Ryzen: Strictly technical

[IRobot23]

Well I tested games myself and I can tell you that on latest win10 update and AMD card basically 2+2 is same as 4+0.
Maybe sometimes games like some cache 2+2 will run better or ther way around, but there is only round 5%

 

[zlobster]

I use to call them stalls, but they are the same thing that other people call freezes for what we talk about.

I'm getting the exact same things while crunching S@H w/ BOINC and Ryzen 1700X. No OC on my CPU; all UEFI settings are left on AUTO (minus the XMP).

 

[amd6502]

https://semiaccurate.com/2017/05/17/amds-details-epyc-server-ambitions/

EPYC will remain on 14LPP . Ryzen and most likely ThreadRipper will get a update on 12LP as it really needs higher clock frequency to compete with Coffeelake. Since Pinnacle Ridge dies are desktop only AMD could tweak fabric speeds specifically for reducing memory latency and improving gaming performance.

Seems unlikely; the link is quite dated (before the news about 12nm ttbomr). The process improvement would allow for more efficient server parts, and thus higher margin models.

It'd be surprising to see PR as a very different part from SR---for example SR with all the server uncore area removed, shrunk (by nr of transistors and by higher density) as a pure desktop product, 8c/16t in a very small area like <140mm2. Maybe it'd even make sense. I'd be very surprised to see this happen though.

 

[DrMrLordX]

Seems unlikely; the link is quite dated (before the news about 12nm ttbomr). The process improvement would allow for more efficient server parts, and thus higher margin models.

It'd be surprising to see PR as a very different part from SR---for example SR with all the server uncore area removed, shrunk (by nr of transistors and by higher density) as a pure desktop product, 8c/16t in a very small area like <140mm2. Maybe it'd even make sense. I'd be very surprised to see this happen though.

Keep in mind that GF isn't that far away from offering 7nm LP. Epyc can make the jump to that node without there being much time between releases.

 

[The Stilt]

I recently updated my "test suite" slightly. The changes mostly affect the version of the test application used (i.e. more up to date) however, there were some workloads which I got rid of completely as I found them either bugged or otherwise not representing the performance on different platforms too well. The workloads which were removed from the original suite were Caselab Euler CFD program and Himeno poisson pressure calculator. These were replaced with open source image processing (filtering, crop & rescale) and audio conversion (Opus) tasks.

Is there any other workload (either a specific one or a type of) you would like to see?

There are few criterias the workload should meet:

- Open source
- Maintained (last code changes must not be older than 2 years)
- Windows compatible
- Can be compiled using either MSVC >= 2015 or GCC toolchain
- Has < 5 library dependencies or pre-requisites (unless the compiled libraries are provided with the source)

 

[xblax]

Openstreetmap Tile Rendering using PostgreSQL + PostGIS + Mapnik could be a interesting workload - should be very I/O heavy. But it's quite a hassle to set up.

Maybe RAW image processing using Darktable (it is Windows compatible now).

Another thing I'd be interested in is idle power consumption at the wall for the complete system using the same components for each (SSD, Power Supply, Graphics Card). Especially comparing X370 against X470 with Ryzen 8-core against Z370 with 8700k. With all power savings enabled and unnecessary components disabled in the bios.

 

[The Stilt]

Maybe RAW image processing using Darktable (it is Windows compatible now).

I've looked into Darktable, but the problem with it is that using CPU for the processing is pretty much pointless.
Tasks like this are up to order of magnitude faster when done on OpenCL capable GPU.

 

[HurleyBird]

Can you compare multi-threaded IPC with *and* without SMT? Wondering how much of AMD's MT (scaling) advantage is down to SMT.

 

[The Stilt]

Can you compare multi-threaded IPC with *and* without SMT? Wondering how much of AMD's MT (scaling) advantage is down to SMT.

I did it in the original review.
Based on the changes I don't think there is reason to expect that the SMT yield has changed in either way in the newer Zen iterations.
I could double check with few tests, but generally I don't think there is any difference.

 

[Drazick]

@The Stilt , Could you use Julia (Programming Language)?

You can create many great tests using it.
You can start with - Benchmark MATLAB & Julia for Matrix Operations.

Also using the Image Package one could test many Image Processing procedures.

Another option is LuaJIT with SciLua (Performance for JIT languages is very interesting).
On the project frontage you'll be able to find a built in test.

 

[The Stilt]

@The Stilt , Could you use Julia (Programming Language)?

You can create many great tests using it.
You can start with - Benchmark MATLAB & Julia for Matrix Operations.

Also using the Image Package one could test many Image Processing procedures.

Another option is LuaJIT with SciLua (Performance for JIT languages is very interesting).
On the project frontage you'll be able to find a built in test.

From the looks of it, it would be too time consuming to test and hard to maintain unfortunately

 

[xblax]

Do you already have an integer linear programming benchmark? That would be interesting.

MIPLIB http://miplib.zib.de/ contains multiple benchmark instances of linear programming problems.

SCIP http://scip.zib.de/ is a non-commercial solver and running these benchmark instances seems to be quite easy: http://scip.zib.de/doc/html/SHELL.php

Apparently it's single threaded only. But might still be interesting for IPC comparisons.

 

[CatMerc]

I did it in the original review.
Based on the changes I don't think there is reason to expect that the SMT yield has changed in either way in the newer Zen iterations.
I could double check with few tests, but generally I don't think there is any difference.

Actually since SMT helps hide latency, I expect SMT yield to drop a bit with Raven and Pinnacle Ridge due to the latency reductions in both caches and memory.

 

[The Stilt]

Do you already have an integer linear programming benchmark? That would be interesting.

MIPLIB http://miplib.zib.de/ contains multiple benchmark instances of linear programming problems.

SCIP http://scip.zib.de/ is a non-commercial solver and running these benchmark instances seems to be quite easy: http://scip.zib.de/doc/html/SHELL.php

Apparently it's single threaded only. But might still be interesting for IPC comparisons.

Latest version from MIPLib is from 2010 so that's not going to happen (due being obsolete).

Also I downloaded the package and couldn't even figure out what the thing is supposed to be as there were no binaries or C / C++ source in it.
Based on that it is probably irrelevant to consumers anyway.

Thanks for the suggestion thou.

 

[Olikan]

Since zen+ is supposed to have an updated brach predictor, it might be interesting to bench it... Queens?

 

[xblax]

Latest version from MIPLib is from 2010 so that's not going to happen (due being obsolete).

MIPLIB is only a collection of integer linear programming problem instances. To solve these problem instances a solver like SCIP (non-commerical) or Gurobi (commercial) is required. SCIP seems to be well maintained.

It's really hard to find 'consumer' workloads that can stress a CPU beside gaming and video encoding. But Ryzen CPUs are also (supposed to be) used for non 'consumer' workloads, e.g. in universities, companies and for students.

Maybe you should consider to use one of the popular javascript/webbrowser benchmarks. That should be somewhat representative for web browsing.

If you're interested in testing the entire AM4 Platform against similar Intel Platforms for workstation workloads you could try the SPECwpc benchmark suite (It's free to use) https://www.spec.org/gwpg/wpc.static/wpc-v2.1-info.html. This would however require similar storage and GPU in all test systems.

Another thing that would be really interesting is measuring system responsiveness for short computation tasks (with and without background load), like rendering a web page. But I don't know any any benchmarks for that.

 

[The Stilt]

Since zen+ is supposed to have an updated brach predictor, it might be interesting to bench it... Queens?

NQueen is already included.

 

[tamz_msc]

That would mean that CFD representation in your suite would be reduced, and I would like to see some more CFD. Would it be possible for you to include OpenFOAM?

 

[The Stilt]

That would mean that CFD representation in your suite would be reduced, and I would like to see some more CFD. Would it be possible for you to include OpenFOAM?

Not available for Windows.

 

[tamz_msc]

Not available for Windows.

No chance it'd work under WSL?

 

[The Stilt]

No chance it'd work under WSL?

It probably would, but that isn't very convinient for benchmarking.
When you have close to 30 different workloads and multiple platforms to test, the time spent to test each workload really adds up.

 

[maddie]

It probably would, but that isn't very convinient for benchmarking.
When you have close to 30 different workloads and multiple platforms to test, the time spent to test each workload really adds up.

Was wondering when you would state this. Appreciate what you do.

 

[tamz_msc]

It probably would, but that isn't very convinient for benchmarking.
When you have close to 30 different workloads and multiple platforms to test, the time spent to test each workload really adds up.

That's understandable.

The reason why I would have liked to have more CFD-based benchmarks is because they are one of the few workloads that are sensitive to memory latency, and given Ryzen's unique issues with latency, it would have been quite useful if there was a way of including them in your testing. Yes Euler3D does the job but I just wanted a little more sophistication.

 

[The Stilt]

That's understandable.

The reason why I would have liked to have more CFD-based benchmarks is because they are one of the few workloads that are sensitive to memory latency, and given Ryzen's unique issues with latency, it would have been quite useful if there was a way of including them in your testing. Yes Euler3D does the job but I just wanted a little more sophistication.

I think I just found the workload.

Eulerian variable density solver, based on Robert Bridson book.
Pure C++ solution, with no dependencies on external math libraries.

Adding the timing for benchmarking use was easy as well and the performance seems to be very constant.

Still need to try which compiler has the best performance across the different platforms thou.
Based on the code I'd say ICL 2018 will be superior, in which case some patching is required.

Anyway, I'll be adding it regardless if you like it or not

 

[tamz_msc]

I think I just found the workload.

Eulerian variable density solver, based on Robert Bridson book.
Pure C++ solution, with no dependencies on external math libraries.

Adding the timing for benchmarking use was easy as well and the performance seems to be very constant.

Still need to try which compiler has the best performance across the different platforms thou.
Based on the code I'd say ICL 2018 will be superior, in which case some patching is required.

Anyway, I'll be adding it regardless if you like it or not

So it involves compressibility in Eulerian flows? No problem. Thanks for including it.