I put a datacenter GPU in my gaming PC
A datacenter V100 GPU with an SXM2-to-PCIe adapter added 16GB VRAM to a gaming PC for £200.
The author acquired a used Tesla V100 SXM2 (16GB HBM2 memory, Volta architecture) for £150 on eBay and purchased an unofficial SXM2-to-PCIe adapter for £50 to integrate it alongside an RTX 4080. The V100 has 900GB/s memory bandwidth, exceeding newer consumer GPUs including the RTX 4080 (736GB/s) and M5 Max (614GB/s). The adapter's fan ran at 82dB; the author controlled it via PWM by wiring it to a motherboard fan header with a JST connector cable.
With 32GB total VRAM split across both GPUs, the setup runs Qwen3.6-27B-MTP at approximately 32 tokens per second and 128k token context. The model includes vision support (928MB overhead). Driver configuration required legacy NVIDIA branch 550.x to support both Ada (4080) and Volta (V100) architectures, kernel 6.6, and CUDA 12.2. One quirk: the V100 occasionally vanishes from enumeration after warm reboots, requiring a cold power cycle.
What HN community is saying
The thread's strongest sentiment: the project is hardware-creative but impractical for typical users due to driver debugging, kernel version pinning, and ACPI issues. One commenter disputes the article's writing style as LLM-generated; the author clarifies they wrote it without AI and notes that LLM detectors are unreliable. Technical alternatives emerged: AMD MI250X GPUs offer 128GB HBM2E for under £1k but lack mainstream driver support and require proprietary HPE firmware in the secondhand market. Broader context: V100 hardware is cheap because it is end-of-life e-waste from 2017; newer RTX 5090 cards have 2x the CUDA cores and 2x the bandwidth but cost over £2,000.