Which is better for scientific and engineering computation/simulations, the Intel i9-10900X or the AMD R9-3950X?

The R9 has 6 more cores, whereas the i9 has doubled memory channels.

Is it possible that in many use cases R9's real performance is limited by DDR bandwidth?

R9's TDP is 105W, i9's TDP is 165W, but there are rumors that i9's real power can be 300W+ at normal full load.

What are the zero and light load whole system power consumption? Hope it is less than ~50W.

Don't know why all AMD's CPU with 4 memory channels are far more expensive or why doesn't AMD have a CPU with 4 memory channel with ~12 cores.

R9 also have more cache: 72MB. https://images.anandtech.com/doci/15062/AMD%20Fall%20Desktop%20Announcement%20Briefing%20Deck-page-002.jpg

R9's DDR frequency is a higher than i9: 3200 VS 2933, DDR bandwith is 51GB/s VS 93GB/s


There are indeed fields of computing where Intel's X299 lineup still performs better than AMD's Ryzen 3000 series. And even offers better performance per dollar. Engineering/scientific simulations like CFD and FEA is one of these areas. Source: I have been working in this field for many years now.

You already mentioned the key specification here: memory bandwidth. AMDs Ryzen CPUs stop scaling at around 6-8 cores due to a lack of memory bandwidth when running parallel CFD codes. See for example the results here: https://www.cfd-online.com/Forums/hardware/198378-openfoam-benchmarks-various-hardware-14.html#post763993

Intel CPUs can draw a lot more power than their rated TDP for short periods of time. That's how turbo boost works. So can AMD CPUs. Unless you are overclocking heavily, you won't see 300W sustained power draw from an I9-10900X. Not even close https://www.anandtech.com/show/15039/the-intel-core-i9-10980xe-review/2

Less than 50W idle power consumption for a whole system on an HEDT platform might be a bit optimistic. You will probably end up in the 70W range with a typical configuration. But this figure does not depend on the CPU alone. Motherboard, drives, GPU... all play a role here.

Why doesn't AMD release a Threadripper CPU with less than 24 cores, at a more reasonable price?

I can only speculate, and there will be a lot more reasons than I can come up with in a short paragraph. But one of the most important ones will be this: AMD is a relatively small company compared to Intel. So far, their goal is not to cover every niche possible with the perfect product. That was Intels goal. And let's face it: scientific and engineering computations on moderately priced workstation PCs is a niche application.

More L3 cache is nice to have, but outside of edge-cases, it can not completely make up for a lack of memory bandwidth.

Btw: if you are looking for a cheaper TR3000 alternative that better fits your requirements, AMD still has something in their lineup: Epyc 7302P. The CPU and the boards are not exactly cheap, but still much better value and higher total performance compared to the cheapest possible TR3000 build.

Edit forgot to mention this: especially with your applications in mind, I would highly recommend faster memory than DDR4-2933 or DDR4-3200 for both AMD Ryzen 3000 and Intel Cascade Lake-X CPUs. DDR4-3600 is around the sweetspot for price/performance.

  • Why do you recommenda DDR4-3600 if the CPU only suppoort 3200 and 2933? Did you mean to overclock the DDR?
    – jw_
    Apr 16 '20 at 1:11
  • Yes, the reason may be that the cost for TR4s CPU module is very high (very complex IO die - not the same with the 39XX IO die?) then it is not cost effective for 2 CPU die+1 complex IO die+ 1 complex substrate.
    – jw_
    Apr 16 '20 at 1:13
  • Recommending faster than "supported" RAM: yes, technically this counts as overclocking the integrated memory controller. But it is so easy to do these days, with practically zero chance of breaking anything. And with performance of the memory subsystem being crucial to these applications, it is well worth it. Since generational improvements of CPUs (at least on Intels side) are very small, running the memory faster gets you 2-3 CPU generations worth of extra performance.
    – user13807
    Apr 16 '20 at 9:26
  • 1
    Worth adding is that 10900X supports AVX-512. Science applications especially are usually good at making use of it. If your application does make use of AVX512, the 10900 will destroy R9 in performance. If you use GPU computing the 10900X also has far more PCIe lanes. Apr 18 '20 at 15:04
  • 1
    @JanDorniak “If your application does make use of AVX512, the 10900 will destroy R9 in performance” Any references? Commonly the boost are disappointing. "10900X also has far more PCIe lanes" Doesn't AMD also have a lot of lanes and are even 4.0 lanes?
    – jw_
    Apr 30 '20 at 1:54

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.