Hobbyist doubles VRAM for $266 with repurposed Tesla V100

An AI enthusiast has bolted a decommissioned enterprise GPU into a consumer desktop, doubling VRAM to 32GB for $266 and running a 27-billion-parameter model at interactive speeds, a data point that underscores how the secondary market is rewriting the economics of local inference.

The hardware arithmetic

Oscar Molnar sourced a Tesla V100 SXM2 with 16GB of HBM2 memory for roughly $140 on eBay, then spent $66 on an SXM2-to-PCIe adapter to make it physically compatible with a standard motherboard. The card sits alongside an RTX 4080 and its 16GB of Ada-generation VRAM, giving the system a combined 32GB across two architectures, Volta and Ada, for a total outlay of $266. A 32GB V100 exists but commands double the price, according to Molnar.

The noise problem

The V100’s stock cooler measured 82 decibels, which Molnar described as somewhere between a garbage disposal and a lawnmower. The fix was a PWM modification: rerouting the fan wires to the motherboard’s fan header so the speed could be controlled in software. At 10% duty cycle the GPU stays under 50°C at full load, turning a server-grade shriek into something a home office can tolerate.

The software glue

Making two GPU generations cooperate required a legacy Nvidia driver with overlapping support for both Volta and Ada, running on NixOS. With that in place, a 27-billion-parameter model delivered 32 tokens per second, fast enough for interactive use, Molnar says, and faster than most cloud API alternatives.

What it signals

The project is a reminder that the constraint on local LLM performance is VRAM capacity, not compute, and that the enterprise upgrade cycle is flooding the secondary market with cards that solve exactly that constraint for pocket change. The prices quoted in the build, $140 for a 16GB HBM2 accelerator, already look optimistic; commenters on the original story report listings closer to $500. But the architecture is sound: if you can tolerate the plumbing, the lawnmower, and the driver archaeology, the economics of inference are moving decisively toward the hobbyist.