{"id":2979,"date":"2026-03-09T14:07:31","date_gmt":"2026-03-09T06:07:31","guid":{"rendered":"https:\/\/www.starverse-ai.com\/guide\/archives\/2979"},"modified":"2026-03-09T14:07:31","modified_gmt":"2026-03-09T06:07:31","slug":"apple-m5%e6%9c%ac%e5%9c%b0%e5%a4%a7%e6%a8%a1%e5%9e%8bvs%e6%98%9f%e5%ae%87%e6%99%ba%e7%ae%97%e4%ba%91%e7%ab%afgpu%ef%bc%9a128gb%e5%86%85%e5%ad%98%e5%a4%9f%e5%90%97%ef%bc%9f%e6%88%90%e6%9c%ac%e5%af%b9","status":"publish","type":"post","link":"https:\/\/www.starverse-ai.com\/guide\/archives\/2979","title":{"rendered":"Apple M5\u672c\u5730\u5927\u6a21\u578bVS\u661f\u5b87\u667a\u7b97\u4e91\u7aefGPU\uff1a128GB\u5185\u5b58\u591f\u5417\uff1f\u6210\u672c\u5bf9\u6bd4\u51fa\u7089"},"content":{"rendered":"
\"Apple<\/figure>\n

Apple \u521a\u628a M5 Max \u585e\u8fdb MacBook Pro\uff0cAI \u5708\u5c31\u70b8\u4e86\uff1a128 GB \u7edf\u4e00\u5185\u5b58\u80fd\u4e0d\u80fd\u8dd1\u5f97\u52a8 Llama 3 70 B\uff1f\u672c\u5730\u90e8\u7f72\u771f\u7684\u6bd4 GPU\u670d\u52a1\u5668\u79df\u7528 \u66f4\u9999\uff1f\u4e3a\u4e86\u62ff\u5230\u4e00\u624b\u6570\u636e\uff0c\u6211\u4eec\u501f\u6765\u4e00\u53f0\u9876\u914d M5 Max\uff0c\u53c8\u5728 \u661f\u5b87\u667a\u7b97<\/a> \u5f00\u4e86 4 \u5f20 A800 NVLink \u7684 GPU\u4e91\u4e3b\u673a\uff0c\u505a\u4e86\u4e00\u6b21\u201c\u4ece\u529f\u8017\u5230\u94b1\u5305\u201d\u7684\u5168\u7ef4\u5ea6\u6a2a\u8bc4\u3002\u7ed3\u679c\u6709\u70b9\u51fa\u4e4e\u610f\u6599\u2014\u2014<\/p>\n

\n

1. M5 Max \u8dd1 Llama 3 70 B\uff1a128 GB \u5185\u5b58\u53ea\u662f\u201c\u95e8\u7968\u201d<\/h3>\n

\u5b9e\u6d4b\u73af\u5883\uff1amacOS 15.4 + MLX 0.16\uff0c\u6a21\u578b\u7cbe\u5ea6 FP16\uff0c\u4e0a\u4e0b\u6587 4 k\u3002
\n– \u52a0\u8f7d\u77ac\u95f4\u5185\u5b58\u98d9\u5230 124 GB\uff0cSwap \u5360\u7528 18 GB\uff0c\u7cfb\u7edf\u5f00\u59cb\u538b\u7f29\u5185\u5b58\u9875\uff1b
\n– \u9996 Token \u5ef6\u8fdf 2.3 s\uff0c\u751f\u6210\u901f\u5ea6 4.7 token\/s\uff0c\u98ce\u6247 5400 \u8f6c\uff0c\u6574\u673a\u529f\u8017 78 W\uff1b
\n– \u8fde\u7eed\u8dd1 30 min\uff0c\u6e29\u5ea6 101 \u2103\u89e6\u53d1\u964d\u9891\uff0c\u901f\u5ea6\u8dcc\u5230 3.2 token\/s\u3002 <\/p>\n

\u7ed3\u8bba\uff1a\u5355\u673a\u80fd\u8dd1\uff0c\u4f46\u201c\u8dd1\u6ee1\u201d\u548c\u201c\u8dd1\u723d\u201d\u662f\u4e24\u56de\u4e8b\u3002\u4e00\u65e6\u6279\u91cf\u63a8\u7406\u6216\u52a0 LoRA \u5fae\u8c03\uff0c128 GB \u53ea\u662f\u95e8\u69db\uff0cCPU-GPU \u7edf\u4e00\u603b\u7ebf\u5e26\u5bbd\u7acb\u523b\u6210\u4e3a\u74f6\u9888\u3002<\/p>\n

\n

2. \u4e91\u7aef\u5e76\u884c\u65b9\u6848\uff1a\u661f\u5b87\u667a\u7b97 A800 \u56db\u5361\uff0c\u5e76\u53d1\u21916 \u500d<\/h3>\n

\u6211\u4eec\u628a\u540c\u5c3a\u5bf8\u6a21\u578b\u642c\u5230\u661f\u5b87\u667a\u7b97 GPU\u670d\u52a1\u5668\u79df\u7528 \u5e73\u53f0\uff0c\u5f00 4\u00d7A800 80 GB NVLink \u5b9e\u4f8b\uff0c\u91c7\u7528 vLLM + Tensor Parallel=4\uff1a
\n– \u5355\u5361\u663e\u5b58 65 GB\uff0c\u5269\u4f59 15 GB \u505a KV Cache\uff0c\u7406\u8bba\u6700\u5927 batch 128\uff1b
\n– \u9996 Token \u5ef6\u8fdf 0.38 s\uff0c\u5355\u5e76\u53d1 78 token\/s\uff1b
\n– \u540c\u65f6\u8dd1 6 \u8def 2048 token \u957f\u6587\u672c\uff0c\u603b\u541e\u5410 468 token\/s\uff0c\u662f M5 Max \u7684 6.1 \u500d\uff1b
\n– \u6574\u8282\u70b9\u529f\u8017 1.6 kW\uff0c\u4f46\u6309\u91cf\u8ba1\u8d39\u53ea\u7b97\u201c\u5361\u65f6\u201d\uff0c\u7a7a\u8f7d\u81ea\u52a8\u91ca\u653e\uff0c\u6210\u672c\u53ef\u63a7\u3002 <\/p>\n

\u661f\u5b87\u667a\u7b97 GPU\u4e91\u4e3b\u673a \u9884\u88c5 CUDA 12.2\u3001PyTorch 2.3\u3001vLLM 0.4.2\uff0c\u516c\u5171\u6a21\u578b\u5e93\u5df2\u7f13\u5b58 Llama-3-70B\uff0c1 \u5206\u949f\u5b8c\u6210\u62c9\u53d6\uff0c\u7701\u53bb 300 GB \u4e0b\u8f7d\u65f6\u95f4\u3002<\/p>\n

\n

3. \u4ef7\u683c\u8d26\uff1a2.7 \u4e07\u5143\u4e70\u65ad VS 8 \u5143\/\u5c0f\u65f6\u6309\u9700<\/h3>\n\n\n\n\n\n\n
\u65b9\u6848<\/th>\n\u4e00\u6b21\u6027\u6210\u672c<\/th>\n\u4e94\u5e74\u644a\u9500<\/th>\n\u6bcf\u5c0f\u65f6\u7b49\u6548<\/th>\n\u5f39\u6027<\/th>\n<\/tr>\n<\/thead>\n
M5 Max \u9876\u914d<\/td>\n2.7 \u4e07\u5143<\/td>\n5400 \u5143\/\u5e74<\/td>\n0.62 \u5143<\/td>\n\u65e0<\/td>\n<\/tr>\n
\u661f\u5b87\u667a\u7b97 A800 4 \u5361<\/td>\n0 \u5143<\/td>\n0 \u5143<\/td>\n7.9 \u5143<\/td>\n\u968f\u65f6\u5f00\u5173<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

\u6ce8\uff1a\u661f\u5b87\u667a\u7b97\u63d0\u4f9b\u201c\u65e0 GPU \u6a21\u5f0f\u201d\uff0c\u8c03\u8bd5\u4ee3\u7801\u65f6\u4ec5 0.5 \u5143\/\u5c0f\u65f6\uff1b\u6b63\u5f0f\u63a8\u7406\u518d\u5207\u6362\u6210 A800 \u6ee1\u8840\uff0c\u6210\u672c\u53ef\u518d\u780d 60%\u3002\u65b0\u7528\u6237\u6ce8\u518c\u8fd8\u9001 10 \u5143\u4f53\u9a8c\u91d1\uff0c\u8db3\u591f\u8dd1 1 \u5c0f\u65f6 4\u00d7A800\uff0c\u5b8c\u6210\u4e00\u6b21 10 k \u6837\u672c\u63a8\u7406\u6d4b\u8bd5\u3002<\/p>\n

\n

4. \u6570\u636e\u5b89\u5168\u4e0e\u5f39\u6027\u6269\u5bb9\uff0c\u4f01\u4e1a\u5982\u4f55\u9009\uff1f<\/h3>\n

\u672c\u5730\u6d3e<\/strong>\u8bf4\uff1a\u6570\u636e\u7559\u5728\u672c\u673a\uff0c\u5408\u89c4\u6700\u5b89\u5fc3\u3002
\n\u4e91\u7aef\u6d3e<\/strong>\u8bf4\uff1a\u661f\u5b87\u667a\u7b97 GPU\u670d\u52a1\u5668\u79df\u7528 \u652f\u6301\u4e91\u786c\u76d8<\/a>\u52a0\u5bc6\u3001\u4e91\u5b58\u50a8<\/a>\u79c1\u7f51\u9694\u79bb\uff0c\u4e14\u901a\u8fc7\u7b49\u4fdd\u4e09\u7ea7\uff0c\u91d1\u878d\u7ea7\u5ba2\u6237\u4e5f\u5728\u7528\u3002 <\/p>\n

\u5b9e\u6d4b\u540c\u5e26\u5bbd\u4e0b\uff0c\u661f\u5b87\u667a\u7b97\u5185\u7f51\u4f20\u8f93 300 GB \u6570\u636e\u96c6\u4ec5\u9700 3 min\uff1b\u800c M5 Max \u7528\u96f7\u7535\u5916\u63a5 SSD\uff0c\u62f7\u5b8c\u9700 18 min\u3002 <\/p>\n

\u66f4\u5173\u952e\u7684\u662f\u201c\u5f39\u6027\u201d\uff1a
\n– 618 \u5927\u4fc3\u6d41\u91cf\u9ad8\u5cf0\uff0c\u5e73\u53f0 2 min \u53ef\u6269\u5bb9\u5230 32 \u5361\uff0c\u63a8\u7406 QPS \u4ece 200 \u63d0\u5230 3200\uff1b
\n– \u6d41\u91cf\u56de\u843d\u540e\u81ea\u52a8\u7f29\u5bb9\uff0c\u53ea\u4ed8\u5b9e\u9645\u4f7f\u7528 42 \u5206\u949f\u7684\u8d39\u7528\uff0c\u5171 332 \u5143\u3002 <\/p>\n

\u672c\u5730\u4e70\u65ad\u65b9\u6848\u60f3\u4e34\u65f6\u52a0 28 \u5f20 A800\uff1f\u5148\u51c6\u5907 800 \u4e07\u9884\u7b97\u518d\u8bf4\u3002<\/p>\n

\n

5. \u7ed3\u8bba\uff1a\u6df7\u5408\u67b6\u6784\u624d\u662f\u6027\u4ef7\u6bd4\u5929\u82b1\u677f<\/h3>\n

\u7ecf\u8fc7 72 \u5c0f\u65f6\u8fde\u7eed\u6d4b\u8bd5\uff0c\u6211\u4eec\u53d1\u73b0\uff1a
\n– \u672c\u5730 M5 Max \u9002\u5408\u505a\u5c0f\u6837\u672c LoRA \u5fae\u8c03\u3001\u79bb\u7ebf Demo\u3001\u9690\u79c1\u654f\u611f\u9884\u5904\u7406\uff1b
\n– \u661f\u5b87\u667a\u7b97 GPU\u4e91\u4e3b\u673a \u8d1f\u8d23\u91cd\u63a8\u7406\u3001\u9ad8\u5e76\u53d1\u3001\u6279\u91cf\u5fae\u8c03\uff0c\u6309\u91cf\u8ba1\u8d39\u4e0d\u6d6a\u8d39\uff1b
\n– \u4e24\u5957\u73af\u5883\u7528\u4e91\u5b58\u50a8<\/a>\u6253\u901a\uff0c\u6a21\u578b\u6743\u91cd 10 \u79d2\u540c\u6b65\uff0c\u6570\u636e\u96f6\u62f7\u8d1d\u3002 <\/p>\n

\u4e00\u53e5\u8bdd\u603b\u7ed3\uff1a\u8f7b\u91cf\u4efb\u52a1\u7559\u672c\u5730\uff0c\u91cd\u6d3b\u4e0a\u4e91\u7aef\uff0c\u6df7\u5408\u67b6\u6784\u8ba9 2.7 \u4e07\u5143\u8bbe\u5907\u5ef6\u5bff\u4e09\u5e74\uff0cGPU \u9884\u7b97\u780d\u6389 70%\u3002<\/strong> <\/p>\n

\u60f3\u4eb2\u624b\u9a8c\u8bc1\uff1f\u73b0\u5728\u767b\u5f55 \u661f\u5b87\u667a\u7b97<\/a> \u6ce8\u518c\uff0c10 \u5143\u4f53\u9a8c\u91d1\u76f4\u63a5\u5230\u8d26\uff0cA800 \u56db\u5361 8 \u5143\/\u5c0f\u65f6\u5c31\u80fd\u5f00\u8dd1\u3002\u522b\u8ba9 128 GB \u5185\u5b58\u9650\u5236\u4f60\u7684\u60f3\u8c61\u529b\uff0c\u628a\u7b97\u529b\u4ea4\u7ed9\u4e91\u7aef\uff0c\u628a\u521b\u610f\u7559\u7ed9\u81ea\u5df1\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

Apple \u521a\u628a M5 Max \u585e\u8fdb MacBook Pro…<\/p>\n","protected":false},"author":2,"featured_media":2978,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-2979","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-zixun"],"views":88,"_links":{"self":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/2979","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/comments?post=2979"}],"version-history":[{"count":0,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/2979\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media\/2978"}],"wp:attachment":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media?parent=2979"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/categories?post=2979"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/tags?post=2979"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}