Nine P106 cards and 54GB of VRAM at 3,000 rubles set a cheap local-LLM floor
Nine P106 cards with 54GB of VRAM were bought for 3,000 rubles, about $35, after the price was cut from 5,000 rubles, and the buyer says the rig can run quantized local LLMs.
This purchase is a memory bargain, not a speed claim. The page lists nine P106 cards, 6GB each, and a negotiated total of 3,000 rubles, but it does not give memory bandwidth, PCIe topology, power draw, or sustained token rate.
The constraint is the buyer, not the fab. A hobbyist gets aggregate VRAM, but Pascal-era cards on x1 links likely limit token speed and prefill. For a local-LLM user, the deal changes the cost floor for large quantized models, not the performance ceiling.
Unmeasured are per-card bandwidth, the host bridge, the SSD or USB boot penalty, and whether the claimed 30+ tokens per second holds over a longer run. The document also omits the exact quantization behind the 70B claim.
The buyer is constrained by PCIe x1 links and Pascal-era bandwidth, not by a lack of VRAM. The document shows a cheap aggregate-memory path, not a fast local-LLM platform.
Lead with the cards and negotiated price, then flag missing bandwidth and topology. Keep the tone skeptical about speed claims.
9 P106 GPUs · 54GB total VRAM · 6GB per GPU · 3,000 rubles final price · 5,000 rubles starting price · 30+ tokens per second
After Wccftech Hardware. We did not report this. The pictures, if any, are theirs.
