How do I sell a used Quadra GV100?

How do I sell a Quadro GV100 that did not work out for my CFD application? I want a potential buyer to rest assured that it works. As the seller, I want to make sure that the GPU worked when I turned it over to the buyer. I am considering Ebay and Craigslist.

Why not? What’s the CFD application? ANSYS Fluent supports GPU acceleration with Quadro, and there are some other packages like OpenFOAM that can take advantage of a GPU as well, I think.

It is a custom CFD code running on a Supermicro workstation with 32 intel processors (64 ranks) and 256GB of memory. The code is running under Ubuntu. The CFD code is written in MPI and Fortran. The data transfer between the ranks and the device is too slow. Once the arrays are on the GPU, it is about 20-30 times faster than a single CPU for a simple Jacobi relaxation of a 3D Laplacian. However, 64 ranks are faster once the time for data transfer is included.

When using GPUs for acceleration, it is important to avoid unnecessary data movement between host and device, and keep data resident on the GPU for as many processing steps as possible. I assume you designed your software accordingly, and also made sure you get close to PCIe gen3 maximum throughput of 12.5 GB/sec for those transfers that can’t be avoided?

I assume the “32 Intel processors” are actually 32 CPU cores, possible spread over two physical CPUs (dual socket system)? The CPU/GPU speedup you observe seems plausible. Out of curiosity: What specific Intel CPUs are being used in this system?

They are 2 x E5-2683V4 XEON 2.1G 16 C 120W CPUs. The workstation is a dual socket Supermicro workstation SYS-7038A-I. The data transfer rate is about 11.7 GB/s. Another problem is that the GPU has to allocate memory for every core that is connected to it. This reduces the amount of available memory to perform calculations. In an ideal world, you would want 1 GPU for every 1 to 2 cores.

A more common ratio suitable for the majority of HPC applications is one GPU per every four CPU cores. But clearly one GPU for a big 32-core platform makes for a poorly balanced system, I’ll give you that.

I am looking forward to PCIe gen4 deployment to double those CPU/GPU transfer rates on x86 hardware. Looks like we will have to wait another year for that :-(