PDF 原檔：報告_OCP2026_SemiAnalysis_從晶片到服務_20261015_original.pdf
講者公開簡報（OCP 官方 Recorded Content 頁提供下載）。文字以 PyMuPDF 逐頁抽取，
### p<n>為頁碼；圖表內容見原檔。
圖片清單（已驗證 2026-10-19）
| 檔名 | 頁 | 分類 | 親眼所見內容 |
|---|---|---|---|
報告_OCP2026_SemiAnalysis_從晶片到服務_20261015_p9.png |
p9 | 真資料圖 | AgentX 重播 393 個匿名 Claude Code 工作階段（自 8,300 筆語料挑選）：flamegraph 示意，p50/p90/p99 輸入 88K/272K/675K tokens，44% 工作階段用子代理，最大 context 1M |
報告_OCP2026_SemiAnalysis_從晶片到服務_20261015_p10.png |
p10 | 真資料圖 | AgentX 2026-08 結果：NVIDIA vs AMD SGLang Qwen3.5 90 tok/s/user 達 20x+、GLM 5.3 150 tok/s/user 成本效率最高 5x、vLLM CPU KV offload +81.7%；右圖為 DeepSeek V4 Pro 每 $ 總 token 對 P90 互動性曲線與各 SKU TCO $/chip-hr |
報告_OCP2026_SemiAnalysis_從晶片到服務_20261015_p12.png |
p12 | 真資料圖 | Rubin 對 GB300 每 MW token 數：DeepSeek V4 Pro 預覽 75／100／125／150／170／200 tok/s/user 分別為 1.45x／2.09x／4.29x／7.19x／5.56x／2.72x；模型化單晶片全含功耗 VR200 3.3 kW、GB300 2.12 kW、MI355X 2.09 kW |
報告_OCP2026_SemiAnalysis_從晶片到服務_20261015_p13.png |
p13 | 真資料圖 | TPU、NVIDIA、AMD 同圖（Qwen3.5 397B FP8 8K/1K，內部 TCO）：TPU7x $1.03/chip-hr，低延遲 B200/GB200 領先、高延遲 TPUv7 領先；各 SKU TCO $/chip-hr 列於圖頂 |
報告_OCP2026_SemiAnalysis_從晶片到服務_20261015_p14.png |
p14 | 文字卡 | TPUv7 Ironwood 預覽範圍：100 tok/s/user 外部 TCO 每百萬 token $0.181 vs B200 $0.222、B300 $0.276；列出後續比較 TPUv8i、分離式 prefill/decode、MTP、AgentX on TPUv7 |
報告_OCP2026_SemiAnalysis_從晶片到服務_20261015_p15.png |
p15 | 文字卡 | InferenceX 後續路線：Rubin NVL72 由預覽轉正式、MI455X UALoE72、TPUv8i、SambaNova SN50、Amazon Trainium3（僅規劃，無結果） |
原始內容
p1
2026 | Confidential | All Rights Reserved Dylan Patel Founder & CEO, SemiAnalysis From Silicon to Services Market Impact & Adoption
p2
2026 | Confidential | All Rights Reserved 2026 | Confidential | All Rights Reserved Who is SemiAnalysis ? INTRODUCTION • SemiAnalysis is the best - of- breed Semiconductor and AI Research boutique. Our coverage spans the entire AI Infrastructure supply chain, from semiconductor fabrication essentials to cutting - edge AI Models, software, and infrastructure for them. • Product - first Approach • Analysts have extensive experience following and working in the Semiconductor and AI industry, allowing us to delve into technical complexities. We attend 80+ conferences annually and interact with industry professionals to understand technological details and the merits of products and services. • Connecting the Dots • No other independent research house spends most of its time in the field, seeking out and connecting publicly available data points and applying deep technical expertise to build bottoms - up models that drive actionable conclusions and insights. Independent semiconductor and AI research, 80+ conferences a year, bottoms - up models
p3
2026 | Confidential | All Rights Reserved Column One Column Two Column Three Column Four 2026 | Confidential | All Rights Reserved What We Build Fourteen institutional models, two open benchmarks, one rating system, one free price index Tokenomics Datacenter – Industrials, Energy, China Datacenter AI Supply - Accelerator & HBM, Memory Model AI Cloud TCO AI Networking Wafter Fab Equipment Core Research Foundry Compute Capital and Markets (launching) Chip Book Space DC TCO Model VR NVL72 Model InferenceX AgentX EndpointX (first edition coming) SandboX (upcoming) ClusterMAX 3.0, September 2026 323 providers tracked 77 reviewed 19 Medallions GPU Spot - Contract Composite Indexes, hourly H100 1- year contract index, free since April 2026 H100 composite on Bloomberg, ticker SAH100SC All other GPUs available to clients INTRODUCTION MODELS BENCHMARKS RATINGS DATA Source: semianalysis.com, inferencex.semianalysis.com, clustermax.ai, gpu - index.semianalysis.com, 30 Sep 2026
p4
2026 | Confidential | All Rights Reserved 2026 | Confidential | All Rights Reserved STEEL SEMIANALYSIS TEARDOWNS LAB Top: sheet cuts, Intel P - core logic and the Samsung logic field labelled CU. Bottom: gate cuts, Intel P - core logic and Samsung NPU logic. Cut directions are matched; cell functions and drive targets differ. Scale bars per panel. Source: SemiAnalysis Intel Core Ultra 7 365 (Panther Lake). Source: SemiAnalysis
p5
2026 | Confidential | All Rights Reserved InferenceX
p6
Friday, March 6, 2026 TITLE GOES HERE IN MASTER STYLE 2026 | Confidential | All Rights Reserved 2026 | Confidential | All Rights Reserved One Open Benchmark, Every Accelerator INFERENCEX NVIDIA, AMD, Google TPU and OpenAI silicon, rerun on public GitHub Actions 13 accelerator SKUs ~19 AgentX models 10+ serving stacks 1,806 GitHub stars Data: inferencex.semianalysis.com homepage and About page, captured Oct 6, 2026; every data point links to its public GitHub Action s run Source: InferenceX , About
p7
Friday, March 6, 2026 TITLE GOES HERE IN MASTER STYLE 2026 | Confidential | All Rights Reserved 2026 | Confidential | All Rights Reserved Backed by the Inference Ecosystem INFERENCEX Model labs, hyperscalers, neoclouds and serving frameworks, plus Lisa Su and Jensen Huang Data: InferenceX supporters page (logos and named quotes), captured Oct 6, 2026 Source: InferenceX supporters
p8
Friday, March 6, 2026 TITLE GOES HERE IN MASTER STYLE 2026 | Confidential | All Rights Reserved 2026 | Confidential | All Rights Reserved Agentic Coding at OpenAI AGENTX Codex accounts for 64% of OpenAI enterprise output tokens in the June 2026 snapshot Data: share of OpenAI enterprise output tokens, ChatGPT vs agentic (Codex), 28 -day trailing average, Aug 2025 to Jun 2026 Data: coding agents alternate model turns and tool calls. Green blocks reuse context from earlier turns Source: SemiAnalysis, AgentX / InferenceX v3
p9
Friday, March 6, 2026 TITLE GOES HERE IN MASTER STYLE 2026 | Confidential | All Rights Reserved 2026 | Confidential | All Rights Reserved AgentX Replays Real Coding Sessions AGENTX 393 anonymized Claude Code sessions, selected from a corpus of 8,300 sessions 88K / 272K / 675K p50/p90/p99 input tokens 393 sessions in AgentX v1.0 44% sessions use subagents 1M max context Data: AgentX flamegraph of one session (turns, subagents, cached vs uncached input); ISL percentiles from DeepSeek V4 Pro 0813 runs ; corpus 8.3K sessions, 3.4M requests, $3M+ spend (Aug 2026) Source: SemiAnalysis, AgentX / InferenceX v3 · Hugging Face dataset
p10
Friday, March 6, 2026 TITLE GOES HERE IN MASTER STYLE 2026 | Confidential | All Rights Reserved 2026 | Confidential | All Rights Reserved NVIDIA Leads Selected AgentX Results AGENTX August 2026 results. The advantage depends on the model, engine and latency target • 20x+ NVIDIA vs AMD on SGLang, Qwen3.5 at 90 tok/s/user • Up to 5x NVIDIA cost efficiency on SGLang, GLM 5.3 at 150 tok/s/user • +81.7% output throughput from vLLM CPU KV offload • GB300 NVL72 Dynamo leads tokens per $ on DeepSeek V4 Pro Data: DeepSeek V4 Pro 1.6T · FP4/FP8 · AgentX · total tokens per $1 TCO (owning, hyperscaler) vs P90 interactivity · InferenceX, up dat ed Aug 21, 2026 Source: SemiAnalysis, AgentX / InferenceX v3 ; Forbes
p11
Friday, March 6, 2026 TITLE GOES HERE IN MASTER STYLE 2026 | Confidential | All Rights Reserved 2026 | SemiAnalysis AGENTX AMD + TileRT Reaches ~470 Tokens/s/User GLM5.2 / GLM5.3 744B AgentX. P90 interactivity. September 24, 2026 benchmark view ~470 tok/s/user peak P90 interactivity on MI355X FP8 vLLM prefill + TileRT decode The displayed GB300 curve uses Dynamo SGLang FP4 Higher TTFT in the source result Peak streaming speed alone does not establish a cost - efficiency win Actual benchmark graph · owning at hyperscale TCO B300 $2.26 / GB300 $2.31 / MI355X $1.50 per chip - hour Source: SemiAnalysis InferenceX · September 24 view, captured October 7, 2026
p12
Friday, March 6, 2026 TITLE GOES HERE IN MASTER STYLE 2026 | Confidential | All Rights Reserved 2026 | Confidential | All Rights Reserved Rubin Delivers Up to 7.2x Tokens per MW AGENTX DeepSeek V4 Pro preview: 2.1x at 100 tok/s/user and 7.2x at 150, versus the best GB300 engine Data: DeepSeek V4 Pro 0813 1.6T AgentX · tok/s per all -in utility MW vs P90 interactivity · modeled all - in power/chip VR200 3.3 kW, GB300 2.12 kW, MI355X 2.09 kW · updated Sep 10, 2026, preview Data: same run, throughput in M tok/s per MW at matched P90 interactivity · Rubin on pre -release software Source: SemiAnalysis, Vera Rubin NVL72 Agentic Inference
p13
Friday, March 6, 2026 TITLE GOES HERE IN MASTER STYLE 2026 | Confidential | All Rights Reserved 2026 | Confidential | All Rights Reserved TPU, NVIDIA and AMD on One Chart INFERENCEX Qwen3.5 397B, FP8, 8K/1K. Internal TCO versus end - to- end latency • $1.03 per TPU chip-hour, using Google's internal TCO • At low E2E latency, B200 and GB200 lead TPUv7 • At higher E2E latency, TPUv7 leads the displayed curves • Same FP8 model and 8K/1K workload. Serving recipes differ by chip Data: Qwen3.5 397B · 8K/1K · FP8 · STP · total tokens per $1 TCO (internal TCO, large hyperscaler volume) vs E2E latency · dashboar d updated Sep 25, 2026 · TPUv7 is a preview Source: InferenceX dashboard ; TPU InferenceX article
p14
Friday, March 6, 2026 TITLE GOES HERE IN MASTER STYLE 2026 | Confidential | All Rights Reserved 2026 | Confidential | All Rights Reserved TPU Preview Scope and Next Steps INFERENCEX FP8 results today. Disaggregation, MTP and AgentX remain follow - up work • Matched precision: The chart compares FP8 across accelerators. Blackwell FP4 results answer a different deployment question • Preview scope: TPUv7 uses aggregated serving and STP. Disaggregated prefill/decode, MTP and AgentX remain future comparisons • Serving software: TorchTPU supports native PyTorch with Pallas kernels. This TPU preview uses vLLM EXTERNAL TCO, 100 TOK/S/USER TPUv7 Ironwood, official preview Qwen3.5 397B · FP8 · 8K/1K · STP vLLM on TorchTPU $0.181 per M total tokens vs $0.222 B200 and $0.276 B300 PLANNED COMPARISONS TPUv8i: native FP4, compared FP4 vs FP4 Disaggregated prefill/decode MTP AgentX on TPUv7 Data: external TPU TCO at 100 tok/s/user, Qwen3.5 FP8, 8K/1K, STP. Source article: Sep 7, 2026 Source: TPU InferenceX article (Sep 7, 2026) ; Inferact TPU partnership post
p15
Friday, March 6, 2026 TITLE GOES HERE IN MASTER STYLE 2026 | Confidential | All Rights Reserved 2026 | Confidential | All Rights Reserved What's Next on InferenceX INFERENCEX Coming next on the same methodology: Rubin NVL72, MI455X, TPUv8i, SN50 and Amazon Trainium3 Rubin NVL72 From preview to official results MI455X UALoE72 Rack - scale AMD TPUv8i Native FP4; FP4 vs FP4 comparisons, and AgentX on TPUv7 first SambaNova SN50 Listed "coming soon" on the InferenceX dashboard Amazon Trainium3 Planned benchmark coverage. Performance results pending Same open methodology for every new chip: public recipes, public CI, same TCO basis Data: roadmap in the Oct 6 draft. Trainium3 is planned. No Trainium3 benchmark result is shown Source: InferenceX · github.com/SemiAnalysisAI/InferenceX
p16
2026 | Confidential | All Rights Reserved QR code linking to semianalysis.com semianalysis.com Scan to visit