AMD: Threadripper 3990X Isn't Better on Windows 10 Enterprise

[harlan4096]

AMD issues a statement.

AnandTech's recent review of the AMD Threadripper 3990X contained an interesting caveat: The website noted that Windows 10 Enterprise provided more performance than the Windows 10 Pro operating system that AMD recommended for testing. But AMD has told us that this isn't the case.

We've spent some time re-testing our Threadripper 3990X with a direct update from Windows 10 Pro to the Enterprise version and noticed little to no performance improvement outside of the expected standard deviation we experience with our benchmarks (test results below). We reached out to AMD to troubleshoot, which provided us with the following statement:

"AMD officially recommends Windows 10 Professional or Linux for the AMD Ryzen Threadripper 3990X. Higher editions/versions of Windows 10 confer no additional performance or compatibility benefits to the processor. We understand that this suggestion has been made in the media, but we believe this to be an error in testing that our team is presently diagnosing.”

AnandTech's rationale behind using Windows 10 Enterprise is pretty simple: AMD has pushed core counts up to 64 with the 3990X, and Simultaneous Multithreading (SMT) brings us up to 128 threads on the HEDT platform. Windows is simply ill-prepared for this massive jump in thread counts, so, as we noted in our own Threadripper 3990X review, it splits up the threads into two "processor groups." Due to the vagaries of the Windows scheduler, the operating system still sees the first 64 threads as one 'processor group,' while anything above that number of threads appears as a second group. In the case of the 3990X, that means threads 65-128 appear as their own processing group to the operating system.

Some applications can span across both groups, but many cannot, and you can either configure the 3990X with one NUMA node or two, adding the complexity and erratic performance trends of multiple NUMA nodes. We configured our system with the AMD-recommended single-NUMA node, as extensive testing with two nodes proved problematic, but we still experienced sub-par scaling in some workloads with both the 3990X and our server test platforms.

AnandTech noticed that while the Windows 10 Professional operating system supports 128 threads (right in the 3990X's sweet spot), the Pro for Workstation and Enterprise versions of Windows support 256 threads, which AnandTech surmised would lead to scheduler enhancements that could improve performance.

And in their test environment, it did. However, in our test environment, a direct upgrade of our Windows 10 Pro operating system did not lead to any performance improvements beyond what we would expect with normal run-to-run variation in the selection of workloads we targeted for re-testing. And they certainly aren't impactful enough to recommend spending the extra cash on the more-expensive Windows 10 Enterprise.

The Threadripper 3990X, like all third-gen Ryzen processors, is extremely sensitive to environmental factors, such as thermal dissipation, so even a poor (or slightly different) cooler mounting can introduce some (mostly minor) variability into our test results. As such, we ensured accuracy for our re-tests by using the same cooler and mount for these tests with the Windows 10 Pro image and the updated Enterprise image, so a few of these numbers are slightly different than those in our review.

AnandTech's purportedly incorrect test results could simply boil down to NUMA configurations of Windows versioning: AMD recommended testing with Windows 10 Pro OS build 18362.592 (or greater) due to unspecified optimizations that ensured maximum performance. As a standard manner of course, our Windows 10 test image remains constant over time (usually no longer than 9 months), but we've had to update more frequently due to a rash of patches for Intel security vulnerabilities. Every update to a new major OS revision requires retesting all comparable processors in our test pool, so we lag one distribution behind to assess any negative impacts from Microsoft's erratic updates.

Luckily, our test image was already on 18362.476, and after a series of regression tests with our update to 18362.592, we found that the slight update did not impact performance outside of our standard deviation for either Intel or AMD architectures, though it did provide better performance for the 64-core 3990X. That implies that whatever changes were made only impact super-high core-count processors, specifically.

It's possible that testing with an older version of Windows than the recommended distribution led to the performance gains AnandTech recorded, but there is currently no visibility into what led to the issue. It remains possible that there are a few scattered advantages for using the Enterprise version of Windows instead of the Pro, but AMD said that its testing, like ours, didn't find any meaningful performance improvements that would merit a step up to the Enterprise or Workstation versions of Windows 10.
...

Continue Reading