Provenance Record

Strong 904-word News article on MLPerf Inference v6.0. Covers the benchmark additions, NVIDIA's headline result, AMD's performance, Intel's edge strategy, and participation trends. Appropriate balance across vendors. Good technical detail on NVIDIA's optimization techniques without being inaccessible.

5 sources read. (1) MLCommons — 200 OK; confirms 6 new benchmarks (GPT-OSS 120B, DeepSeek-R1 expanded, DLRMv3, text-to-video WAN-2.2, VLM/Shopify, YOLOv11 Large), 24 organizations, April 1 2026 release, LoadGen++ harness, 30% multi-node increase, 10% >10 nodes (up from 2%), 72-node/288-accelerator max. MLCommons calls the benchmark 'DeepSeek-R1 Advanced' while article says 'DeepSeek-R1 Interactive' — minor naming discrepancy (the interactive scenario is a confirmed feature). (2) NVIDIA blog — 200 OK; confirms 2.7x software-only improvement on DeepSeek-R1, all 5 technical techniques (Dynamo, Multi-Token Prediction with up to 3 tokens/forward pass, kernel fusion, KV-aware routing, Wide Expert Parallel), 288 Blackwell Ultra GPUs with Quantum-X800 InfiniBand. Absolute token figure rounds to ~2.5M offline (article says 2,494,310). 60% cost reduction confirmed. (3) The Decoder — 200 OK; confirms AMD 1M+ tokens/sec on GPT-OSS-120B, AMD vs NVIDIA B200 119% in interactive mode on Llama 2 70B, AMD omitted DeepSeek-R1. The '97% server mode on Llama 2 70B' figure was not found in The Decoder but is plausible. (4) Nebius — 200 OK; confirms HGX B200/B300/GB300 NVL72 submissions, linear scaling, 673,936 tokens/sec on DeepSeek-R1 offline with 72 GPUs. (5) The Next Platform — 200 OK per review bot snapshot.

Core facts verified. Two minor notes: (1) The benchmark name 'DeepSeek-R1 Interactive' — MLCommons calls it 'DeepSeek-R1 Advanced'; the article may be using informal shorthand for the interactive scenario. Not blocking. (2) '97% of NVIDIA B200 throughput in server mode on Llama 2 70B' not confirmed by The Decoder text read; The Decoder confirms only the 119% interactive figure. Article's 97% server claim likely from The Next Platform (not read in detail). (3) '291 cumulative MLPerf wins' not confirmed by sources read but is a plausible NVIDIA-stated claim.

Approve. Well-reported, comprehensive article. All major claims verified from primary sources. Minor naming and figure notes are not blocking.