Analysis of the Results

The P106-090s only support PCIe 1.1 with x4 channels for communication. In our mining rig with x1 risers, they were therefore only connected with PCIe 1.1 x1. In a PCIe 4.0 x16 environment, the same cards can be addressed with four times the throughput. The fact that the BOINC computing times hardly changed when switching from the PCIe 1.1 x1 to PCI4.0 x16 on the faster system reflected the fact that the projects we selected use the GPU almost exclusively. In the style typical of BOINC, these manageable computational jobs are designed to be computed independently of each other – they do not need to be synchronized with the computations on the neighboring GPU.

To our astonishment, the A100 cards could hardly exploit their advantages in the BOINC test. Even the speed-up factor of 10 achieved in a prime number search seems low compared to a factor of 100 if you compare the hardware price.

Although the mining rig might have been competitive in computations per euro, the P106-090 (75W per card) is clearly inferior to the maximum 250W per professional card in terms of performance per consumed watt – after all, you would have to spend between 475 and 750W for the same computing performance with data-parallel requirements. However, in commercial use, it is important to note that the real cost could be in the longer wait time. Things you can compute in an hour on a large card take a whole workday on a small one.

The machine learning test with PyTorch was different. The small cards of the mining rig completely failed to process larger batches, which specifically benefit from parallelization. The weak bus connection in the rig and slow communication due to PCIe 1.1 ate up the advantage of parallelization in the test.

There were no big surprises in the test. Although the training phase took longer than classifying the taught model, the relative times of the different hardware configurations matched. We also tried neural networks with different sizes; this had an effect on the maximum batch size, but with roughly equal relative speeds of the systems to each other. This is why we are only showing the figures for training with the smallest model in Figure 3.

Figure 3: The benchmark results for the machine learning test for image classification.

Conclusions

For data-parallel requirements like BOINC, the eight cards in the mining rig roughly match the performance of a single pro card, but cost less, even taking into account the higher power consumption. For machine learning, however, a good, modern graphics card with plenty of memory is preferable. With the support of the 16-bit floating-point numbers frequently used in machine learning compared to integer operations with 8- or even only 4-bit width, the newer cards extend their lead.

Cloud services are an interesting footnote for this study. Although not every hosting service provider offers special GPU computers yet, you can already find offers with 8 and 16 cards. Prices vary depending on the number and type of GPUs selected. In some scenarios, you might come up with a configuration where the mining rig serves as a local installation that is useful for preparing projects to run on faster systems in the cloud, as long as you are allowed to store the data in the cloud and the latencies for data transfer are compatible with the project goals.