Compute Benchmarks: RTX 5090 vs 6000 PRO
panchovix · reddit · 2026-07-12
The author provides a detailed performance comparison of the RTX 5090, 6000 PRO MaxQ, and 6000 PRO WS/SE, covering image generation and local LLM inference.
They performed a shunt mod on the 6000 PRO MaxQ, combined with water cooling, pushing the power limit to around 600W, with temperatures ranging between 45°C and 60°C. A rented 6000 PRO WS from runpod was used as a baseline.
Test software and setups include:
- Various PyTorch versions
- SageAttention 2.1
- Forge neo
- RTX upscaling extensions and additional sampler extensions
- torch compile using max autotune without cudagraphs
Image generation tests were run with fixed samplers, steps, and prompts. For LLMs, llama.cpp was used with partial model offloading to the CPU, bottlenecking the main GPU to test local inference performance. Overall, the post highlights how different GPUs, power limits, and cooling solutions impact AI compute throughput.
Related event: Performance Comparison: RTX 5090 vs RTX 6000 PRO(2 posts)→
More from Infra
- AWS on agent prompt caching: explicit + implicit patterns to slash token costs — AWS ML Blog · 2026-10-27
- Thieves spoofing GPS steal $100M+ in data center cargo since June — Ars Technica AI · 2026-10-27
- Lam Research: weekly token usage up ~250x in 18 months — Beth_Kindig · 2026-10-27
- AWS shows how to build enterprise agentic retrieval on Bedrock with CDK — AWS ML Blog · 2026-10-27
- Encrypted GPT-2 speedrun challenges FHE inference: best is 3.031 s/token, goal sub-1s — tensorqt · 2026-10-27
- Open-source Latch compiles PyTorch into training megakernels, 2.75x faster on H100 — sguada · 2026-10-27