After loading into Windows 10 and ensuring all the drivers were up to date, I verified that the hardware was what was promised and also running correctly. The relevant screenshots of CPU-Z, GPU-Z, and HWiNFO confirm as much. Some quick benchmarks followed to see if the cooling system was up to par, which we will get to on the next page, before I packed up the system to take to my university for more rigorous tests and comparisons. After all, the GRANDO RM-S in the test configuration is for and justified by demanding loads that are supposed to run for longer periods of time. To do this, the same Windows build was loaded onto an external NVMe SSD along with some programs of interest, including COMSOL, protein folding, Aspen HYSIS, a game and graphics rendering engine, and a machine learning tool with emphasis on real-time language translation coupled with two-way encryption and decryption. These actual tests can not be reproduced by anyone given the sensitive nature of the data sets, and I was thus able to add two other systems for comparison whose specifications are seen below.
Test System "Supercomputer"
CPU:
20x Intel Xeon Gold 6148 CPU @ 2.40 GHz with 20 cores each
Memory:
3.84 TB DDR4 2666 MHz
Cooling:
Air-cooled server rack in an environmental chamber
GPU:
16x NVIDIA Tesla V100
Storage:
25 TB PCIe Gen 3 NVMe SSDs
Power Supply:
Various PSUs from Delta
Operating System:
Windows 10 Professional 64-bit
Note: This is a supercomputer cluster with several compute and GPU nodes, I selected 10x compute and 2x graphics nodes.
Test System "Work PC"
Processor:
AMD Ryzen 5950X, PBO Max enabled 16 cores, 32 threads 3.4 GHz base (4.9 GHz Boost)
Unsurprisingly, the GRANDO RM-S slots itself firmly between the work PC and supercomputer every single time. Performance also clearly does not scale linearly here, with different tests favoring different configurations. Note that every single test here was chosen to scale with available compute performance, and in some mock situations, differences between the three were minimal. But those are also unlikely to be used by clients of the GRANDO RMS-S, more of which will become apparant through the genereal benchmarks on the following page.