PC with three mining GPUs for 3D work

About a year ago, mining video cards started appearing on second-hand marketplaces at throwaway prices, around $1000-1$5.95 each (mining on them became unprofitable). The idea arose: why not build a powerful PC from old mining GPUs. What came out of it can be read below the cut.

First, here is the link to the article about the million-ruble computer: Building a deep learning server for half a million, maybe even a child can do it. Or not? The title mentions half a million, but four 3090 GPUs at 220 thousand rubles each and 164 thousand rubles for the other components add up to around 1 million. Overall, the explanation is correct if you're building a universal PC for training neural networks and photogrammetry. But I have a narrower task: only photogrammetry in the MeshroomCL program. For this task, you can assemble a comparable PC at 10-20 times lower cost using mining GPUs that are no longer needed by anyone else.

Let’s start with the assortment of mining cards. Nvidia released two series of cards. The first series looks like this:

The second series consists of video cards made on 12nm (TU) and 8nm (GA) processes, so they have better energy efficiency. Mining on 12nm video cards has near-zero profitability. 8nm video cards seem to still bring profit, but it heavily depends on electricity prices. In Russia, electricity prices are indexed every year on July 1st, and this year, most likely, the 12nm video cards will become unprofitable and will also appear on Avito for $1000-1$5.95. You can check which GPUs are profitable and which are unprofitable on the website hashrate.no (enter the video card model and electricity cost to see the revenue/profit/loss)

NVIDIA CMP series:
CMP 170HX: GA100, 4480 CUDA, 8GB HBMe, equivalent to A100
CMP 90HX: GA102, 6400 CUDA, 10GB G6X, equivalent to 3080
CMP 70HX: GA104, 6144 CUDA, 8GB G6X, equivalent to 3070ti
CMP 50HX: TU102, 3584 CUDA, 10GB G6, equivalent to 2080ti
CMP 40HX: TU106, 2304 CUDA, 8GB G6, equivalent to 2070
CMP 30HX: TU116, 1408 CUDA, 6GB G6, equivalent to 1660ti

To evaluate the performance of all these video cards with photogrammetry software MeshroomCL, you need to use the benchmark opencl benchmark. This benchmark has several hundred GPUs. Here are some results that are relevant to my task:

For some reason, this benchmark doesn’t have CMP170HX and CMP90HX, but these cards are still being used for mining and are not being sold at throwaway prices on Avito.

From the benchmark, it can be concluded that six R102-100 graphics cards are equivalent in power to one RTX5090. For some unknown reason, R102-100 cards are not yet being sold cheaply on marketplaces (perhaps we need to wait until electricity prices increase on July 1st). Therefore, I bought three P106-100 cards to test the performance of such a setup.

The motherboard Asrock h510 pro btc+ (supports 6 graphics cards) 5$0.00 on Avito

Memory 32GB (system requirements for MeshroomCL) 5$0.00 in the store

Power supply Aerocool Cylon 700W (for four PCIe 6+2 pin) 5$0.00 in the store

Intel i3 10100 (processor with integrated graphics is necessary, since mining cards don't have video outputs) 13$0.00 in the store

P106-100 three units at 1$5.95 each on Avito

SDD 1000Gb also around 5$0.00 in the store

Total around 45$0.00

Additionally, I have a laptop with an RTX4050 (74459 points in openCL benchmark). This laptop performs the photogrammetry task at about the same speed as a PC with two P106-100 cards (36142 points in openCL benchmark each). So the benchmark results completely correspond to the actual performance of these graphics cards in MeshroomCL.

Summary: The PC based on 16nm graphics cards turned out not to be very powerful. We need to wait until mining on 12nm graphics cards becomes unprofitable (I think this will happen after July 1, 2025, with the electricity price increase), and then it will be possible to assemble a PC with six CMP40HX cards (94577 points in openCL each). These cards will outperform, for example, one RTX5090 (371504 in openCL)

The purpose of such a strange PC configuration is described in my article on Creating a 3D model of terrain from drone video.

For the need of creating 3D terrain models, you can read here: In the USA, drones are starting to be used without GPS thanks to the Raptor system from Maxar and here: A method for drone navigation without GPS: using "terrain fingerprints"

P.S. In the next article, I want to test mining cards for training neural networks. If anyone has already tried it, leave links or your opinions on this topic in the comments.