{"id":6419,"date":"2025-10-31T15:40:47","date_gmt":"2025-10-31T07:40:47","guid":{"rendered":"https:\/\/www.killtest.hk\/?p=6419"},"modified":"2025-10-31T15:40:47","modified_gmt":"2025-10-31T07:40:47","slug":"2025%e8%8b%b1%e5%81%89%e9%81%94gpu%e6%9e%b6%e6%a7%8b%e6%8a%80%e8%a1%93%e9%81%b8%e7%94%a8%e6%8c%87%e5%8d%97","status":"publish","type":"post","link":"https:\/\/www.killtest.hk\/index.php\/killtest\/2025%e8%8b%b1%e5%81%89%e9%81%94gpu%e6%9e%b6%e6%a7%8b%e6%8a%80%e8%a1%93%e9%81%b8%e7%94%a8%e6%8c%87%e5%8d%97\/","title":{"rendered":"2025\u82f1\u5049\u9054GPU\u67b6\u69cb\u6280\u8853\u9078\u7528\u6307\u5357"},"content":{"rendered":"\n

\u5c0d\u65bcAI\u4f3a\u670d\u5668\u8a2d\u8a08\u4eba\u54e1\u800c\u8a00\uff0c\u9078\u64c7\u5408\u9069\u7684GPU\u662f\u6c7a\u5b9a\u4fc2\u7d71\u6548\u80fd\u3001\u80fd\u6e90\u6548\u7387\u548c\u7e3d\u64c1\u6709\u6210\u672c\u7684\u95dc\u9375\u3002
\u82f1\u5049\u9054<\/a>\u4f5c\u70baAI\u904b\u7b97\u9818\u57df\u7684\u9818\u5c0e\u8005\uff0c\u5176GPU\u67b6\u69cb\u7684\u6bcf\u4e00\u6b21\u8fed\u4ee3\u90fd\u6a19\u8a8c\u8457\u904b\u7b97\u80fd\u529b\u7684\u98db\u8e8d\u3002
\u7406\u89e3\u5f9eTesla\u5230Hopper\u7684\u67b6\u69cb\u6f14\u9032\uff0c\u4e0d\u50c5\u95dc\u4e4e\u8b58\u5225\u5cf0\u503c\u7b97\u529b\uff0c\u66f4\u95dc\u4e4e\u65bc\u7406\u89e3\u5176\u6838\u5fc3\u8a08\u7b97\u7bc4\u5f0f\uff08\u5982Tensor Core\u7684\u5f15\u5165\uff09\u3001\u5167\u5b58\u5b50\u7cfb\u7d71\u8207\u4e92\u806f\u6280\u8853\u7684\u9032\u6b65\u3002
\u672c\u6587\u5c07\u4ee5\u904b\u7b97\u67b6\u69cb\u70ba\u4e2d\u5fc3\uff0c\u6574\u7406\u5176\u6f14\u9032\u8def\u5f91\u3001\u95dc\u9375\u7522\u54c1\u898f\u683c\uff0c\u4e26\u70ba\u9078\u578b\u63d0\u4f9b\u7cfb\u7d71\u6027\u5efa\u8b70\u3002
\u4e00\u3001\u904b\u7b97\u67b6\u69cb\u7684\u5960\u57fa\u8207\u840c\u82bd\u671f
1. Tesla\u67b6\u69cb (2006-2009)
\u5b9a\u4f4d\u8207\u7279\u6027\uff1a \u9019\u662f\u82f1\u5049\u9054GPGPU\u7684\u8d77\u9ede\u3002
Tesla\u67b6\u69cb\u9996\u6b21\u5f15\u9032\u4e86\u7d71\u4e00\u7684\u8457\u8272\u5668\u67b6\u69cb\uff0c\u53ef\u900f\u904eCUDA\uff082006\u5e74\u767c\u5e03\uff09\u9032\u884c\u901a\u7528\u5e73\u884c\u904b\u7b97\u3002
\u6b64\u6642\u5c1a\u7121\u5c08\u7528\u7684AI\u8a08\u7b97\u55ae\u5143\uff0c\u6240\u6709\u8a08\u7b97\u5747\u5728CUDA Core\u4e0a\u5b8c\u6210\u3002
\u4ee3\u8868\u7522\u54c1\uff1a\u00a0Tesla C1060 \/ T10
\u95dc\u9375\u898f\u683c\uff1a
\u6d41\u8655\u7406\u5668\uff08CUDA Core\uff09\uff1a 240\u500b
\u55ae\u7cbe\u5ea6\u6d6e\u7b97\u529b\uff08FP32\uff09\uff1a 933 GFLOPs
\u8a18\u61b6\u9ad4\uff1a 4GB GDDR3
\u4e92\u806f\uff1a PCIe 2.0
\u9078\u578b\u610f\u7fa9\uff1a \u6b77\u53f2\u610f\u7fa9\u5927\u65bc\u5be6\u7528\u610f\u7fa9\uff0c\u6a19\u8a8c\u8457GPU\u53ef\u7528\u65bc\u8a08\u7b97\u3002
2. Fermi\u67b6\u69cb (2010-2012)
\u5b9a\u4f4d\u8207\u7279\u6027\uff1a \u9996\u500b\u5b8c\u6574\u7684GPU\u904b\u7b97\u67b6\u69cb\u3002
\u5f15\u5165\u4e86\u771f\u6b63\u7684\u5feb\u53d6\u5c64\u6b21\u7d50\u69cb\uff08L1\/L2 Cache\uff09\uff0c\u652f\u63f4ECC\u986f\u5b58\uff0c\u63d0\u5347\u4e86\u96d9\u7cbe\u5ea6\uff08FP64\uff09\u6548\u80fd\u3002\u5b83\u662f\u771f\u6b63\u70ba\u8cc7\u6599\u4e2d\u5fc3\u8a2d\u8a08\u7684\u67b6\u69cb\u96db\u5f62\u3002
\u4ee3\u8868\u7522\u54c1\uff1a\u00a0Tesla M2090<\/p>\n\n\n\n

\u95dc\u9375\u898f\u683c\uff1a
CUDA Core\uff1a 512\u500b
FP32\u7b97\u529b\uff1a 1.33 TFLOPsFP64
\u7b97\u529b\uff1a 665 GFLOPs \uff081:2 FP64\/FP32\u6bd4\u7387\uff09
\u8a18\u61b6\u9ad4\uff1a 6GB GDDR5 with ECC\uff081:2 FP64\/FP32\u6bd4\u7387\uff09
\u8a18\u61b6\u9ad4\uff1a 6GB GDDR5 with ECC
\u4e92\u806f\uff1a PCIe 2.002.002.002 \u8cc7\u683c\u8a08\u7b97\u548c\u57fa\u790e\u79d1\u5b78\u3002
\u4e8c\u3001\u73fe\u4ee3AI\u8a08\u7b97\u7684\u555f\u8499\u8207\u7206\u767c\u671f
3. Kepler\u67b6\u69cb (2012-2014)
\u5b9a\u4f4d\u8207\u7279\u6027\uff1a \u8ffd\u6c42\u6027\u80fd\u8207\u529f\u8017\u7684\u5e73\u8861\u3002
\u5f15\u9032\u4e86GPUDirect\u6280\u8853\uff0c\u964d\u4f4e\u4e86GPU\u8207\u7b2c\u4e09\u65b9\u8a2d\u5099\uff08\u5982\u7db2\u8def\u5361\uff09\u7684\u901a\u8a0a\u5ef6\u9072\u3002
\u96d6\u7136\u4ecd\u7121\u5c08\u7528AI\u6838\u5fc3\uff0c\u4f46\u5176\u5f37\u5927\u7684FP32\u6027\u80fd\u70ba\u65e9\u671f\u6df1\u5ea6\u5b78\u7fd2\u7814\u7a76\u63d0\u4f9b\u4e86\u53ef\u80fd\u3002
\u4ee3\u8868\u7522\u54c1\uff1a\u00a0Tesla K80\u00a0\uff08\u96d9\u82af\u986f\u793a\u5361\uff09
\u95dc\u9375\u898f\u683c\uff08\u6bcf\u9846GPU\uff09\uff1a
CUDA Core\uff1a 2496\u500b
FP32\u7b97\u529b\uff1a 2.91 TFLOPs
\u8a18\u61b6\u9ad4\uff1a 12GB GDDR5 \uff08\u6bcfGPU\uff0c\u95c6\u5361\u517124GB\uff09
\u4e92\u806f\uff1a PCIe 3.0
\u9078\u578b\u610f\u7fa9\uff1a\u8a31\u591a\u65e9\u671fAI\u6a21\u578b\uff08\u5982AlexNet\uff09\u5728Kepler\u53ca\u985e\u4f3c\u67b6\u69cb\u7684GPU\u4e0a\u88ab\u8b49\u660e\u662f\u53ef\u884c\u7684\uff0c\u958b\u555f\u4e86AI\u7684\u300c\u66b4\u529b\u8a08\u7b97\u300d\u6642\u4ee3\u3002
4. Maxwell\u67b6\u69cb (2014-2016)
\u5b9a\u4f4d\u8207\u7279\u6027\uff1a \u6975\u81f4\u80fd\u6548\u6700\u4f73\u5316\u3002
\u900f\u904e\u512a\u5316\u8abf\u5ea6\u5668\u548c\u5feb\u53d6\u7d50\u69cb\uff0c\u5728\u76f8\u540c\u7684\u529f\u8017\u4e0b\u63d0\u4f9b\u4e86\u6bd4Kepler\u66f4\u9ad8\u7684\u6548\u80fd\u3002
\u70ba\u5f8c\u7e8c\u67b6\u69cb\u7684\u80fd\u6e90\u6548\u7387\u8a2d\u8a08\u92ea\u5e73\u4e86\u9053\u8def\u3002
\u4ee3\u8868\u7522\u54c1\uff1a\u00a0Tesla M40
\u95dc\u9375\u898f\u683c\uff1a
CUDA Core\uff1a 3072\u500b
FP32\u904b\u7b97\u529b\uff1a 7 TFLOPs
\u8a18\u61b6\u9ad4\uff1a 12GB \/ 24GB GDDR5
\u4e92\u806f\uff1a PCIe 3.0
\u9078\u578b\u610f\u7fa9\uff1a M40\u5f8c\u671f\u88ab\u5ee3\u6cdb\u7528\u65bcAI\u63a8\u7406\uff0c\u56e0\u5176\u5728INT8\u7cbe\u5ea6\u4e0b\u5177\u6709\u826f\u597d\u7684\u80fd\u6548\u3002\u4e09\u3001AI\u904b\u7b97\u7684\u5c08\u696d\u5316\u8207\u5178\u7bc4\u78ba\u7acb\u671f5. Pascal\u67b6\u69cb (2016-2017) \uff1a
AI\u904b\u7b97\u7684\u7b2c\u4e00\u6b21\u98db\u8e8d\u5b9a\u4f4d\u8207\u7279\u6027\uff1aNVLink 1.0\uff1a \u9996\u6b21\u5f15\u9032\u9ad8\u901fGPU\u4e92\u806f\u6280\u8853\uff0c\u5927\u5e45\u63d0\u5347\u591aGPU\u9593\u983b\u5bec\u3002
HBM2\uff1a \u5728\u65d7\u8266\u5361\u4e0a\u5f15\u5165\u9ad8\u983b\u5bec\u5167\u5b58\uff0c\u89e3\u6c7a\u5167\u5b58\u7246\u554f\u984c\u3002
16nm FinFET\u88fd\u7a0b\uff1a \u5e36\u4f86\u5de8\u5927\u7684\u80fd\u6e90\u6548\u7387\u63d0\u5347\u3002
\u4ee3\u8868\u7522\u54c1\uff1a\u00a0Tesla P100\u00a0\uff08\u6709NVLink\u548cPCIe\u5169\u500b\u7248\u672c\uff09
\u95dc\u9375\u898f\u683c\uff1a
CUDA Core\uff1a 3584\u500b
FP32\u7b97\u529b\uff1a 10.6 TFLOPsFP16
\u7b97\u529b\uff1a21.2 TFLOPs \uff08\u7121\u5c08\u7528\u55ae\u5143\uff0c\u50c5\u4f7f\u7528NV \uff08160 GB\/s\uff09 \/ PCIe 3.0
\u9078\u578b\u610f\u7fa9\uff1a P100\u662f\u9996\u6b3e\u70baAI\u548cHPC\u91cf\u8eab\u5b9a\u5236\u7684\u8cc7\u6599\u4e2d\u5fc3GPU\uff0c\u5176NVLink\u548cHBM2\u5960\u5b9a\u4e86\u73fe\u4ee3AI\u4f3a\u670d\u5668\u7684\u57fa\u672c\u5f62\u614b\u3002
6. Volta\u67b6\u69cb (2017-2020) \uff1a
\u9769\u547d\u6027\u7684Tensor Core\u5230\u4f86
\u5b9a\u4f4d\u8207\u7279\u6027\uff1a
Tensor Core\uff1a \u9769\u547d\u6027\u5275\u65b0\uff01\u5c08\u70ba\u77e9\u9663\u904b\u7b97\u8a2d\u8a08\u7684\u6838\u5fc3\uff0c\u652f\u63f4\u6df7\u5408\u7cbe\u6e96\u5ea6\u8a13\u7df4\uff0c\u5927\u5e45\u63d0\u5347\u4e86FP16\u548cINT8\u7684\u7b97\u529b\u3002
NVLink 2.0\uff1a \u983b\u5bec\u7ffb\u500d\u3002
HBM2\uff1a \u5bb9\u91cf\u548c\u983b\u5bec\u7e7c\u7e8c\u63d0\u5347\u3002\u9996\u6b21\u6574\u5408NVSwitch\u65bcDGX-2\u4e2d\uff0c\u5be6\u73fe\u591aGPU\u5168\u4e92\u806f\u3002
\u4ee3\u8868\u7522\u54c1\uff1a\u00a0Tesla V100\u00a0\uff08SXM2\u548cPCIe\u5f62\u614b\uff09
\u95dc\u9375\u898f\u683c\uff1a
CUDA Core\uff1a 5120\u500b
\u5f35\u91cf\u6838\u5fc3\uff1a 640\u500b
FP32\u7b97\u529b\uff1a 15.7 TFLOPsFP16550\u500b
\u5f35\u91cf\u6838\u5fc3\uff1a 640\u500bFP32
\u904b\u7b97\u529b\uff1a 15.7 TFLOPsFP165056250 TGB~ HBM2
\u8a18\u61b6\u9ad4\u983b\u5bec\uff1a 900 GB\/s\u4e92\u806f\uff1a NVLink 2.0 \uff08300 GB\/s\uff09
\u9078\u578b\u610f\u7fa9\uff1a\u00a0V100\u662fAI\u8a13\u7df4\u9818\u57df\u7684\u91cc\u7a0b\u7891\u3002
\u5176Tensor Core\u4f7f\u5f97\u8a13\u7df4\u5927\u578b\u6a21\u578b\u5f9e\u201c\u53ef\u80fd\u201d\u8b8a\u70ba\u201c\u9ad8\u6548\u80fd\u201d\uff0c\u81f3\u4eca\u4ecd\u5728\u8a31\u591a\u8cc7\u6599\u4e2d\u5fc3\u670d\u5f79\u3002
\u9078\u64c7V100\u610f\u5473\u8457\u64c1\u62b1\u4e86\u6210\u719f\u7684AI\u904b\u7b97\u7bc4\u5f0f\u3002
\u56db\u3001AI\u8a08\u7b97\u7684\u898f\u6a21\u5316\u8207\u666e\u53ca\u671f7. Turing\u67b6\u69cb (2018-2020) \uff1a
\u63a8\u7406\u7684\u9769\u65b0\u5b9a\u4f4d\u8207\u7279\u6027\uff1a
\u5f35\u91cf\u6838\u5fc3\u5347\u7d1a\uff1a \u5f15\u5165INT4\u548cINT1\u7cbe\u5ea6\u652f\u6301\uff0c\u4e26\u589e\u5f37INT8\/FP16\u6027\u80fd\u3002
RT Core\uff1a \u5c08\u6ce8\u65bc\u5149\u8ffd\uff0c\u5c0dAI\u4f3a\u670d\u5668\u9078\u578b\u610f\u7fa9\u4e0d\u5927\u3002
\u96d6\u7136\u4e3b\u6253\u6d88\u8cbb\u5e02\u5834\uff0c\u4f46\u5176\u5c08\u696d\u5361\u5728\u63a8\u7406\u5834\u666f\u8868\u73fe\u51fa\u8272\u3002
\u4ee3\u8868\u7522\u54c1\uff1a\u00a0Tesla T4\u00a0\uff08\u4f4e\u529f\u8017\u63a8\u7406\u5361\uff09
\u95dc\u9375\u898f\u683c\uff1a
CUDA Core\uff1a 2560\u4e2a<\/p>\n\n\n\n

\u5f20\u91cf\u6838\u5fc3\uff1a 320\u4e2a<\/p>\n\n\n\n

FP32\u7b97\u529b\uff1a 8.1 TFLOPs<\/p>\n\n\n\n

INT8\u7b97\u529b\uff08Tensor Core\uff09<\/strong>\uff1a 130 TOPS<\/strong><\/p>\n\n\n\n

INT4\u7b97\u529b\uff08Tensor Core\uff09<\/strong>\uff1a 260 TOPS<\/strong><\/p>\n\n\n\n

\u5185\u5b58\uff1a 16GB GDDR6<\/p>\n\n\n\n

\u529f\u8017\uff1a \u4ec570W
\u9078\u578b\u610f\u7fa9\uff1a\u00a0T4\u662f\u908a\u7de3\u63a8\u7406\u548c\u96f2\u4e0a\u63a8\u7406\u7684\u6a19\u7aff\u7522\u54c1\u3002\u5176\u6975\u4f73\u7684\u80fd\u6e90\u6548\u7387\u6bd4\u975e\u5e38\u9069\u5408\u9ad8\u5bc6\u5ea6\u3001\u4f4e\u5ef6\u9072\u7684\u63a8\u7406\u4f3a\u670d\u5668\u3002
8. Ampere\u67b6\u69cb (2020-2022) \uff1a
\u901a\u7528\u6027\u8207AI\u6027\u80fd\u7684\u5b8c\u7f8e\u7d50\u5408\u5b9a\u4f4d\u8207\u7279\u6027\uff1a
\u7b2c\u4e09\u4ee3Tensor Core\uff1a \u652f\u63f4TF32\u3001FP64 Tensor Core\uff0c\u4ee5\u53ca\u7a00\u758f\u5316\u52a0\u901f\uff0cAI\u6027\u80fd\u66b4\u589e\u3002\u7d50\u69cb\u6027\u7a00\u758f\uff1a \u5229\u75282:4\u7684\u7a00\u758f\u6a21\u5f0f\uff0c\u7406\u8ad6\u4e0a\u8b93\u7a00\u758f\u6a21\u578b\u7684\u7b97\u529b\u589e\u52a0\u4e00\u500d\u3002 NVLink 3.0\uff1a \u983b\u5bec\u518d\u6b21\u5927\u5e45\u63d0\u5347\u3002
\u591a\u91cd\u57f7\u884c\u500b\u9ad4GPU\uff08MIG\uff09\uff1a \u53ef\u5c07\u4e00\u584a\u5be6\u9ad4GPU\u5206\u5272\u70ba\u591a\u500b\u7368\u7acb\u3001\u5b89\u5168\u7684\u5be6\u4f8b\uff0c\u4e26\u63d0\u5347\u8cc7\u6e90\u4f7f\u7528\u7387\u3002
A100 40GB\/80GB\uff1a \u63a1\u7528HBM2e\u3002
\u4ee3\u8868\u7522\u54c1\uff1a\u00a0Tesla A100 40GB\/80GB PCIe & SXM4
\u95dc\u9375\u898f\u683c\uff08A100 80GB SXM\uff09\uff1a
\u5173\u952e\u89c4\u683c<\/strong>\uff08A100 80GB SXM\uff09\uff1a<\/p>\n\n\n\n