Three 'e-waste' Nvidia V100s hit 100 tokens per second on Qwen 3.6 35B in a $1,042 homelab rig
- ID
- 48256
- Status
- summarized
- Published
- 19 Nov 2026, 7:00 PM
- Fetched
- 19 Nov 2026, 8:02 PM
- Provider
- Tom's Hardware
- Category
- technology
- Original URL
- https://www.tomshardware.com/tech-industry/artificial-intelligence/three-e-waste-nvidia-v100s-hit-100-tokens-per-second-on-qwen-3-6-35b-in-a-usd1-042-homelab-rig-decommissioned-gpus-outpace-a-usd5-399-m5-max-mac-studio-at-long-contexts
- Source URL
- https://www.tomshardware.com/feeds/all
Summary
- Score
- 6.0
- Created
- 19 Nov 2026, 8:03 PM
- Tags
- Audience
- developersai_ml_learnersvibe_coders
What happened
Tom's Hardware reports a homelab build using three decommissioned Nvidia V100 GPUs, put together for $1,042, reaching roughly 100 tokens per second on Qwen 3.6 35B — and claims it outpaces a $5,399 M5 Max Mac Studio specifically at long contexts. The excerpt available here is headline and page furniture only: no methodology, quantization level, context length, batch size, or power draw is included.
Why it matters
The concrete claim is a ~5x price gap ($1,042 vs $5,399) with better long-context throughput, which is the number to check before assuming a new Mac Studio is the default local-inference buy. But since the text gives no quantization, context length, or power/cooling costs, treat 100 tok/s as unverified — if you're speccing a self-hosted box for a 35B model, ask for those four numbers before spending anything.
Discussion angle
Before anyone copies this build: what quantization and context length produced 100 tok/s, and does the $1,042 include PSU, cooling, and the electricity for three datacenter cards? Ask what the V100s cost to run per month versus a Mac Studio's wall draw.