The author demonstrates that a £200 V100 with 16GB HBM2 outperforms a £550 RTX 4070 for LLM inference, running Llama-2-13B at interactive speeds and fitting 70B models in 4-bit quantization. By documenting the full SXM2-to-PCIe adapter and cooling hack, they argue that hyperscaler decommissioning has created a secondary market where serious compute is available at scrap prices to anyone willing to do the integration work.
By submitting this story and driving it to 225 points on the HN front page, birdculture signal-boosts the argument that the V100's specs (5,120 CUDA cores, 640 tensor cores, 900 GB/s bandwidth) at a fraction of consumer card pricing exposes how decoupled Nvidia's retail prices have become from actual silicon cost. The community traction validates this as a legitimate path rather than a fringe stunt.
The editorial frames the build as evidence that Nvidia's consumer pricing has decoupled from manufacturing cost, pointing out a 16GB RTX 4060 Ti retails for £450 while a V100 with four times the HBM2 stack and higher bandwidth sells for less than half that. The thin resale market for SXM-form-factor accelerators, combined with hyperscalers aging out Volta-generation silicon, creates a structural arbitrage that the consumer GPU market hasn't priced in.
A developer writing on blog.tymscar.com documented the full process of dropping a Tesla V100 SXM2 — Nvidia's 2017 datacenter flagship — into a consumer gaming PC for a total outlay of roughly £200. The build hit the front page of Hacker News at 225 points, which is the community signaling that this is no longer a fringe stunt. The V100 in question shipped with 16GB of HBM2, 5,120 CUDA cores, 640 tensor cores, and a 900 GB/s memory bandwidth figure that current mid-range consumer cards still cannot match.
The catch, of course, is that SXM2 is a mezzanine form factor designed for DGX servers — not a PCIe card you plug into an ATX motherboard. The author sourced a SXM2-to-PCIe adapter board from AliExpress, fabricated a cooling solution by bolting a blower fan to the original passive heatsink (the card was designed to sit in a 200-CFM server chassis airflow path), and dealt with the fact that the V100 has no display outputs whatsoever — it exists purely as a compute accelerator, so the gaming GPU stays in the system for rendering.
The result: llama.cpp running Llama-2-13B at interactive speeds, with enough VRAM headroom to load 70B models in 4-bit quantization. The same workload on a new RTX 4070 (12GB, £550) would OOM on the 70B model entirely.
The economics here are the story, not the hack. Nvidia's consumer pricing has decoupled from manufacturing cost — a 16GB RTX 4060 Ti retails at £450, while the V100 module contains four times the HBM2 stack, more tensor cores, and substantially higher memory bandwidth, all for less than half the price on the secondary market. This isn't because the V100 is technically inferior for inference; it's because hyperscalers are aging out Volta-generation silicon en masse as they refresh to Hopper and Blackwell, and the resale market for SXM-form-factor accelerators is thin enough that prices collapse to scrap-metal-plus-shipping.
The community reaction on HN broke into two camps. The first pointed out that V100s lack FP8 and the newer transformer-engine optimizations that make H100s scream — so this is a capacity play, not a throughput play. The second camp, which dominated the higher-voted comments, made the practical point: for local LLM inference, VRAM capacity per dollar is the only metric that actually gates which models you can run, and the V100 currently wins that benchmark by a wide margin. A 70B parameter model in 4-bit needs ~40GB; you can stack two V100s for £400 and clear that bar. You cannot stack two RTX 4090s for £400.
The deeper signal is what this says about the consumer GPU market's pricing pressure. Nvidia has spent two generations strategically capping VRAM on consumer cards specifically to prevent them from being viable AI inference platforms — the 4080 ships with 16GB, the 4070 with 12GB, despite the silicon supporting more. The secondary market for datacenter GPUs is now actively arbitraging that artificial segmentation. Every quarter, more Volta and Ampere accelerators hit eBay as cloud providers refresh, and the floor keeps dropping. P100s are already in the £100 range; V100s will follow within 18 months.
There's a quieter implication for the homelab and indie-AI crowd: the production hardware that ran the first generation of GPT-scale models — the literal silicon that trained BERT and early GPT-3 derivatives — is now landfill-priced. If you spent 2023 watching the RTX 4090 sell out and concluded local AI was for the wealthy, the market has moved underneath you.
If you're running local LLM inference for development workflows, code completion, or RAG over private data, the build-vs-buy math has shifted hard toward build. A two-V100 rig with 32GB combined VRAM costs less than a single RTX 4070 Ti and can run models the 4070 Ti physically cannot load. The tradeoffs are real: you need a case with airflow tolerance for blower-style cooling, a PSU with the right EPS connectors, and the willingness to physically fabricate adapter brackets. You also lose modern features — no DLSS, no AV1 encode, no display output, and Nvidia's consumer drivers don't officially support Tesla cards on Windows without registry hacks. On Linux with the open-source kernel modules, it's mostly plug-and-play.
For teams evaluating whether to self-host inference vs. paying per-token API fees, this changes the breakeven calculation. A £400 dual-V100 inference box, amortized over 18 months, costs ~£0.75/day in hardware. At current Anthropic Haiku or OpenAI gpt-4o-mini rates, that's roughly 500K-1M tokens of equivalent output per day depending on which model you're displacing. If your workload is steady-state and privacy-sensitive (legal, medical, internal RAG), the secondary-market hardware route now beats API fees inside three months for any non-trivial volume.
The operational caveat is power: V100s pull 300W under sustained load, and SXM2 cards weren't designed for residential power budgets. Two of them plus a host system will hit 800W wall-draw, which matters for UK/EU users on 13A circuits and matters more for anyone calculating opex at £0.30/kWh.
The secondary-market datacenter GPU pipeline is going to keep widening. AWS, GCP, and Azure all began Hopper refreshes in 2024, and the A100s those replaced will hit the channel through 2026-2027. Expect 40GB A100s to break the £1,000 floor by end of 2026, at which point single-card 70B inference at full precision becomes a sub-£1,500 home build. The arbitrage between artificial VRAM segmentation on consumer cards and the actual cost of HBM silicon is now a structural feature of the market, not a temporary glitch — and the developer community has figured out how to exploit it.
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.