What this is
A Reddit user crammed 8 AMD Radeon Pro V620 cards (legacy enterprise cloud-gaming GPUs on the RDNA2 architecture, each with 32GB VRAM) into one machine, for a combined 256GB of VRAM. He had Claude (Anthropic's AI coding assistant) write from scratch a set of GPU kernels (low-level instructions that actually drive the GPU's compute) tailored to RDNA2, plugged them into vLLM (the open-source inference framework that dominates LLM serving today), and ran a quantized variant of a Qwen model. The result: 60–100 tokens/sec on decode (text generation) and 3000+ tokens/sec on prefill (input processing) — roughly 8× faster than his previous llama.cpp setup. Total hardware cost: $2,800.
Industry view
What makes this build worth our attention isn't the speed numbers — it's the convergence of two things: legacy GPUs are being resurrected, and AI wrote the low-level code that lets legacy GPUs run AI. AMD has long played catch-up in the AI accelerator market, and mainstream frameworks (vLLM, SGLang, etc.) remain deeply bound to the NVIDIA CUDA ecosystem. The fact that RDNA2 can now deliver production-grade speeds tells us the potential of cheap hardware has been chronically underestimated.
But the counter-argument deserves airtime too: this is a one-off engineering feat, not an out-of-the-box product. The code is a personal fork of vLLM (a modified branch), with zero production validation; RDNA2 is two generations old, and AMD's appetite for AI software investment is limited; for actual enterprise deployment, renting cloud H100s remains the lower-friction option. Cloud pricing won't budge because of this.
Impact on regular people
For enterprise IT: For budget-constrained SMBs and departmental projects, running LLMs locally is moving from "theoretically viable" to "worth running the numbers" — a one-time $2,800 outlay for 256GB of VRAM may pay for itself inside 3 years faster than expected.
For individual professionals: Independent developers and researchers can now prototype on a local inference rig for a few thousand dollars, without watching the API bill climb on every call; for sensitive data (contracts, internal documents), there's finally a genuinely "never leaves your premises" option on the table.
For the consumer market: Falling inference costs will eventually filter through to consumer AI pricing. Subscription-based AI services billed "per token" may loosen further — and we may see more one-time-purchase, small-footprint AI devices appear.