{"id":1676,"date":"2026-11-13T17:10:51","date_gmt":"2026-11-13T08:10:51","guid":{"rendered":"https:\/\/www.hpc.co.jp\/ai-deeplearning\/?page_id=1676"},"modified":"2026-11-13T17:10:52","modified_gmt":"2026-11-13T08:10:52","slug":"rtx-pro-4500-vs-6000-max-q","status":"publish","type":"page","link":"https:\/\/www.hpc.co.jp\/ai-deeplearning\/rtx-pro-4500-vs-6000-max-q\/","title":{"rendered":"RTX PRO 4500 Blackwellvs. RTX PRO 6000 Blackwell Max-Q"},"content":{"rendered":"\n\n
\u751f\u6210\u901f\u5ea6<\/a>\u540c\u6642\u5229\u7528<\/a>\u8a55\u4fa1\u5b9f\u6a5f<\/a>\u5b9f\u6a5f\u306e\u76f8\u8ac7 \u2197<\/a><\/div><\/div><\/div>

<\/span>LOCAL LLM BENCHMARK REPORT<\/p>

\u751f\u6210\u901f\u5ea6\u306e\u5dee\u3068\u3001
\u540c\u6642\u5229\u7528<\/span>\u306e\u5dee\u3002<\/h1>

RTX PRO 4500 Blackwell
vs.<\/span>RTX PRO 6000 Blackwell Max-Q<\/p>

\u5358\u72ec\u306e\u751f\u6210\u901f\u5ea6\u306f\u8fd1\u304f\u3066\u3082\u3001\u5229\u7528\u304c\u91cd\u306a\u3063\u305f\u3068\u304d\u306e\u5f85\u3061\u6642\u9593\u306f\u3069\u3046\u5909\u308f\u308b\u304b\u3002Qwen3.8-27B\u3092\u7528\u3044\u305f\u5b9f\u6e2c\u304b\u3089\u3001KV\u30ad\u30e3\u30c3\u30b7\u30e5\u5bb9\u91cf\u3068\u904b\u7528\u306e\u9055\u3044\u3092\u6bd4\u8f03\u3057\u307e\u3059\u3002<\/p>

\u5b9f\u6e2c\u30ec\u30dd\u30fc\u30c8\u3092\u8aad\u3080 \u2193<\/a>\u8a55\u4fa1\u306b\u4f7f\u3063\u305f\u5b9f\u6a5f\u3092\u898b\u308b \u2193<\/a><\/p>

Qwen3.8-27B-NVFP4 \/ vLLM 0.32.0 \/ KV NVFP4<\/p><\/div>

16k\u5165\u529b\u30fb8\u4e26\u5217\u306e\u5b9f\u6e2c<\/p>

\u751f\u6210\u4e2d\u306e\u5dee\u306f\u7d047%\u3002
\u56de\u7b54\u958b\u59cb\u5f85\u3061\u306f\u7d047.8\u500d\u3002<\/h2>

\u56de\u7b54\u958b\u59cb\u307e\u3067\u306eTTFT p95
\u79d2\u30fb\u5c0f\u3055\u3044\u307b\u3069\u826f\u597d<\/span><\/p>

Pro 4500 \/ 32GB <\/span><\/p>

<\/div><\/div>

53.16\u79d2<\/span><\/p><\/div><\/div>

Pro 6000 Max-Q \/ 96GB <\/span><\/p>

<\/div><\/div>

6.79\u79d2<\/span><\/p><\/div><\/div>

\u751f\u6210\u4e2d\u306eTPOT p95\uff1a31.9 ms \/ 29.7 ms
Pro 4500 \/ Pro 6000 Max-Q\u306e\u9806<\/span><\/p><\/div><\/div><\/div><\/div><\/div>

REPORT OVERVIEW<\/p><\/div>

\u5b9f\u6e2c\u304b\u3089\u898b\u308b\u3001GPU\u306e\u9078\u3073\u65b9\u3002<\/h2>

Pro 4500\u306f\u3001\u5c11\u4eba\u6570\u3067\u4f7f\u3046\u751f\u6210\u901f\u5ea6\u3092\u5341\u5206\u306b\u78ba\u4fdd\u3067\u304d\u307e\u3059\u3002\u4e00\u65b9\u3001\u9577\u3044\u5165\u529b\u3084\u8907\u6570\u306e\u8981\u6c42\u304c\u91cd\u306a\u308b\u6761\u4ef6\u3067\u306f\u3001Pro 6000 Max-Q\u306e\u5927\u304d\u306aKV\u9818\u57df\u304c\u904b\u7528\u4e0a\u306e\u4f59\u88d5\u306b\u3064\u306a\u304c\u308a\u307e\u3059\u3002<\/p>

\u672c\u30ec\u30dd\u30fc\u30c8\u306f\u3001\u30aa\u30d5\u30a3\u30b9\u3078\u306e\u8a2d\u7f6e\u3092\u60f3\u5b9a\u3057\u305f2U\u30fb\u5965\u884c400mm\u306e\u30b3\u30f3\u30bb\u30d7\u30c8\u30e2\u30c7\u30eb\u3067\u5b9f\u65bd\u3057\u307e\u3057\u305f\u3002\u307e\u305aGPU\u306e\u6bd4\u8f03\u7d50\u679c\u3092\u793a\u3057\u3001\u5f8c\u534a\u3067\u8a55\u4fa1\u306b\u4f7f\u3063\u305f\u5b9f\u6a5f\u3092\u7d39\u4ecb\u3057\u307e\u3059\u3002<\/p><\/div><\/div><\/div><\/div>

<\/div>

01 \/ SINGLE-REQUEST DECODE<\/p>

1\u4eba\u3067\u4f7f\u3046\u3068\u3001
\u30c7\u30b3\u30fc\u30c9\u901f\u5ea6\u306e\u5dee\u306f\u7d0416\u301c20%\u3002<\/h2>

\u751f\u6210\u958b\u59cb\u5f8c\u306e\u901f\u3055\u3092\u793a\u3059TPOT\u3092\u6bd4\u8f03\u3059\u308b\u3068\u3001\u4eca\u56de\u306e\u30e2\u30c7\u30eb\u30fb\u8a2d\u5b9a\u3067\u306fPro 4500\u3082\u8fd1\u3044\u901f\u5ea6\u3092\u793a\u3057\u307e\u3057\u305f\u3002\u500b\u4eba\u5229\u7528\u3084\u3001\u540c\u6642\u51e6\u7406\u6570\u304c\u5c11\u306a\u3044\u7528\u9014\u3067\u306f\u3001\u3053\u306e\u9055\u3044\u3092\u3069\u3046\u8a55\u4fa1\u3059\u308b\u304b\u304c\u9078\u5b9a\u306e\u51fa\u767a\u70b9\u306b\u306a\u308a\u307e\u3059\u3002<\/p>

\u5358\u72ec\u5229\u7528\uff08Concurrency 1\uff09\u3067\u306e\u751f\u6210\u901f\u5ea6<\/caption>
\u30ef\u30fc\u30af\u30ed\u30fc\u30c9\uff08\u5165\u529b \/ \u51fa\u529btokens\uff09<\/th>4500
TPOT p95<\/th>		6000 Max-Q
TPOT p95<\/th>		6000\u306e\u901f\u5ea6\u6bd4\u203b<\/th><\/tr><\/thead>
Chat\uff081,024 \/ 256\uff09<\/th>23.3 ms<\/td>19.5 ms<\/td>1.19\u500d<\/td><\/tr>
\u696d\u52d9\u6587\u66f8QA\uff084,096 \/ 512\uff09<\/th>22.2 ms<\/td>19.1 ms<\/td>1.16\u500d<\/td><\/tr>
RAG\uff088,192 \/ 512\uff09<\/th>22.5 ms<\/td>19.4 ms<\/td>1.16\u500d<\/td><\/tr>
\u30b3\u30fc\u30c9\u751f\u6210\uff0832,768 \/ 4,096\uff09<\/th>19.4 ms<\/td>16.2 ms<\/td>1.20\u500d<\/td><\/tr><\/tbody><\/table><\/div>

TPOT\uff1a1\u30c8\u30fc\u30af\u30f3\u3092\u751f\u6210\u3059\u308b\u6642\u9593\u3002\u5c0f\u3055\u3044\u307b\u3069\u9ad8\u901f\u3067\u3059\u3002\u203b4500\u306eTPOT \u00f7 6000\u306eTPOT\u3002GPU\u306e\u6f14\u7b97\u6027\u80fd\u5168\u822c\u306e\u500d\u7387\u3067\u306f\u3042\u308a\u307e\u305b\u3093\u3002<\/p><\/div><\/div>

<\/div>

02 \/ MEASURED PERFORMANCE<\/p>

\u5229\u7528\u304c\u91cd\u306a\u308b\u3068\u3001
\u56de\u7b54\u958b\u59cb\u307e\u3067\u306e\u6642\u9593\u306b\u5927\u304d\u306a\u5dee\u3002<\/h2>

\u5358\u72ec\u306e\u30c7\u30b3\u30fc\u30c9\u901f\u5ea6\u306b\u6bd4\u3079\u3066\u3001\u540c\u6642\u5229\u7528\u6642\u306e\u5f85\u3061\u6642\u9593\u306b\u306f\u5927\u304d\u306a\u5dee\u304c\u73fe\u308c\u307e\u3057\u305f\u3002\u751f\u6210\u4e2d\u306e\u901f\u5ea6\u3068\u3001\u9806\u756a\u5f85\u3061\u3092\u542b\u3080\u56de\u7b54\u958b\u59cb\u307e\u3067\u306e\u6642\u9593\u3092\u5206\u3051\u3066\u6bd4\u8f03\u3057\u307e\u3059\u3002<\/p>

CASE 1 \uff0f \u77ed\u3044\u30c1\u30e3\u30c3\u30c8<\/p>

\u77ed\u3044\u30c1\u30e3\u30c3\u30c8\u3092\u300116\u4ef6\u540c\u6642\u306b\u3002<\/h3><\/div>

\u5165\u529b1,024 \/ \u51fa\u529b256 tokens<\/p><\/div><\/div>

<\/div>

\u56de\u7b54\u304c\u59cb\u307e\u308b\u307e\u3067\u306e\u6642\u9593
TTFT p95 \/ \u79d2\u30fb\u5c0f\u3055\u3044\u307b\u3069\u826f\u597d\uff08\u76ee\u76db\u308a 0\u301c12\u79d2\uff09<\/span><\/p>

RTX PRO 4500 32GB<\/span><\/p>

<\/div><\/div>

10.16\u79d2<\/span><\/p><\/div><\/div>

RTX PRO 6000 Max-Q 96GB<\/span><\/p>

<\/div><\/div>

3.01\u79d2<\/span><\/p><\/div><\/div><\/div>

\u56de\u7b54\u958b\u59cb\u5f85\u3061\u306e\u5dee<\/p>

3.4\u500d<\/span><\/p>

\u751f\u6210\u4e2d\u306e\u901f\u5ea6\u306f\u6bd4\u8f03\u7684\u8fd1\u304f\u3066\u3082\u3001\u56de\u7b54\u304c\u59cb\u307e\u308b\u307e\u3067\u306e\u5f85\u3061\u6642\u9593\u306f\u5909\u308f\u308a\u307e\u3059\u3002<\/p><\/div><\/div><\/div>
\u540c\u3058\u4e26\u5217\u6570\u3067\u306e\u6bd4\u8f03<\/th>Pro 4500<\/th>Pro 6000 Max-Q<\/th><\/tr>
\u751f\u6210\u4e2d\u306eTPOT p95<\/th>30.5 ms\/token<\/td>29.3 ms\/token<\/td><\/tr>
KV\u30ad\u30e3\u30c3\u30b7\u30e5\u4f7f\u7528\u7387 max<\/th>90.1%<\/td>16.7%<\/td><\/tr><\/tbody><\/table><\/div>

CASE 2 \uff0f \u9577\u3044\u6587\u66f8\u306e\u8981\u7d04<\/p>

16k\u5165\u529b\u306e\u6587\u66f8\u51e6\u7406\u3092\u30018\u4ef6\u540c\u6642\u306b\u3002<\/h3><\/div>

\u5165\u529b16,384 \/ \u51fa\u529b2,048 tokens<\/p><\/div><\/div>

<\/div>

\u56de\u7b54\u304c\u59cb\u307e\u308b\u307e\u3067\u306e\u6642\u9593
TTFT p95 \/ \u79d2\u30fb\u5c0f\u3055\u3044\u307b\u3069\u826f\u597d\uff08\u76ee\u76db\u308a 0\u301c65\u79d2\uff09<\/span><\/p>

RTX PRO 4500 32GB<\/span><\/p>

<\/div><\/div>

53.16\u79d2<\/span><\/p><\/div><\/div>

RTX PRO 6000 Max-Q 96GB<\/span><\/p>

<\/div><\/div>

6.79\u79d2<\/span><\/p><\/div><\/div><\/div>

\u56de\u7b54\u958b\u59cb\u5f85\u3061\u306e\u5dee<\/p>

7.8\u500d<\/span><\/p>

\u751f\u6210\u4e2d\u306e\u901f\u5ea6\u5dee\u306f\u7d047%\u3002\u540c\u6642\u5229\u7528\u3067\u306f\u3001\u5f85\u3061\u6642\u9593\u306b\u5927\u304d\u306a\u5dee\u304c\u51fa\u307e\u3059\u3002<\/p><\/div><\/div><\/div>
\u540c\u3058\u4e26\u5217\u6570\u3067\u306e\u6bd4\u8f03<\/th>Pro 4500<\/th>Pro 6000 Max-Q<\/th><\/tr>
\u751f\u6210\u4e2d\u306eTPOT p95<\/th>31.9 ms\/token<\/td>29.7 ms\/token<\/td><\/tr>
KV\u30ad\u30e3\u30c3\u30b7\u30e5\u4f7f\u7528\u7387 max<\/th>94.1%<\/td>12.2%<\/td><\/tr><\/tbody><\/table><\/div>

CASE 3 \uff0f \u30b3\u30fc\u30c9\u751f\u6210<\/p>

\u9577\u3044\u30b3\u30fc\u30c9\u751f\u6210\u3092\u30018\u4ef6\u540c\u6642\u306b\u3002<\/h3><\/div>

\u5165\u529b32,768 \/ \u51fa\u529b4,096 tokens<\/p><\/div><\/div>

<\/div>

\u56de\u7b54\u304c\u59cb\u307e\u308b\u307e\u3067\u306e\u6642\u9593
TTFT p95 \/ \u79d2\u30fb\u5c0f\u3055\u3044\u307b\u3069\u826f\u597d\uff08\u76ee\u76db\u308a 0\u301c160\u79d2\uff09<\/span><\/p>

RTX PRO 4500 32GB<\/span><\/p>

<\/div><\/div>

131.52\u79d2<\/span><\/p><\/div><\/div>

RTX PRO 6000 Max-Q 96GB<\/span><\/p>

<\/div><\/div>

13.97\u79d2<\/span><\/p><\/div><\/div><\/div>

\u56de\u7b54\u958b\u59cb\u5f85\u3061\u306e\u5dee<\/p>

9.4\u500d<\/span><\/p>

4500\u306f\u751f\u6210\u3092\u59cb\u3081\u305f\u8981\u6c42\u306e\u901f\u5ea6\u3092\u4fdd\u3063\u3066\u3082\u3001\u9806\u756a\u3092\u5f85\u3064\u8981\u6c42\u304c\u5897\u3048\u3066\u3044\u307e\u3059\u3002<\/p><\/div><\/div><\/div>
\u540c\u3058\u4e26\u5217\u6570\u3067\u306e\u6bd4\u8f03<\/th>Pro 4500<\/th>Pro 6000 Max-Q<\/th><\/tr>
\u751f\u6210\u4e2d\u306eTPOT p95<\/th>28.6 ms\/token<\/td>31.5 ms\/token<\/td><\/tr>
KV\u30ad\u30e3\u30c3\u30b7\u30e5\u4f7f\u7528\u7387 max<\/th>99.0%<\/td>16.1%<\/td><\/tr><\/tbody><\/table><\/div>

Qwen3.8-27B-NVFP4 \/ vLLM 0.32.0 \/ KV NVFP4\u3002\u672c\u30da\u30fc\u30b8\u5f8c\u534a\u3067\u7d39\u4ecb\u3059\u308b2U\u30fb\u5965\u884c400mm\u306e\u30b3\u30f3\u30bb\u30d7\u30c8\u30e2\u30c7\u30eb\u3067\u5b9f\u65bd\u3057\u305fGPU\u6bd4\u8f03\u306e\u7d50\u679c\u3067\u3059\u3002<\/p>

\u5b9f\u6e2c\u5024\u306e\u898b\u65b9\u30fb\u8a55\u4fa1\u6761\u4ef6<\/summary>

TTFT\u306f\u56de\u7b54\u958b\u59cb\u307e\u3067\u306e\u6642\u9593\u3001TPOT\u306f\u751f\u6210\u958b\u59cb\u5f8c\u306e1\u30c8\u30fc\u30af\u30f3\u3042\u305f\u308a\u306e\u6642\u9593\u3067\u3059\u3002p95\u306f\u3001\u8a08\u6e2c\u8981\u6c42\u306e95%\u304c\u305d\u306e\u5024\u4ee5\u4e0b\u3068\u306a\u308b\u76ee\u5b89\u3067\u3059\u3002\u4e26\u5217\u6570\u306f\u540c\u6642\u51e6\u7406\u8981\u6c42\u6570\u3067\u3042\u308a\u3001\u767b\u9332\u30e6\u30fc\u30b6\u30fc\u6570\u3068\u306f\u7570\u306a\u308a\u307e\u3059\u3002<\/p>

KV\u4f7f\u7528\u7387\u306f\u5404\u69cb\u6210\u306e\u5272\u5f53\u9818\u57df\u306b\u5bfe\u3059\u308b\u5272\u5408\u3067\u3059\u3002KV\u5727\u8feb\u3068\u5f85\u6a5f\u5897\u52a0\u306e\u76f8\u95a2\u304c\u898b\u3089\u308c\u307e\u3059\u304c\u3001\u5fdc\u7b54\u5dee\u3092KV\u5bb9\u91cf\u306e\u307f\u306e\u52b9\u679c\u3068\u3057\u3066\u5206\u96e2\u3057\u305f\u8a66\u9a13\u3067\u306f\u3042\u308a\u307e\u305b\u3093\u3002\u4e21GPU\u306eSLO\u306f\u672a\u8a2d\u5b9a\u3067\u3001\u5feb\u9069\u306b\u4f7f\u3048\u308b\u4e0a\u9650\u4eba\u6570\u306f\u672a\u78ba\u5b9a\u3067\u3059\u3002<\/p>

\u5358\u72ec\u5229\u7528\u6642\u306e\u901f\u5ea6\u5dee\u306f\u3001Chat\u30fb\u696d\u52d9\u6587\u66f8QA\u30fbRAG\u30fb\u30b3\u30fc\u30c9\u751f\u6210\u306eTPOT p95\u306e\u9006\u6570\u306b\u3088\u308b\u6bd4\u8f03\u3067\u3059\u3002\u8a55\u4fa1\u5b9f\u6a5f\u306e\u30cf\u30fc\u30c9\u30a6\u30a7\u30a2\u69cb\u6210\u306f\u672c\u30da\u30fc\u30b8\u5f8c\u534a\u306b\u63b2\u8f09\u3057\u3066\u3044\u307e\u3059\u3002\u8a55\u4fa1\u6761\u4ef6\u306e\u8a73\u7d30\u306f\u5f8c\u65e5\u8ffd\u52a0\u4e88\u5b9a\u3067\u3059\u3002<\/p><\/div><\/details><\/div><\/div>

<\/div>

03 \/ KV CACHE & QUEUE<\/p>

\u751f\u6210\u3092\u7d9a\u3051\u308b\u901f\u3055\u3068\u3001
\u6b21\u306e\u8981\u6c42\u3092\u53d7\u3051\u5165\u308c\u308b\u4f59\u88d5\u3002<\/h2>

Pro 4500\u306eChat\u3067\u306f\u30018\u4e26\u5217\u304b\u308916\u4e26\u5217\u306b\u5897\u3084\u3057\u3066\u3082\u3001\u5168\u4f53\u306e\u51e6\u7406\u91cf\u306f\u307b\u307c\u982d\u6253\u3061\u3002\u4e00\u65b9\u3001\u56de\u7b54\u958b\u59cb\u307e\u3067\u306e\u6642\u9593\u306f2.64\u79d2\u304b\u308910.16\u79d2\u3078\u5897\u52a0\u3057\u307e\u3057\u305f\u3002<\/p>

Chat\uff1a\u4e26\u5217\u6570\u3092\u5897\u3084\u3057\u305f\u3068\u304d\u306e\u51e6\u7406\u91cf\u3068\u56de\u7b54\u958b\u59cb\u5f85\u3061<\/caption>
\u4e26\u5217\u6570<\/th>4500
\u5168\u4f53\u306e\u51fa\u529b<\/th>		6000 Max-Q
\u5168\u4f53\u306e\u51fa\u529b<\/th>		4500
TTFT p95<\/th>		6000 Max-Q
TTFT p95<\/th><\/tr><\/thead>
4<\/th>176.0 tok\/s<\/td>212.5 tok\/s<\/td>0.56\u79d2<\/td>0.32\u79d2<\/td><\/tr>
8<\/th>264.9 tok\/s<\/td>377.4 tok\/s<\/td>2.64\u79d2<\/td>1.57\u79d2<\/td><\/tr>
16<\/th>272.6 tok\/s<\/td>615.6 tok\/s<\/td>10.16\u79d2<\/td>3.01\u79d2<\/td><\/tr>
32<\/th>271.5 tok\/s<\/td>806.6 tok\/s<\/td>25.40\u79d2<\/td>4.40\u79d2<\/td><\/tr><\/tbody><\/table><\/div>

\u5165\u529b1,024 \/ \u51fa\u529b256 tokens\u3002\u51fa\u529b\u51e6\u7406\u91cf\u306f\u5168\u8981\u6c42\u306e\u5408\u8a08\u3002\u4e26\u5217\u6570\u306f\u767b\u9332\u4eba\u6570\u3067\u306f\u306a\u304f\u3001\u540c\u6642\u51e6\u7406\u8981\u6c42\u6570\u3067\u3059\u3002<\/p>

\u30e2\u30c7\u30eb\u3092\u8f09\u305b\u305f\u5f8c\u306b\u3001
\u3069\u308c\u3060\u3051\u4f59\u88d5\u304c\u6b8b\u308b\u304b\u3002<\/h3><\/div>

GPU\u30e1\u30e2\u30ea\u306f\u30e2\u30c7\u30eb\u306e\u91cd\u307f\u3084\u5b9f\u884c\u7528\u30e1\u30e2\u30ea\u306b\u3082\u4f7f\u3044\u307e\u3059\u3002\u540c\u3058\u30e2\u30c7\u30eb\u3092\u52d5\u304b\u3057\u3066\u3082\u300132GB\u306896GB\u3067\u306fKV\u30ad\u30e3\u30c3\u30b7\u30e5\u306b\u56de\u305b\u308b\u4f59\u88d5\u304c\u5909\u308f\u308a\u307e\u3059\u3002<\/p>

\u540c\u3058\u8ca0\u8377\u306b\u304a\u3051\u308bKV\u4f7f\u7528\u7387\u306f\u3001Chat\u30fb4\u4e26\u5217\u306751.5%\u30684.2%\u3001\u30b3\u30fc\u30c9\u751f\u6210\u30fb4\u4e26\u5217\u306799.0%\u30688.0%\u3067\u3057\u305f\u3002Pro 4500\u306f\u65e9\u3044\u6bb5\u968e\u3067KV\u9818\u57df\u306e\u5727\u8feb\u304c\u898b\u3089\u308c\u307e\u3059\u3002<\/p>

\u5b9f\u884c\u4e2d\u306e\u8981\u6c42\u306e\u751f\u6210\u901f\u5ea6\u304c\u4fdd\u305f\u308c\u3066\u3044\u3066\u3082\u3001\u5f85\u6a5f\u3059\u308b\u8981\u6c42\u304c\u5897\u3048\u308b\u3068\u3001\u5229\u7528\u8005\u5168\u4f53\u306e\u5f85\u3061\u6642\u9593\u306f\u9577\u304f\u306a\u308a\u307e\u3059\u3002KV\u4f7f\u7528\u7387\u3068\u5f85\u6a5f\u3092\u3001\u751f\u6210\u901f\u5ea6\u3068\u4f75\u305b\u3066\u898b\u308b\u5fc5\u8981\u304c\u3042\u308a\u307e\u3059\u3002<\/p><\/div><\/div><\/div>

96GB\u306e\u4f59\u88d5\u3068\u3001\u51e6\u7406\u80fd\u529b\u306e\u9650\u754c\u306f\u5225\u306b\u78ba\u8a8d<\/h3>

Pro 6000 Max-Q\u3067\u3082\u3001RAG\u30fb16\u4e26\u5217\u3067\u306fKV\u4f7f\u7528\u7387max\u304c20.2%\u306a\u304c\u3089\u3001TTFT p95\u306f13.34\u79d2\u3067\u3057\u305f\u3002\u5b9f\u7528\u4e0a\u306e\u4e26\u5217\u6570\u306f\u3001\u5165\u529b\u306e\u9577\u3055\u3068\u3001\u8a31\u5bb9\u3059\u308bTTFT\u30fbTPOT\u3092\u6c7a\u3081\u305f\u3046\u3048\u3067\u5224\u65ad\u3057\u307e\u3059\u3002<\/p>

KV\u4f7f\u7528\u7387\u306f\u5404\u69cb\u6210\u306e\u5272\u5f53\u9818\u57df\u306b\u5bfe\u3059\u308b\u5272\u5408\u3067\u3059\u3002\u5b9f\u969b\u306eKV\u5272\u5f53\u5bb9\u91cf\u306f\u672a\u53d6\u5f97\u306e\u305f\u3081\u3001\u5bb9\u91cf\u306e\u500d\u7387\u306f\u65ad\u5b9a\u3057\u307e\u305b\u3093\u3002\u5fdc\u7b54\u5dee\u3092KV\u5bb9\u91cf\u3060\u3051\u306e\u52b9\u679c\u3068\u3057\u3066\u5206\u96e2\u3057\u305f\u8a66\u9a13\u3067\u306f\u3042\u308a\u307e\u305b\u3093\u3002<\/p><\/div><\/div><\/div>

04 \/ RIGHT-SIZING<\/p>

Pro 4500\u3067\u3088\u3044\u5834\u5408\u3082\u3002
Pro 6000\u3092\u52e7\u3081\u305f\u3044\u5834\u5408\u3082\u3002<\/h2>

\u300c\u4f55\u4eba\u304c\u767b\u9332\u3059\u308b\u304b\u300d\u306b\u52a0\u3048\u3001\u300c\u3069\u3093\u306a\u51e6\u7406\u304c\u4f55\u4ef6\u91cd\u306a\u308b\u304b\u300d\u3067\u8003\u3048\u307e\u3059\u3002<\/p>

32GB VRAM<\/p>

RTX PRO 4500 Blackwell<\/h3>

\u5c11\u4eba\u6570\u3067\u3001\u5229\u7528\u7bc4\u56f2\u3092\u6c7a\u3081\u3089\u308c\u308b\u904b\u7528\u306b\u3002<\/p>