RTX PRO 4000 Blackwell 70W→120W mod boosts Qwen 27B generation 67%
iam31337 · reddit · 2026-10-26
On a 70W RTX PRO 4000 Blackwell SFF (24GB GDDR7, no aux power), the author loaded a modified VBIOS into memory (no flashing/soldering) to raise the limit to 120W and benchmarked LLMs: Qwen3.8 27B + DFlash2 went from 35.04 to 58.53 tok/s (+67%) with 120W, VRAM +20%, core +180MHz.- 27B kept gaining up to 120W; 35B A3B was flat between 106-120W- VRAM +33.3% slightly hurt 27B decode (57.70→56.91 tok/s), suggesting shared power budget; yet gave +19.8% decode on 35B A3B- Takeaway: tune for your actual workload (prefill vs decode), not panel numbers- Caveats: 120W exceeds the 75W PCIe slot budget; only 1-3 requests per config, ECC OFF, different quantization settings. Full write-up and CSVs on the blog.
More from Infra
- Cloud Architecture Is Being Rewritten by AI: GPUs, Token Economics Take Center Stage — DavidLinthicum · 2026-10-26
- Local MoE Inference Sees <10% Prefill Boost From TFLOPS Uplift, Decode Unchanged — johnowhitaker · 2026-10-26
- DeepSeek is having its Linux moment: open infrastructure as ecosystem strategy — TheTuringPost · 2026-10-26
- SpaceX seen as a 'hyper-neocloud' that will force 200+ neoclouds to consolidate — JOBhakdi · 2026-10-26
- From Big Tech SWE to AI Infra Startup: One Year of SGLang Contributions, Detailed — DrDatta_AIIMS · 2026-10-26
- Silicon Data plans tradable LLM token indexes with CME, weighted by usage — Beth_Kindig · 2026-10-26