54GB of VRAM for $35: How One Person Scored 9 P106 Mining Cards

An Unexpected $35 Find

As local AI inference costs remain stubbornly high, one Reddit user shared a jaw-dropping bargain: for just 3,000 rubles (roughly $35), they picked up an abandoned mining rig from the Russian secondhand marketplace Avito — walking away with a total of 54GB of GPU memory.

According to the original post, the seller had purchased a garage a few years back and discovered a mining setup left inside. Unsure whether it still worked, they listed it for 5,000 rubles (~$60). The buyer negotiated it down to 3,000 rubles, powered it on, and found all 9 GPUs working perfectly. The only thing missing was a hard drive — easily solved with a USB boot drive.

That works out to less than $0.65 per GB of VRAM — a price that's simply unimaginable through any legitimate retail channel. For budget-conscious local AI enthusiasts, this is practically a textbook example of a perfect deal.

What Exactly Is the P106?

These 9 cards are P106 6GB units — essentially the "mining edition" of the GTX 1060 6GB. The P106 is an NVIDIA model designed specifically for cryptocurrency mining, built on the Pascal architecture and sharing the same core as the consumer-grade 1060. The key difference: video output ports are physically removed or disabled at the factory. That means no display output whatsoever — these function purely as compute cards.

Specs and Positioning

Each P106 features 1,280 CUDA cores and 6GB of GDDR5 memory, with a memory bandwidth of roughly 192 GB/s. By today's standards, the Pascal architecture is quite dated — it lacks Tensor Cores and offers limited acceleration for FP16 and INT8 workloads. That said, it still holds some value for inference tasks on small-to-medium-sized models.

Stack nine of these together and you get 54GB of VRAM in total — enough, in theory, to load reasonably large language models. This is the core logic behind mining cards being rediscovered for AI inference: in many local deployment scenarios, VRAM capacity matters more than single-card compute performance.

What are Tensor Cores? Tensor Cores are dedicated matrix computation units that NVIDIA introduced starting with the Volta architecture (2017), designed specifically for mixed-precision deep learning. They execute FP16 matrix multiplications with far greater efficiency than traditional CUDA Cores while accumulating results in FP32 for accuracy — which is exactly what modern LLM inference acceleration depends on. Ampere and newer architectures further support INT8 and INT4 quantized inference, potentially doubling or more the throughput with minimal accuracy loss. Because Pascal lacks Tensor Cores, the P106 will generate tokens significantly slower than RTX 30/40-series cards with equivalent VRAM, and the benefits of quantization acceleration are largely out of reach.

The Hidden Costs Behind the Bargain

Getting 54GB of VRAM for $35 sounds almost like getting it for free — but the hidden costs of a multi-GPU mining rig setup are very real.

Power Draw and Cooling

A single P106 draws around 120W under full load. Running all 9 simultaneously could push peak power consumption beyond 1,000W. In regions with higher electricity rates, the ongoing power bill could quickly dwarf the original purchase price. Mining rigs were built for 24/7 high-load operation, so cooling and power supply infrastructure need to be matched accordingly.

Performance Bottlenecks in Multi-GPU Splits

That 54GB is stitched together across 9 cards — not a single large pool of memory. When a model is split across multiple GPUs, PCIe bandwidth and inter-card communication become significant bottlenecks. Mining motherboards typically use PCIe x1 risers, which are extremely bandwidth-constrained. This means real-world inference speeds will fall far short of what that total VRAM figure implies.

PCIe x1 vs. x16 bandwidth: A PCIe x1 lane offers theoretical bandwidth of only ~500 MB/s (PCIe 3.0), whereas a PCIe x16 slot provides ~16 GB/s. Mining motherboards use x1 riser cables to accommodate as many GPUs as possible, which severely limits inter-card data exchange. During tensor-parallel inference of large language models, every forward pass requires synchronizing activations across cards — and that bandwidth bottleneck directly tanks overall throughput. By contrast, professional AI servers use NVLink for direct GPU-to-GPU connections at hundreds of GB/s. In practice, the 54GB spread across 9 P106s could actually deliver worse latency in tensor-parallel inference than a single card with adequate bandwidth.

Software Compatibility with Aging Architecture

The Pascal architecture is old enough that many modern inference frameworks and quantization schemes offer little optimization for it. While projects like llama.cpp can still run on this hardware, users often need to manually work through driver, CUDA version, and framework compatibility issues.

CUDA Compute Capability and compatibility: The CUDA Compute Capability version determines which features and instruction sets a GPU can use. Pascal corresponds to Compute Capability 6.x, while many modern inference frameworks — such as vLLM and the bitsandbytes quantization library — have raised their minimum supported version to 7.0 (Volta/Turing), with some features requiring 8.0 (Ampere) or higher. This means the P106 cannot use mainstream acceleration techniques like FlashAttention or AWQ, and is limited to more broadly compatible but less efficient fallback paths. llama.cpp, which grew out of CPU-focused optimization and maintains wide compatibility with older GPUs, is essentially the primary viable framework for this hardware. Even so, users must manually match the correct CUDA toolchain version to avoid compatibility breakage between newer drivers and the older architecture.

The Bigger Picture: Mining Cards Find New Life in Local AI

This story reflects a broader trend: as the cryptocurrency mining wave recedes, large quantities of secondhand mining cards are flooding the market at rock-bottom prices — just as local LLM deployment is creating new demand for exactly the kind of raw VRAM these cards offer.

For tinkerers who don't mind the hassle, aren't sensitive to electricity costs, and are chasing the ultimate bang-for-buck ratio, a mining card array is a cheap way to build an experimental inference platform. But it's not the right fit for anyone who needs stability, low noise, or good power efficiency in a production environment. After that $35 entry ticket, the real costs hide in electricity bills, time spent troubleshooting, and performance compromises.

The real value of stories like this isn't so much a recommendation to go hunt down mining rigs — it's a reminder that in the world of local AI hardware, the trade-off between VRAM capacity and cost efficiency is far more worth thinking about than simply piling on raw compute.