Are there any benchmarks for LC0 on the new AMD 7nm GPUs?

[Dann Corbit]

I am guessing that the Nvida cards win, because of the single cycle matrix operations

[smatovic]

I am guessing that the Nvida cards win, because of the single cycle matrix operations

Hmm, according to the web AMD still works on proper OpenCL drivers for Navi:

https://www.anandtech.com/show/14618/th ... -review/13
https://www.phoronix.com/scan.php?page= ... uxgl&num=1

Assuming someone would optimize LC0 for AMD gpus and considering that AMD has
doubled FP16 throughput, and LC0 uses only ~30% of the TFLOPs the RTX
TensorCores provide, then the race should be won by Nvidia by an factor of two
or so for the current high-end gpu series.

--
Srdja

[Dann Corbit]

AMD support for their own GPU cards is not good and it has not been good.
I had a $799 card that stopped working when they upgraded their software (it simply killed OpenCL completely, though you could still play video games with it,

Since I do not play video games, it did not make me happy,

A wonderful card is not so wonderful with no software to drive it.

Can't analyze mersenne prime candidates for possible facotors.
Can't fold proteins
Can't do chess,.
But you can shoot down space invaders.
Gee, thanks

[mar]

But you can shoot down space invaders.
Gee, thanks

Do you realize what a game engine does? Obviously not, because you'd show more respect towards game engine programmers otherwise.
"rendering triangles" barely scratches the surface. And you have 16 milliseconds for everything, much less when doing VR.

This reminds of my former superior at work who also disdained games, yet he had to download a code snippet to compute
a plane equation from 3 points, which I found rather amusing considering he was a PhD.

Games were the primary driving factor towards early GPUs, and thanks to them we have GPGPUs today.

As for drivers, regressions happen and this isn't specific to AMD, I remember at least two issues with new nVidia drivers in the past,
but we only use GPUs for rendering, which isn't of course as noble as playing chess or folding proteins (by playing fold-it I assume , considering that G still stands for Graphics

[Dann Corbit]

But you can shoot down space invaders.
Gee, thanks

Do you realize what a game engine does? Obviously not, because you'd show more respect towards game engine programmers otherwise.
"rendering triangles" barely scratches the surface. And you have 16 milliseconds for everything, much less when doing VR.

This reminds of my former superior at work who also disdained games, yet he had to download a code snippet to compute
a plane equation from 3 points, which I found rather amusing considering he was a PhD.

Games were the primary driving factor towards early GPUs, and thanks to them we have GPGPUs today.

As for drivers, regressions happen and this isn't specific to AMD, I remember at least two issues with new nVidia drivers in the past,
but we only use GPUs for rendering, which isn't of course as noble as playing chess or folding proteins (by playing fold-it I assume , considering that G still stands for Graphics

Believe it or not, I worked as a graphics programmer for several years.
(Not games, I wrote font systems, presentation graphics packages, and geographic information systems).
I appreciate the wonder of rendering triangles, ray tracing, etc.
It is just that I bought the cards for a purpose. I am past the age of Qbert and Xevius. I want to do serious math.

I do not disdain games. I recognize that it is a multi-billion dollar industry. And I know several adult men who enjoy playing them.
But *I* personally have no interest whatsoever in playing them.

[dragontamer5788]

AMD support for their own GPU cards is not good and it has not been good.
I had a $799 card that stopped working when they upgraded their software (it simply killed OpenCL completely, though you could still play video games with it,

Since I do not play video games, it did not make me happy,

A wonderful card is not so wonderful with no software to drive it.

Can't analyze mersenne prime candidates for possible facotors.
Can't fold proteins
Can't do chess,.
But you can shoot down space invaders.
Gee, thanks

AMD's OpenCL implementation was never very good.

1. Constant regressions as new drivers led to grossly different OpenCL Compiler bugs. One driver version may infinite-loop on compiling your OpenCL code, while a 2nd driver version may crash your code.

2. The optimizer output grossly incorrect code. -O1 or -O2 was completely useless. -O0 was the only sane way to use AMD's OpenCL compiler. Fortunately, AMD GPUs have so much girth that the crypto-coin community managed to get everything done without the compiler-optimizer working (yeah, GPUs have THAT much compute power available).

3. OpenCL 2.0 support was shoddy. A few important features were implemented (SVM, Device-side kernel queue, etc. etc.) but important features like... a debugger... never was fully implemented. OpenCL 1.2 is basically the last version of OpenCL AMD really supported well (aside from the compiler bugs).

-------------

ROCm still has compiler bugs on OpenCL. I'm having moderate amounts of success with their C++ HCC code, but then AMD just deprecated that interface. C++ HIP is very similar to CUDA, so I've been using it with moderate amounts of success.

ROCm HIP is how I'd recommend people do GPU programming on AMD these days. Otherwise... yeah... AMD's software stack leaves much to be desired.

The hardware is great though. Its just a complete mess on the software side.

[DustyMonkey]

AMD bet on OpenCL but it "lost" so its no wonder their investment in it became sketchy.

Their OpenCL 1.0 SDK would seamlessly mix CPU and GPU which was what I was after early on, as I had one of the first 6 core consumer chips available (The Thuban line of Phenom II processors) - this was when the nVidia 8800GT was killing it (which I also had) so CPUs actually had somewhat comparable grunt.

But a few versions of the SDK later and the CPU support was dropped, after even Apple lost any desire to keep up work on OpenCL.

[smatovic]

Yes, it's sad.
Apple recently announced to push their own Metal API instead of OpenCL, and I
think it is getting even worse with Intel entering the discrete gpu market in
2020/2021, they already push their OneAPI thing, one API to rule them all...
I hope the vendors will at least support OpenCL 1.x to keep old code running.

--
Srdja

[dragontamer5788]

Yes, it's sad.
Apple recently announced to push their own Metal API instead of OpenCL, and I
think it is getting even worse with Intel entering the discrete gpu market in
2020/2021, they already push their OneAPI thing, one API to rule them all...
I hope the vendors will at least support OpenCL 1.x to keep old code running.

--
Srdja

With NVidia, AMD, and Intel supporting OpenCL 1.2 so much, it seems unlikely that anyone would move off of OpenCL 1.2.

Intel actually has one of the more interesting OpenCL implementations, for Altera FPGAs: https://www.intel.com/content/www/us/en ... rview.html, as well as some interesting hardware such as the Xeon CPU +Altera FPGA chip: https://itpeernetwork.intel.com/intel-p ... #gs.oxsqee

Its really NVidia that stopped implementing OpenCL. OpenCL was first specified in 2013, and NVidia barely has any of those features implemented. AMD has a proper excuse: the company simply couldn't afford to implement many OpenCL 2.0 features at that time (AMD was well on its way to bankruptcy in 2013 and 2014).

AMD has started over with their ROCm compiler stack, and they're clearly going for OpenCL 2.0 support. But Blender's OpenCL code doesn't work on ROCm yet, and the ROCm OpenCL compiler seems to crash on a lot of other example code. So OpenCL is new and very experimental right now... features and stability continues to improve, but AMD is starting from scratch again, so it may take a while before that stabilizes.

Intel's implementation is the most decent. Intel has OpenCL 2.0 across Xeon+FPGA, FPGA, iGPU and probably will support OpenCL2.0+ on discrete chips.

[dragontamer5788]

Its really NVidia that stopped implementing OpenCL. OpenCL was first specified in 2013, and NVidia barely has any of those features implemented. AMD has a proper excuse: the company simply couldn't afford to implement many OpenCL 2.0 features at that time (AMD was well on its way to bankruptcy in 2013 and 2014).

I mean OpenCL 2.0 across this paragraph. OpenCL 1.2 has good support across the companies (with AMDPRO drivers not with ROCm yet). On NVidia... OpenCL 1.2 is slower than CUDA but it is at least supported.