Jun Yamog did not wait for a new card allocation.
He mounted four Tesla V100s in one machine.
Those are 32GB HBM2 parts from 2017 with 900 GB/s of bandwidth each, joined through an NVLink carrier board that adds 300 GB/s across six NVLink 2.0 links.
To house them he milled down an A100 heatsink until it fit the V100 die, which pulls 300W per chip, then ran the carrier board into a PCIe x8 connection with passthrough adapters.
Then he ran Qwen3.8-Flash-Next at 4-bit quantization against two DGX Sparks with 128GB of unified memory and a single M5 Ultra.
On long prompts the four V100s came in 40 percent faster than the DGX Spark pair and 230 percent faster than the M5 Ultra.
Jun notes that the Apple Silicon gap is likely software optimization rather than silicon.
Short-prompt token generation ran at 100.6 tok/s against 44.5 on the Mac, a 121 percent lead over the Sparks.
Nine-year-old compute, ahead of hardware released this year.
Small teams cannot control allocations.
What they can control is what runs this week, and buying the newest card is only one way to get there.
A 32GB V100 is a $744 purchase on eBay today, and renting an RTX 3090 is $0.15/hr with no build, no heatsink mods, no carrier board.
Buy Now Pay Later means zero upfront, payment after one month.
What is the oldest GPU still doing real work in your stack?
https://ai.bhkcloud.com/?utm_source=linkedin&utm_medium=social&utm_campaign=bhk_social_2026w42&utm_content=evening
Originally posted on linkedin_personal