2014\u2014\u8584\u819c\u94cc\u9178\u9502\u3002<\/strong><\/p>\n\n\n\n

\u4e0e\u5927\u91cf\u4f9d\u8d56\u70ed\u8c03\u5236\u7684\u7845\u5149\u5668\u4ef6\u4e0d\u540c\uff0c\u94cc\u9178\u9502\u53ef\u4ee5\u901a\u8fc7\u7535\u573a\u6539\u53d8\u5149\u5b66\u7279\u6027\uff0c\u4ece\u800c\u51cf\u5c11\u70ed\u4e32\u6270\uff0c\u5e76\u5177\u6709\u9ad8\u901f\u8c03\u5236\u6f5c\u529b\u3002<\/p>\n\n\n\n

\u5176 Native Processing Server \u5df2\u7ecf\u505a\u6210\u6807\u51c6 19 \u82f1\u5bf8\u670d\u52a1\u5668<\/strong>\uff0c\u5185\u90e8\u914d\u7f6e PCIe \u5149\u5b50 NPU\u3002<\/p>\n\n\n\n

2026 \u5e74 6 \u6708\uff0cQ.ANT \u53c8\u5c55\u793a\u4e86\u5728\u7b2c\u4e8c\u4ee3 NPU \u4e0a\u8fd0\u884c\uff1a<\/p>\n\n\n\n

\u6269\u6563\u6a21\u578b<\/strong><\/p>\n\n\n\n

\u4ee5\u53ca<\/p>\n\n\n\n

xLSTM \u5faa\u73af\u795e\u7ecf\u7f51\u7edc\u3002<\/strong><\/p>\n\n\n\n

\u8fd9\u610f\u5473\u7740\u5546\u4e1a\u5149\u5b50\u786c\u4ef6\u6b63\u5728\u4ece\u4f20\u7edf\u5206\u7c7b\u4efb\u52a1\u8fdb\u5165\u751f\u6210\u5f0f AI \u548c\u65f6\u95f4\u5e8f\u5217 AI \u5de5\u4f5c\u8d1f\u8f7d\u3002<\/p>\n\n\n\n

\n\n\n\n

\u5341\u56db\u3001\u4e0b\u4e00\u4e2a\u6218\u573a\u53ef\u80fd\u662f GPU \u4e0e\u5185\u5b58\u4e4b\u95f4<\/h1>\n\n\n\n

Photonics \u7684\u53e6\u4e00\u4e2a\u91cd\u8981\u65b9\u5411\u6b63\u5728\u51fa\u73b0\uff1a<\/p>\n\n\n\n

GPU \u2194 \u5185\u5b58\u3002<\/strong><\/p>\n\n\n\n

Volantis \u5c31\u662f\u4e00\u4e2a\u5178\u578b\u4f8b\u5b50\u3002<\/p>\n\n\n\n

\u5b83\u4e0d\u662f\u5236\u9020\u5149\u8ba1\u7b97\u6838\u5fc3\uff0c\u800c\u662f\u8bd5\u56fe\u5229\u7528 VCSEL \u7b49\u5149\u4e92\u8fde\u6280\u672f\uff0c\u628a GPU \u4e0e\u66f4\u591a\u5185\u5b58\u5668\u4ef6\u8fde\u63a5\u8d77\u6765\uff0c\u4ece\u800c\u7a81\u7834\u7535\u4e92\u8fde\u8ddd\u79bb\u548c\u5c01\u88c5\u89c4\u6a21\u9650\u5236\u3002<\/p>\n\n\n\n

\u8fd9\u4e2a\u601d\u8def\u80cc\u540e\u7684\u903b\u8f91\u5176\u5b9e\u5f88\u7b80\u5355\uff1a<\/p>\n\n\n\n

\u4eca\u5929 AI \u9762\u4e34\u7684\u5f88\u591a\u95ee\u9898\uff0c\u4e0d\u662f GPU \u7b97\u4e0d\u52a8\uff0c\u800c\u662f\uff1a<\/p>\n\n\n\n

\u6570\u636e\u5582\u4e0d\u8fdb\u53bb\u3002<\/strong><\/p>\n\n\n\n

\u6240\u4ee5\u672a\u6765\u5149\u4e92\u8fde\u5f88\u53ef\u80fd\u4ece\uff1a<\/p>\n\n\n\n

\u670d\u52a1\u5668 \u2194 \u670d\u52a1\u5668<\/p>\n\n\n\n

\u9010\u6e10\u8fdb\u5165\uff1a<\/p>\n\n\n\n

GPU \u2194 GPU<\/p>\n\n\n\n

\u2193<\/p>\n\n\n\n

Package \u2194 Package<\/p>\n\n\n\n

\u2193<\/p>\n\n\n\n

Chiplet \u2194 Chiplet<\/p>\n\n\n\n

\u2193<\/p>\n\n\n\n

GPU \u2194 HBM \/ DRAM<\/strong><\/p>\n\n\n\n

\u751a\u81f3\u6700\u7ec8\u8fdb\u5165\u7247\u5185\u7f51\u7edc\u3002<\/p>\n\n\n\n

Intel \u7684 Optical Compute Interconnect\uff08OCI\uff09<\/strong> \u5c31\u5df2\u7ecf\u6cbf\u7740\u8fd9\u4e00\u65b9\u5411\u53d1\u5c55\u3002<\/p>\n\n\n\n

Intel \u7b2c\u4e00\u4ee3 OCI chiplet \u652f\u6301 4 Tbps \u53cc\u5411\u8fde\u63a5<\/strong>\uff0c\u5185\u90e8\u96c6\u6210\u7845\u5149 PIC\u3001DWDM \u6fc0\u5149\u5668\u3001\u5149\u653e\u5927\u5668\u4ee5\u53ca CMOS EIC\uff0c\u5e76\u8ba1\u5212\u5411\u5355\u5668\u4ef6\u6570\u5341 Tbps \u6f14\u8fdb\u3002<\/p>\n\n\n\n

\n\n\n\n

\u5341\u4e94\u3001\u4e3a\u4ec0\u4e48\u4e0d\u80fd\u770b\u5230\u201c100 TOPS\/W\u201d\u5c31\u8ba4\u4e3a GPU \u8981\u88ab\u6dd8\u6c70\uff1f<\/h1>\n\n\n\n

\u8fd9\u662f\u7406\u89e3\u5149\u5b50 AI \u6700\u91cd\u8981\u7684\u95ee\u9898\u4e4b\u4e00\u3002<\/p>\n\n\n\n

\u8bba\u6587\u4e2d\u7684\u5149\u8ba1\u7b97\u6027\u80fd\u7ecf\u5e38\u975e\u5e38\u60ca\u4eba\u3002<\/p>\n\n\n\n

\u539f\u56e0\u662f\uff1a<\/p>\n\n\n\n

\u5149\u5b8c\u6210\u4e00\u6b21\u77e9\u9635\u53d8\u6362\uff0c\u672c\u8eab\u53ef\u80fd\u51e0\u4e4e\u4e0d\u9700\u8981\u50cf\u6570\u5b57\u6676\u4f53\u7ba1\u90a3\u6837\u4e0d\u65ad\u5f00\u5173\u3002<\/p>\n\n\n\n

\u4f46\u662f\u5b8c\u6574\u5149\u8ba1\u7b97\u7cfb\u7edf\u8fd8\u6709\u5927\u91cf\u5916\u56f4\u5668\u4ef6\uff1a<\/p>\n\n\n\n

Laser<\/strong><\/p>\n\n\n\n

\u2193<\/p>\n\n\n\n

DAC<\/strong><\/p>\n\n\n\n

\u2193<\/p>\n\n\n\n

Modulator<\/strong><\/p>\n\n\n\n

\u2193<\/p>\n\n\n\n

Photonic Matrix<\/strong><\/p>\n\n\n\n

\u2193<\/p>\n\n\n\n

Photodetector<\/strong><\/p>\n\n\n\n

\u2193<\/p>\n\n\n\n

TIA<\/strong><\/p>\n\n\n\n

\u2193<\/p>\n\n\n\n

ADC<\/strong><\/p>\n\n\n\n

\u2193<\/p>\n\n\n\n

Digital Memory \/ Control<\/strong><\/p>\n\n\n\n

\u771f\u6b63\u7684\u670d\u52a1\u5668\u80fd\u6548\uff0c\u5fc5\u987b\u8ba1\u7b97\u6574\u6761\u94fe\u3002<\/p>\n\n\n\n

\u5982\u679c\u53ea\u7edf\u8ba1\u4e2d\u95f4\u7684 optical MAC\uff0cTOPS\/W \u4f1a\u975e\u5e38\u6f02\u4eae\u3002<\/p>\n\n\n\n

\u4f46\u5982\u679c\u628a\uff1a<\/p>\n\n\n\n

\u6fc0\u5149\u5668\u3001ADC\/DAC\u3001\u6e29\u63a7\u3001PCIe\u3001HBM\u3001\u63a7\u5236\u82af\u7247\u548c\u6570\u636e\u642c\u8fd0<\/p>\n\n\n\n

\u5168\u90e8\u8ba1\u7b97\u8fdb\u53bb\uff0c\u4f18\u52bf\u5f80\u5f80\u4f1a\u660e\u663e\u964d\u4f4e\u3002<\/p>\n\n\n\n

\u6240\u4ee5\u8bc4\u4ef7\u672a\u6765\u5149\u5b50 AI \u82af\u7247\uff0c\u4e0d\u80fd\u53ea\u770b\uff1a<\/p>\n\n\n\n

TOPS\/W\u3002<\/strong><\/p>\n\n\n\n

\u8fd8\u5e94\u8be5\u770b\uff1a<\/p>\n\n\n\n

tokens\/s<\/strong><\/p>\n\n\n\n

tokens\/J<\/strong><\/p>\n\n\n\n

\u6a21\u578b\u7cbe\u5ea6<\/strong><\/p>\n\n\n\n

\u7aef\u5230\u7aef\u670d\u52a1\u5668\u529f\u8017<\/strong><\/p>\n\n\n\n

ADC\/DAC \u529f\u8017<\/strong><\/p>\n\n\n\n

\u5149\u6e90\u529f\u8017<\/strong><\/p>\n\n\n\n

\u5185\u5b58\u548c\u901a\u4fe1\u529f\u8017<\/strong><\/p>\n\n\n\n

\u4ee5\u53ca<\/p>\n\n\n\n

\u5b9e\u9645\u8ba1\u7b97\u5229\u7528\u7387\u3002<\/strong><\/p>\n\n\n\n

\u8fd9\u4e5f\u662f\u4e3a\u4ec0\u4e48 ACCEL\u3001FLARE\u3001INSPIRE \u7b49\u7814\u7a76\u5982\u6b64\u503c\u5f97\u5173\u6ce8\u3002<\/p>\n\n\n\n

\u5b83\u4eec\u5df2\u7ecf\u4e0d\u518d\u5355\u7eaf\u8ffd\u6c42\uff1a<\/p>\n\n\n\n

\u201c\u8ba9\u4e2d\u95f4\u90a3\u4e2a\u5149\u5b66\u77e9\u9635\u66f4\u5feb\u3002\u201d<\/strong><\/p>\n\n\n\n

\u800c\u662f\u5728\u89e3\u51b3\u5b8c\u6574\u7cfb\u7edf\u4e2d\u7684\uff1a<\/p>\n\n\n\n

AD \u8f6c\u6362\u3001\u8bad\u7ec3\u3001\u5b58\u50a8\u3001\u5668\u4ef6\u8bef\u5dee\u548c\u7cfb\u7edf\u9002\u5e94\u6027\u3002<\/strong><\/p>\n\n\n\n

\n\n\n\n

\u5341\u516d\u3001\u5149\u5b50 AI \u771f\u6b63\u8fd8\u7f3a\u4ec0\u4e48\uff1f<\/h1>\n\n\n\n

\u76ee\u524d\u5269\u4e0b\u7684\u6838\u5fc3\u95ee\u9898\u5df2\u7ecf\u6bd4\u8f83\u6e05\u695a\u3002<\/p>\n\n\n\n

\u9996\u5148\u662f\u7cbe\u5ea6\u4e0e\u7a33\u5b9a\u6027<\/strong>\u3002<\/p>\n\n\n\n

\u5149\u5b50\u6a21\u62df\u8ba1\u7b97\u5929\u7136\u53d7\u5230\u70ed\u6f02\u79fb\u3001\u5668\u4ef6\u8bef\u5dee\u3001\u6fc0\u5149\u566a\u58f0\u3001\u5de5\u827a\u53d8\u5316\u5f71\u54cd\u3002<\/p>\n\n\n\n

INSPIRE \u8fd9\u79cd\u539f\u4f4d\u7269\u7406\u5b66\u4e60\u7684\u91cd\u8981\u610f\u4e49\uff0c\u5c31\u662f\u5c1d\u8bd5\u628a\u8fd9\u4e9b\u201c\u4e0d\u5b8c\u7f8e\u201d\u76f4\u63a5\u7eb3\u5165\u5b66\u4e60\u8fc7\u7a0b\u3002<\/p>\n\n\n\n

\u7b2c\u4e8c\u662f\u975e\u7ebf\u6027<\/strong>\u3002<\/p>\n\n\n\n

\u77e9\u9635\u4e58\u6cd5\u975e\u5e38\u9002\u5408\u5149\uff0c\u4f46 ReLU\u3001Softmax\u3001Normalization\u3001\u903b\u8f91\u63a7\u5236\u7b49\u8fd0\u7b97\u5e76\u4e0d\u5929\u7136\u9002\u5408\u5149\u3002<\/p>\n\n\n\n

\u56e0\u6b64\u81f3\u5c11\u5728\u53ef\u9884\u89c1\u7684\u672a\u6765\uff1a<\/p>\n\n\n\n

Optical + Electronic<\/strong><\/p>\n\n\n\n

\u4ecd\u7136\u6bd4\u7eaf\u5149\u65b9\u6848\u66f4\u52a0\u73b0\u5b9e\u3002<\/p>\n\n\n\n

\u7b2c\u4e09\u662f\u5b58\u50a8<\/strong>\u3002<\/p>\n\n\n\n

\u5927\u578b\u6a21\u578b\u6700\u7ec8\u9700\u8981\u5927\u91cf\u53c2\u6570\u548c KV Cache\u3002<\/p>\n\n\n\n

FLARE \u5df2\u7ecf\u5c55\u793a\u5149\u5b50\u8bb0\u5fc6\u795e\u7ecf\u5143\uff0c\u4f46\u8ddd\u79bb TB \u7ea7\u5927\u6a21\u578b\u53c2\u6570\u5b58\u50a8\u4ecd\u7136\u975e\u5e38\u8fdc\u3002<\/p>\n\n\n\n

\u7b2c\u56db\u662f\u8f6c\u6362\u5f00\u9500<\/strong>\u3002<\/p>\n\n\n\n

\u5982\u679c\u6bcf\u6b21\u77e9\u9635\u8ba1\u7b97\u524d\u540e\u90fd\u5fc5\u987b\u8fdb\u884c\uff1a<\/p>\n\n\n\n

Digital \u2192 Analog \u2192 Optical \u2192 Analog \u2192 Digital<\/strong><\/p>\n\n\n\n

\u90a3\u4e48\u5916\u56f4\u8f6c\u6362\u5668\u53ef\u80fd\u5403\u6389\u5927\u91cf\u6536\u76ca\u3002<\/p>\n\n\n\n

\u7b2c\u4e94\u662f\u8f6f\u4ef6\u751f\u6001<\/strong>\u3002<\/p>\n\n\n\n

CUDA \u771f\u6b63\u5f3a\u5927\u7684\u5730\u65b9\u4e0d\u53ea\u662f GPU\uff0c\u800c\u662f\u56f4\u7ed5 GPU \u5efa\u7acb\u7684\uff1a<\/p>\n\n\n\n

PyTorch\u3001\u7f16\u8bd1\u5668\u3001kernel library\u3001runtime\u3001\u8c03\u5ea6\u5de5\u5177\u4ee5\u53ca\u5f00\u53d1\u8005\u751f\u6001\u3002<\/p>\n\n\n\n

\u5149\u5b50\u8ba1\u7b97\u5fc5\u987b\u5c3d\u53ef\u80fd\u505a\u5230\uff1a<\/p>\n\n\n\n

\u7528\u6237\u4e0d\u9700\u8981\u61c2\u5149\u5b66\u3002<\/strong><\/p>\n\n\n\n

\u8fd9\u4e5f\u662f PACE 2\u3001Lumai Iris\u3001Q.ANT NPS \u90fd\u5f00\u59cb\u91c7\u7528 PCIe\u3001Python API\u3001\u6807\u51c6\u670d\u52a1\u5668\u5f62\u6001\u7684\u91cd\u8981\u539f\u56e0\u3002<\/p>\n\n\n\n

\u7b2c\u516d\u662f\u5c01\u88c5\u4e0e\u5236\u9020<\/strong>\u3002<\/p>\n\n\n\n

\u8fdb\u5165 CPO \u65f6\u4ee3\u4ee5\u540e\uff0c\u771f\u6b63\u56f0\u96be\u7684\u95ee\u9898\u5305\u62ec\uff1a<\/p>\n\n\n\n

PIC\u3001ASIC\u3001laser\u3001fiber attach\u3001thermal management\u3001yield\u3001repairability\u3002<\/p>\n\n\n\n

\u5f88\u591a\u95ee\u9898\u5df2\u7ecf\u4e0d\u518d\u662f\u201c\u5149\u5b66\u8bba\u6587\u95ee\u9898\u201d\uff0c\u800c\u662f\u6807\u51c6\u7684\uff1a<\/p>\n\n\n\n

\u534a\u5bfc\u4f53\u4ea7\u4e1a\u5de5\u7a0b\u95ee\u9898\u3002<\/strong><\/p>\n\n\n\n

\n\n\n\n

\u5341\u4e03\u3001\u672a\u6765\u771f\u6b63\u7684\u5f62\u6001\u53ef\u80fd\u4e0d\u662f\u201c\u5149\u5b50 GPU\u201d<\/h1>\n\n\n\n

\u7efc\u5408\u76ee\u524d\u79d1\u7814\u4e0e\u4ea7\u4e1a\u4fe1\u606f\uff0c\u6211\u8ba4\u4e3a\u672a\u6765 5\u201410 \u5e74\u5149\u5b50 AI \u6700\u53ef\u80fd\u51fa\u73b0\u7684\u5f62\u6001\uff0c\u5e76\u4e0d\u662f\uff1a<\/p>\n\n\n\n

\u4e00\u9897\u5149\u5b50\u82af\u7247\u7a81\u7136\u53d6\u4ee3 NVIDIA GPU\u3002<\/strong><\/p>\n\n\n\n

\u66f4\u53ef\u80fd\u662f\uff1a<\/p>\n\n\n\n

\n

GPU \/ XPU + HBM + CMOS + Photonic Matrix Engine + Optical I\/O + CPO + Optical Fabric<\/strong><\/p>\n<\/blockquote>\n\n\n\n

\u9010\u6e10\u878d\u5408\u6210\u4e00\u4e2a\u65b0\u7684\u5f02\u6784 AI \u8ba1\u7b97\u7cfb\u7edf\u3002<\/p>\n\n\n\n

\u5176\u4e2d\u54ea\u4e9b\u529f\u80fd\u8fc1\u79fb\u5230\u5149\u57df\uff0c\u5c06\u53d6\u51b3\u4e8e\u5149\u5728\u54ea\u4e9b\u73af\u8282\u771f\u6b63\u5177\u6709\u7cfb\u7edf\u7ea7\u4f18\u52bf\u3002<\/p>\n\n\n\n

\u5927\u81f4\u6f14\u8fdb\u987a\u5e8f\u53ef\u80fd\u662f\uff1a<\/p>\n\n\n\n

\u6570\u636e\u4e2d\u5fc3\u5149\u7f51\u7edc<\/strong><\/p>\n\n\n\n

\u2193<\/p>\n\n\n\n

\u673a\u67b6\u95f4\u5149\u4e92\u8fde<\/strong><\/p>\n\n\n\n

\u2193<\/p>\n\n\n\n

GPU \u95f4 Optical I\/O<\/strong><\/p>\n\n\n\n

\u2193<\/p>\n\n\n\n

Package \/ Chiplet \u5149\u4e92\u8fde<\/strong><\/p>\n\n\n\n

\u2193<\/p>\n\n\n\n

GPU\u2014Memory \u5149\u8fde\u63a5<\/strong><\/p>\n\n\n\n

\u2193<\/p>\n\n\n\n

\u90e8\u5206\u77e9\u9635\u8ba1\u7b97\u5149\u5316<\/strong><\/p>\n\n\n\n

\u2193<\/p>\n\n\n\n

\u7247\u4e0a\u8bad\u7ec3<\/strong><\/p>\n\n\n\n

\u2193<\/p>\n\n\n\n

\u539f\u4f4d\u7269\u7406\u5b66\u4e60<\/strong><\/p>\n\n\n\n

\u2193<\/p>\n\n\n\n

\u5149\u5b50\u5b58\u50a8\u4e0e\u611f\u5b58\u7b97\u4e00\u4f53<\/strong><\/p>\n\n\n\n

\u6309\u7167\u5f53\u524d\u4ea7\u4e1a\u6210\u719f\u5ea6\uff0c\u53ef\u4ee5\u7c97\u7565\u5224\u65ad\uff1a<\/p>\n\n\n\n

\u6280\u672f<\/th>2026 \u5e74\u6210\u719f\u5ea6<\/th>\u72b6\u6001<\/th><\/tr>
\u6570\u636e\u4e2d\u5fc3\u4f20\u7edf\u5149\u6a21\u5757<\/td>\u2605\u2605\u2605\u2605\u2605<\/td>\u5927\u89c4\u6a21\u6210\u719f<\/td><\/tr>
CPO \/ Optical I\/O<\/td>\u2605\u2605\u2605\u2605\u2606<\/td>\u6b63\u8fdb\u5165\u89c4\u6a21\u90e8\u7f72<\/td><\/tr>
GPU Scale-out \u5149\u7f51\u7edc<\/td>\u2605\u2605\u2605\u2605\u2606<\/td>NVIDIA \u7b49\u63a8\u8fdb\u4ea7\u4e1a\u5316<\/td><\/tr>
Photonic Chiplet \/ Interposer<\/td>\u2605\u2605\u2605\u2606\u2606<\/td>\u5feb\u901f\u53d1\u5c55<\/td><\/tr>
\u5149\u5b66\u77e9\u9635 AI \u52a0\u901f\u5668<\/td>\u2605\u2605\u2606\u2606\u2606\uff5e\u2605\u2605\u2605\u2606\u2606<\/td>\u5df2\u51fa\u73b0\u5546\u4e1a\u7cfb\u7edf<\/td><\/tr>
LLM \u5149\u8ba1\u7b97\u63a8\u7406<\/td>\u2605\u2605\u2606\u2606\u2606<\/td>\u5f00\u59cb\u8fdb\u5165\u8bc4\u4f30\u9636\u6bb5<\/td><\/tr>
\u5149\u5b50\u7247\u4e0a\u8bad\u7ec3<\/td>\u2605\u2605\u2606\u2606\u2606<\/td>\u5b9e\u9a8c\u9a8c\u8bc1<\/td><\/tr>
\u539f\u4f4d\u7269\u7406\u5b66\u4e60<\/td>\u2605\u2605\u2606\u2606\u2606<\/td>\u524d\u6cbf\u79d1\u7814\u7a81\u7834<\/td><\/tr>
\u5149\u5b50\u5b58\u7b97\u4e00\u4f53<\/td>\u2605\u2605\u2606\u2606\u2606<\/td>\u524d\u6cbf\u79d1\u7814\u9636\u6bb5<\/td><\/tr>
\u5168\u5149\u901a\u7528 AI \u5904\u7406\u5668<\/td>\u2605\u2606\u2606\u2606\u2606<\/td>\u957f\u671f\u76ee\u6807<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
\n\n\n\n

\u5341\u516b\u3001\u6700\u503c\u5f97\u5173\u6ce8\u7684\u53d8\u5316\uff1a\u5149\u5b50\u82af\u7247\u5f00\u59cb\u4ece\u201c\u8ba1\u7b97\u5668\u201d\u53d8\u6210\u201c\u5b66\u4e60\u673a\u5668\u201d<\/h1>\n\n\n\n

\u5982\u679c\u53ea\u770b\u4e09\u5e74\u524d\uff0c\u5149\u5b50 AI \u6700\u91cd\u8981\u7684\u95ee\u9898\u662f\uff1a<\/p>\n\n\n\n

\u5149\u80fd\u4e0d\u80fd\u9ad8\u6548\u505a\u77e9\u9635\u4e58\u6cd5\uff1f<\/strong><\/p>\n\n\n\n

\u4eca\u5929\u8fd9\u4e2a\u95ee\u9898\u5df2\u7ecf\u8d8a\u6765\u8d8a\u4e0d\u662f\u6838\u5fc3\u77db\u76fe\u3002<\/p>\n\n\n\n

\u73b0\u5728\u771f\u6b63\u91cd\u8981\u7684\u662f\uff1a<\/p>\n\n\n\n

\u5149\u8ba1\u7b97\u80fd\u4e0d\u80fd\u6210\u4e3a\u5b8c\u6574 AI \u7cfb\u7edf\u7684\u4e00\u90e8\u5206\uff1f<\/strong><\/p>\n\n\n\n

\u800c\u8fc7\u53bb\u4e24\u5e74\u7684\u6280\u672f\u8fdb\u5c55\u5df2\u7ecf\u9010\u6e10\u5f62\u6210\u4e86\u4e00\u6761\u5f88\u6e05\u6670\u7684\u8def\u5f84\uff1a<\/p>\n\n\n\n

Lightmatter\uff1a\u771f\u5b9e AI \u6a21\u578b\u80fd\u591f\u8fd0\u884c<\/strong><\/p>\n\n\n\n

\u2193<\/p>\n\n\n\n

Nokia Bell Labs\uff1a\u82af\u7247\u80fd\u591f\u53cd\u5411\u4f20\u64ad\u8bad\u7ec3<\/strong><\/p>\n\n\n\n

\u2193<\/p>\n\n\n\n

INSPIRE\uff1a\u771f\u5b9e\u7269\u7406\u82af\u7247\u80fd\u591f\u76f4\u63a5\u83b7\u5f97\u68af\u5ea6\u5e76\u539f\u4f4d\u5b66\u4e60<\/strong><\/p>\n\n\n\n

\u2193<\/p>\n\n\n\n

FLARE\uff1a\u5149\u5b50\u795e\u7ecf\u5143\u5f00\u59cb\u5177\u6709\u8bb0\u5fc6<\/strong><\/p>\n\n\n\n

\u8fd9\u6761\u8def\u7ebf\u6bd4\u5355\u7eaf\u5237\u65b0\u4e00\u4e2a TOPS\/W \u6570\u5b57\u66f4\u52a0\u91cd\u8981\u3002<\/p>\n\n\n\n

\u4e0e\u6b64\u540c\u65f6\uff0c\u4ea7\u4e1a\u7aef\u53c8\u5f62\u6210\u53e6\u4e00\u6761\u66f4\u52a0\u73b0\u5b9e\u7684\u8def\u7ebf\uff1a<\/p>\n\n\n\n

\u5149\u6a21\u5757<\/strong><\/p>\n\n\n\n

\u2193<\/p>\n\n\n\n

CPO<\/strong><\/p>\n\n\n\n

\u2193<\/p>\n\n\n\n

Optical I\/O<\/strong><\/p>\n\n\n\n

\u2193<\/p>\n\n\n\n

Photonic Interposer<\/strong><\/p>\n\n\n\n

\u2193<\/p>\n\n\n\n

GPU \/ Memory \u5149\u4e92\u8fde<\/strong><\/p>\n\n\n\n

\u2193<\/p>\n\n\n\n

\u5149\u7535\u6df7\u5408\u8ba1\u7b97<\/strong><\/p>\n\n\n\n

\u4e24\u6761\u8def\u7ebf\u6700\u7ec8\u5f88\u53ef\u80fd\u6c47\u5408\u3002<\/p>\n\n\n\n

\u6240\u4ee5\uff0c2026 \u5e74\u89c2\u5bdf\u5149\u5b50 AI\uff0c\u6700\u503c\u5f97\u5173\u6ce8\u7684\u5df2\u7ecf\u4e0d\u662f\uff1a<\/p>\n\n\n\n

\n

\u201c\u5149\u8ba1\u7b97\u4f1a\u4e0d\u4f1a\u66ff\u4ee3 GPU\uff1f\u201d<\/strong><\/p>\n<\/blockquote>\n\n\n\n

\u800c\u662f\uff1a<\/p>\n\n\n\n

\n

\u201cAI \u7cfb\u7edf\u4e2d\u7a76\u7adf\u6709\u591a\u5c11\u8ba1\u7b97\u3001\u901a\u4fe1\u3001\u5b58\u50a8\u548c\u5b66\u4e60\u8fc7\u7a0b\uff0c\u4f1a\u9010\u6b65\u8fc1\u79fb\u5230\u5149\u57df\uff1f\u201d<\/strong><\/p>\n<\/blockquote>\n\n\n\n

\u4ece NVIDIA \u7684 CPO\u3001Marvell \u6536\u8d2d Celestial AI\uff0c\u5230 Ayar Labs\u3001Lightmatter \u7684 Optical I\/O\uff0c\u518d\u5230 Lightelligence\u3001Lumai\u3001Q.ANT \u7684\u5149\u8ba1\u7b97\u7cfb\u7edf\uff0c\u4ee5\u53ca Nokia\u3001FLARE \u548c\u6700\u65b0 INSPIRE \u7684\u7247\u4e0a\u8bad\u7ec3\u3001\u8bb0\u5fc6\u4e0e\u539f\u4f4d\u5b66\u4e60\uff0c\u7b54\u6848\u5df2\u7ecf\u8d8a\u6765\u8d8a\u6e05\u695a\uff1a<\/p>\n\n\n\n

\u5149\u5b50\u6280\u672f\u6b63\u5728\u4ece AI \u7cfb\u7edf\u5916\u56f4\u7684\u901a\u4fe1\u6280\u672f\uff0c\u9010\u6e10\u6df1\u5165 AI \u8ba1\u7b97\u67b6\u6784\u5185\u90e8\u3002<\/strong><\/p>\n\n\n\n

\u5b83\u5927\u6982\u7387\u4e0d\u4f1a\u53d6\u4ee3\u7535\u5b50\u8ba1\u7b97\u3002<\/p>\n\n\n\n

\u771f\u6b63\u53ef\u80fd\u53d1\u751f\u7684\u662f\uff1a<\/p>\n\n\n\n

\n

\u7535\u5b50\u4e0e\u5149\u5b50\u7684\u8fb9\u754c\u4e0d\u65ad\u5411\u82af\u7247\u5185\u90e8\u79fb\u52a8\u3002<\/strong><\/p>\n<\/blockquote>\n\n\n\n

\u800c\u4e00\u65e6\u5149\u771f\u6b63\u540c\u65f6\u8fdb\u5165 Interconnect\u3001Compute\u3001Memory \u548c Learning<\/strong> \u56db\u4e2a\u5c42\u9762\uff0c\u6240\u8c13\u201c\u5149\u5b50 AI \u82af\u7247\u201d\u624d\u4f1a\u4ece\u4e00\u79cd\u7279\u6b8a\u52a0\u901f\u5668\uff0c\u9010\u6e10\u6f14\u5316\u6210\u4e00\u79cd\u65b0\u7684 AI \u8ba1\u7b97\u4f53\u7cfb\u7ed3\u6784\u3002<\/p>\n\n\n\n

\n\n\n\n

\u53c2\u8003\u8d44\u6599<\/h1>\n\n\n\n
    \n
  1. Photonic accelerators for AI across cloud and edge platforms<\/strong> \u2014 Nature Photonics<\/em>, 2026
    Nature Photonics\uff1aPhotonic accelerators for AI across cloud and edge platforms<\/a><\/li>\n\n\n\n
  2. All-analog photoelectronic chip for high-speed vision tasks<\/strong> \u2014 Nature<\/em>, 2023
    Nature\uff1aACCEL \u5149\u7535\u5168\u6a21\u62df\u8ba1\u7b97\u82af\u7247<\/a><\/li>\n\n\n\n
  3. Large-scale photonic chiplet Taichi empowers 160-TOPS\/W artificial general intelligence<\/strong> \u2014 Science<\/em>, 2024
    PubMed\uff1aTaichi \u5149\u5b50 Chiplet<\/a><\/li>\n\n\n\n
  4. Universal photonic artificial intelligence acceleration<\/strong> \u2014 Nature<\/em>, 2025
    Nature\uff1aUniversal photonic artificial intelligence acceleration<\/a><\/li>\n\n\n\n
  5. Integrated photonic neural network with on-chip backpropagation training<\/strong> \u2014 Nature<\/em>, 2026
    Nature\uff1aIntegrated photonic neural network with on-chip backpropagation training<\/a><\/li>\n\n\n\n
  6. Photonic neuromorphic learning via generalized in situ physical gradient descent<\/strong> \u2014 Nature Computational Science<\/em>, 2026
    Nature Computational Science\uff1aINSPIRE \u539f\u4f4d\u7269\u7406\u68af\u5ea6\u4e0b\u964d<\/a><\/li>\n\n\n\n
  7. Photonic Mixture-of-Experts for scalable multi-task on-chip optical neural networks<\/strong> \u2014 Nature Communications<\/em>, 2026
    Nature Communications\uff1aPhotonic Mixture-of-Experts<\/a><\/li>\n\n\n\n
  8. Programmable Three-dimensional Photonic Neural Network Chip<\/strong> \u2014 Nature Communications<\/em>, 2026
    Nature Communications\uff1a3D Photonic Neural Network Chip<\/a><\/li>\n\n\n\n
  9. Fully in-memory photonic computing<\/strong> \u2014 Nature Sensors<\/em>, 2026
    Nature Sensors\uff1aFully in-memory photonic computing \/ FLARE<\/a><\/li>\n\n\n\n
  10. NVIDIA Vera Rubin Ramps Into Full Production to Power Agentic AI Factories Worldwide<\/strong> \u2014 NVIDIA, 2026
    NVIDIA\uff1aVera Rubin \u4e0e Spectrum-X Ethernet Photonics<\/a><\/li>\n\n\n\n
  11. Marvell Completes Acquisition of Celestial AI<\/strong> \u2014 Marvell, 2026
    Marvell\uff1a\u5b8c\u6210\u6536\u8d2d Celestial AI<\/a><\/li>\n\n\n\n
  12. Lightmatter Vision \/ Photonic Computing and Interconnect Roadmap<\/strong> \u2014 Lightmatter
    Lightmatter\uff1aPhotonic Computing \u4e0e Interconnect \u8def\u7ebf\u56fe<\/a><\/li>\n\n\n\n
  13. Lightmatter Expands Photonic Interconnect Roadmap with Passage L20 Unified Optical Engine<\/strong> \u2014 Lightmatter, 2026
    Lightmatter\uff1aPassage L20<\/a><\/li>\n\n\n\n
  14. Industry Leaders Formally Launch CPO System Architecture Initiative within the Open Compute Project<\/strong> \u2014 Lightmatter, 2026
    Lightmatter\uff1aOpen Silicon Photonics for AI Systems<\/a><\/li>\n\n\n\n
  15. Ayar Labs Expands 2026 Funding to $650 Million to Scale Manufacturing-Ready CPO for AI Scale-Up Beyond the Rack<\/strong> \u2014 Ayar Labs, 2026
    Ayar Labs\uff1a2026 \u5e74\u878d\u8d44\u4e0e CPO \u91cf\u4ea7\u8ba1\u5212<\/a><\/li>\n\n\n\n
  16. TeraPHY Optical I\/O Engine<\/strong> \u2014 Ayar Labs
    Ayar Labs\uff1aTeraPHY Optical I\/O<\/a><\/li>\n\n\n\n
  17. PACE 2 Optoelectronic Accelerated Computing Card<\/strong> \u2014 Lightelligence
    Lightelligence\uff1aPACE 2 \u5149\u7535\u8ba1\u7b97\u52a0\u901f\u5361<\/a><\/li>\n\n\n\n
  18. Lumai Launches the World\u2019s First Optical Computing System for Real-Time, Billion-Parameter LLM Inference<\/strong> \u2014 Lumai, 2026
    Lumai\uff1aIris \u5149\u8ba1\u7b97 LLM \u63a8\u7406\u670d\u52a1\u5668<\/a><\/li>\n\n\n\n
  19. The Photonic AI Accelerator \/ Native Processing Server<\/strong> \u2014 Q.ANT
    Q.ANT\uff1aNative Processing Server \u4e0e\u5149\u5b50 NPU<\/a><\/li>\n\n\n\n
  20. Q.ANT Runs Generative AI on Photonic Hardware<\/strong> \u2014 Q.ANT, 2026
    Q.ANT\uff1a\u6269\u6563\u6a21\u578b\u548c xLSTM \u8fd0\u884c\u4e8e\u5149\u5b50\u786c\u4ef6<\/a><\/li>\n\n\n\n
  21. Volantis raises $88 million for technology to connect AI memory chips<\/strong> \u2014 Reuters<\/em>, 2026
    Reuters\uff1aVolantis GPU\u2014Memory \u5149\u4e92\u8fde<\/a><\/li>\n\n\n\n
  22. Intel Silicon Photonics \/ Optical Compute Interconnect<\/strong> \u2014 Intel
    Intel\uff1aSilicon Photonics \u4e0e OCI<\/a><\/li>\n\n\n\n
  23. Intel Demonstrates First Fully Integrated Optical I\/O Chiplet<\/strong> \u2014 Intel
    Intel\uff1a4 Tbps Optical Compute Interconnect Chiplet<\/a><\/li>\n<\/ol>\n","protected":false},"excerpt":{"rendered":"

    \u622a\u81f3 2026 \u5e74 10 \u6708 6 \u65e5\uff0c\u5149\u5b50 AI \u82af\u7247\u6b63\u5728\u7ecf\u5386\u4e00\u4e2a\u975e\u5e38\u5173\u952e\u7684\u9636\u6bb5\u53d8\u5316\u3002 \u8fc7\u53bb\uff0c\u4eba\u4eec\u8ba8\u8bba\u5149\u8ba1\u7b97\uff0c […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"default","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","ast-disable-related-posts":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"set","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[8,36],"tags":[39,75],"class_list":["post-7385","post","type-post","status-publish","format-standard","hentry","category-tech","category-36","tag-ai","tag-ic"],"rttpg_featured_image_url":null,"rttpg_author":{"display_name":"NullThought","author_link":"https:\/\/nullthought.net\/?author=1"},"rttpg_comment":0,"rttpg_category":"Tech<\/a> \u79d1\u5b66<\/a>","rttpg_excerpt":"\u622a\u81f3 2026 \u5e74 10 \u6708 6 \u65e5\uff0c\u5149\u5b50 AI \u82af\u7247\u6b63\u5728\u7ecf\u5386\u4e00\u4e2a\u975e\u5e38\u5173\u952e\u7684\u9636\u6bb5\u53d8\u5316\u3002 \u8fc7\u53bb\uff0c\u4eba\u4eec\u8ba8\u8bba\u5149\u8ba1\u7b97\uff0c…","_links":{"self":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/7385","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=7385"}],"version-history":[{"count":2,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/7385\/revisions"}],"predecessor-version":[{"id":7389,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/7385\/revisions\/7389"}],"wp:attachment":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=7385"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=7385"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=7385"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}