Ryzen Performance on Scientific Applications

[karlo13]

Hi, I would like to ask if anyone had tested Ryzen in scientific programs specifically in Numerical Weather Prediction applications such as Weather Research and Forecasting Model (WRF). Is there any bench available?

http://www2.mmm.ucar.edu/wrf/users/download/get_source.html

[NineEyeRon]

“Will it run Chemstation” is the new “Will it run Crysis”

[NunoLava1998]

If it's multi-threaded, Ryzen (and also Threadripper) will run much better than i.e. an Intel processor.

It likely is multi-threaded lol

[NunoLava1998]

The Ryzen ThreadRipper 2950X has quite good multi-core, about 1.6x better than a 9900K and 2.2x better than a 8700K. MSRP is only 829$, too, and fucktons of PCIe lanes.

[karlo13]

Yes its heavily multi-threaded application but my concern is the program is kinda optimized to intel processor, it is due to amd haven't keep up before in these field. now i'm planning to get the upcoming ryzen 3000. currently in my i7 6700k setup, a resolution of 5km for a forecast of 5 days in WRF, takes a day or two to finish. since ryzen offers a very good perf/price ratio it is a very good choice, but i couldn't find any articles where ryzen was tested in these applications so im doubting to get it since my first priority is to get WRF run faster.

[leadeater]

3 minutes ago, NunoLava1998 said:

If it's multi-threaded, Ryzen (and also Threadripper) will run much better than i.e. an Intel processor.

It likely is multi-threaded lol

That's not something that can be assumed. Many scientific programs and computations heavily utilize AVX and Intel has much high performance in that area, significantly so. Many HPC benchmarks have Intel with twice the per core performance.

 

Where AMD does well, mainly only with EPYC but can be the case with TR on Intel price equivalence, is in workloads that are memory bandwidth bound and use large memory datasets.

 

This is an area where you must know the workload being run and look at the performance benchmarks closest to that, or if possible of that actual workload.

[NunoLava1998]

a

3 minutes ago, karlo13 said:

Yes its heavily multi-threaded application but my concern is the program is kinda optimized to intel processor, it is due to amd haven't keep up before in these field. now i'm planning to get the upcoming ryzen 3000. currently in my 6700k setup, a resolution of 5km for a forecast of 5 days in WRF, takes a day or two to finish. since ryzen offers a very good perf/price ratio it is a very good choice, but i couldn't find any articles where ryzen was tested in these applications so im doubting to get it since my first priority is to get WRF run faster.

If it's heavily multi threaded, the 2950X (workstation) is 300% better in multicore, and the Ryzen 7 2700X (mainstream) is 100% (so 2x) better in multicore.

If it uses AVX, Intel could be better though.

[NunoLava1998]

1 minute ago, leadeater said:

That's not something that can be assumed. Many scientific programs and computations heavily utilize AVX and Intel has much high performance in that area, significantly so. Many HPC benchmarks have Intel with twice the per core performance.

 

Where AMD does well, mainly only with EPYC but can be the case with TR on Intel price equivalence, is in workloads that are memory bandwidth bound and use large memory datasets.

 

This is an area where you must know the workload being run and look at the performance benchmarks closest to that, or if possible of that actual workload.

Makes sense

[NunoLava1998]

Seems like it actually has issues on non Intel CPUs and is optimized for Intel, so I'd probably go with a 7920X or 7940X as an alternative to the TR 2920X and 2950X respectively.

 

If you want a less expensive upgrade, something like the 8700K should do okay

[bloodthirster]

IIRC, Zen2 has a lot higher FLOPS than Zen.  Ryzen 3000 and Rome are going to be pretty good for scientific computation.

[LogicalDrm]

You could poke Wendel about this. Level1 made testing with AMD Epyc. Most of it is about VMs, video/3D rendering and stuff like that. But I think they might be best bet for this. Also if you know youtuber who might work with the software, collab between the two could be a thing. Level1 did just recently some collab for Adobe Premier stuff.

[melahi1994]

Hello, I bought a Ryzen 9 3950x (16 cores and 32 threads) some time ago, but I was really surprzed when my old Xeon of 8 cores with 8 threads of second generation had a better time of calculus sith a lower clock speed. In a reference experiment, my old Xeon did the experiment in 12 hours and for the new Ryzen 9 the same experiment taked 18 hours. This was a great dissapointment. Now I'm going to try to sell this computer and build other with Intel core i9.‎

‎Therfore, for WRF model and numerical calculus is better intel. I think thtat the main problem with AMD is that it's processors are not optimised for some compilers.‎

[LogicalDrm]

4 hours ago, melahi1994 said:

Hello, I bought a Ryzen 9 3950x (16 cores and 32 threads) some time ago, but I was really surprzed when my old Xeon of 8 cores with 8 threads of second generation had a better time of calculus sith a lower clock speed. In a reference experiment, my old Xeon did the experiment in 12 hours and for the new Ryzen 9 the same experiment taked 18 hours. This was a great dissapointment. Now I'm going to try to sell this computer and build other with Intel core i9.‎

‎Therfore, for WRF model and numerical calculus is better intel. I think thtat the main problem with AMD is that it's processors are not optimised for some compilers.‎

Please don't revive dead threads. Locked.