Hello all, I just finished a build of a Xeon server. The parts as listed:
CPU: Intel Xeon Silver 4310 with appropriate Noctua cooler.
Motherboard: Supermicro X12SPL-F
Memory: 1 x Kingston Premier Series 32GB ECC Registered DDR4 3200 (PC4 25600) Server Memory Model KSM32RD4/32HDR
GPU: Nvidia RTX A2000
Storage: Intel 670p Series M.2 2280 1TB PCIe NVMe 3.0 x4 QLC Internal Solid State Drive (SSD) SSDPEKNU010TZX1
OS: Windows 10 for benchmarking purposes.
The system scored a 7,070 on passmark.
The CPU scored a 19,336. Which is 17.2% slower than the global average of 23,354 for the same CPU.
The worst was the memory score, 2,143. Very low. Passmark does not have an entry for my memory to compare against.
The memory is 3200 but the CPU only supports up to 2667. I figured that would hurt, but this much? Could it be because it’s one DIMM? A large DIMM?
Icelake doesn’t really move the needle over broadwell in single threaded CPU perf and in some cases is a regression of skylake (due to clock limitations).
It it a big step up in the IO connected to the CPU though.
if you populated more memory channels your score should go up (although like 2/3 of passmark’s memory benchmark is single threaded and doesn’t really scale with memory channels), also the higher cpu scores are likely that of dual socket boards that can offload background tasks to the non-benchmarked CPU while one socket is being benchmarked.
I have a Threadripper system, but 17,2% is in the ballpark of what I loose when I use fever memory channels, according to CinebenchR23 and Geekbench in this case. I have populated 8 memory channels and with only two channels populated I saw a decrease in performance of about 20-25%.
Like all benchmarks, Passmark is one whose results you have to take with a grain of salt. For example, my dual CPU Broadwell system scores appreciably higher on the memory benchmark than my dual CPU Skylake system, despite achieving significantly lower total memory bandwidth (listed as “memory threaded”). These scores will all be entirely different when run under Linux as opposed to Windows, as well.
And even then, performance as measured by Passmark is a fraction of the bandwidth achieved by running STREAM TRIAD.
While I agree with the other posters that having only a single DIMM populated is hurting your performance, Passmark is also just not a particularly good benchmark
I added three more sticks, although the eBay store sent me the wrong sticks. Anyway, the sticks are close enough and no stability issues so far, so I ran the benchmarks again.
23086 CPU Benchmark
and
2693 Memory Benchmark
Much improvement with 3 more DIMMS for 4 DIMMS in total.
The RAM they sent me was KSM32RD4/32MEI and I was trying to order more KSM32RD4/32HDR. I don’t see a problem with it. Other than the KSM32RD4/32HDR is Hynix and has more IDD specifications, which I don’t know what those are, and the KSM32RD4/32MEI is Micron. If anyone could explain what IDD specifications are, that’d be cool. Google didn’t really bring up much I could comprehend.
