The Current Radeon RX Vega 64 / Radeon VII Linux OpenCL Performance Against NVIDIA

[nuetzel]

@Michael

Stay tuned for some fresh tests with ROCm 2.5 as soon as I'm back in the office.

ROCm 2.6

[bridgman]

Yeah, some of the results are really odd. I'd expect latency to be lower than any kind of non-HBM solution - but its higher. Either I have a wrong assumption, or something is broken.

Latency in this case does not refer to memory latency but rather to kernel launch, ie the combination of SW code paths and HW responses when an app has loaded a shader program along with all the source data then triggers a "dispatch" command to the GPU to execute an instance of the shader program (aka kernel program) on every element in the destination array.

[bridgman]

I started writing a response along the lines of "the test suite has FP64 coverage; Michael ran the test during initial Radeon VII analysis" but when I went back and looked at the original article I didn't see the FP64 results.

They might be in another compute-specific article instead but I couldn't find them... not sure what happened there.

[vegabook]

With all due respect to the Phoronix Test Suite, this is a terminally flawed benchmark.

The only place where you're seeing AMD beating Nvidia is on your memory bandwidth benchmarks, where obviously HBM2 is going to beat GDDR6.

What is not being tested here, yet what is used all day long in scientific computing, is FP64. If you had proper FP64 testing, which is what all science that needs high precision uses (and remember that floating point errors propagate so this is a super important issue), then you'd see Radeon 7 beating all Nvidia consumer-card comers by 4x at least.

Phoronix Test Suite has a strong enough reputation now that this glaring error needs to be addressed, lest you pollute the internet with fake news (because you will get linked).

[Shevchen]

Yeah, some of the results are really odd. I'd expect latency to be lower than any kind of non-HBM solution - but its higher. Either I have a wrong assumption, or something is broken.

[bridgman]

AMD's numbers are weirdly all over the place. The Vega 64 isn't that much worse than the VII, yet in a lot of those tests it was really falling behind. Seems like some progress is going well, but also lots of room for improvement.

Not sure, but going from memory the tests with a big difference between Vega64 and Radeon VII (eg SHOC FFT) probably depend on memory bandwidth as much as compute throughput, and in that case Radeon VII should be ~2x as fast. Doesn't explain everything but I think it covers most of the seemingly odd results.

It would have been fun to see the FP64 results again even though we know what to expect

[schmidtbag]

AMD's numbers are weirdly all over the place. The Vega 64 isn't that much worse than the VII, yet in a lot of those tests it was really falling behind.

Seems like some progress is going well, but also lots of room for improvement.

[Peter Fodrek]

well, latest available to whoever is able to run these I'm a beggar here, I can't be too demanding... but generally latest available, or most "out-of-the-box" like official supported on Ubuntu 18.04 LTS maybe?

But AMD is only Platinum sponsor of openSUSE

Platinum sponsors of openSUSE

AMD







The combined forces of you, AMD and openSUSE are helping lead the Linux operating system community in 64-bit innovation.

Consistent with AMD's theme of customer-centric innovation, openSUSE provides developers with a tool to compile, release, and publish their software for the broad user audience. AMD is a proud Platinum Sponsor of openSUSE.

SUSE has consistently been among the first Linux distributions to support advanced features in AMD microprocessors such as AMD Direct Connect Architecture with NUMA memory architecture and HyperTransport technology, AMD Virtualization™ technology (AMD-V™), and AMD PowerNow!™ technology.

---

Silver Sponsors of openSUSE

https://en.opensuse.org/Sponsors


So there is not being too demanding when demanding best AMD support inside SUSE, or is it realy too demanding?

[TOMBOMBADIL]

Which ROCm version?
I have no ROCm 2.5 packages for OpenSuse Tumbleweed in in software.opensuse.org monitored repository, yet

well, latest available to whoever is able to run these I'm a beggar here, I can't be too demanding... but generally latest available, or most "out-of-the-box" like official supported on Ubuntu 18.04 LTS maybe?

[Peter Fodrek]

Thanks for this nice comparison. Looking forward to seeing Navi thrown into that parkour next month.
What would be great also, would be adding a test run for Tensorflow Cuda vs ROCm,

Which ROCm version?

I have no ROCm 2.5 packages for OpenSuse Tumbleweed in in software.opensuse.org monitored repository, yet


But

. The new compute unit is designed for 2x instruction rate by doubling of the scalar units and schedulers. There is also a single-cycle issue for the compute unit via Wave32 on SIMD32. Two adjacent compute unites can also be combined as a larger workgroup for resource pooling.

https://www.phoronix.com/scan.php?page=article&item=amd-zen2-rx5700-linux&num=1

May help more for openCL then gaming as I see

It looks like

AMD Graphics Core Next
...
--Similar to traditional CPUs (unlike the predecessors)

https://gcc.gnu.org/wiki/cauldron2016?action=AttachFile&do=view&target=Porting+GCC+to+GCN.pdf

linked in

An Early Port Of GCC To AMD's GCN Architecture
Written by Michael Larabel in GNU on 22 September 2016

https://www.phoronix.com/scan.php?page=news_item&px=GCC-For-AMD-GCN-Early

become more relevant for RDNA. Because RDNA looks more like superscalar CPU then any other GPU eve as of Navi related information released, yet