New Processor Choice

[0db]

That's pretty poor.

It's a Mame project. Lot of people complain about its slowness.
Which is the trick to speed-up the emulation of RISC-V CPUs?
Have you spent time on QEMU? And improved it specifically for RISC-V?

[brucehoult]

That's pretty poor.

It's a Mame project. Lot of people complain about its slowness.
Which is the trick to speed-up the emulation of RISC-V CPUs?

Mostly RISC-V is easy to emulate fast because it doesn't have condition codes. It's always a real b*tch to emulate condition codes between CPU families because they're never quite the same. But MIPS should be the same for that.

Have you spent time on QEMU? And improved it specifically for RISC-V?

I've done quite a lot of work on Spike (very slow ~100 to 200 MIPS interpretive RISC-V reference emulator) but not much on qemu. I submitted a few small bugfixes back in January 2017 when compressed instruction support was first being implemented. I think I just did those in an email thread, not even a proper pull request e.g. see https://github.com/qemu/qemu/commit/a52a1ca0666f0ff92e322007ca0beea29cad34b4

[NiHaoMike]

Generally, old computer emulators not only have to emulate the CPU but also the peripherals, in some cases they have to emulate old analog audio hardware in such a way as to sound "correct".

[Mattjd]

For a time I liked the idea of using Intel so I had access to the AVX-512 instruction set for development work. However, Intel have made such a mess of AVX-512 it doesn't seem to be worth bothering with right now.

I haven't made it through the entire thread but have you see this?


https://www.pugetsystems.com/labs/hpc/How-To-Use-MKL-with-AMD-Ryzen-and-Threadripper-CPU-s-Effectively-for-Python-Numpy-And-Other-Applications-1637/


AVX-512 wont be enabled but you do get some of the AVX2 instructions.

[coppice]

For a time I liked the idea of using Intel so I had access to the AVX-512 instruction set for development work. However, Intel have made such a mess of AVX-512 it doesn't seem to be worth bothering with right now.

I haven't made it through the entire thread but have you see this?


https://www.pugetsystems.com/labs/hpc/How-To-Use-MKL-with-AMD-Ryzen-and-Threadripper-CPU-s-Effectively-for-Python-Numpy-And-Other-Applications-1637/


AVX-512 wont be enabled but you do get some of the AVX2 instructions.

I think the recent AMD chips all fully support AVX2. AVX2 has been around for a long time. It certainly should be supported now.

[Mattjd]

For a time I liked the idea of using Intel so I had access to the AVX-512 instruction set for development work. However, Intel have made such a mess of AVX-512 it doesn't seem to be worth bothering with right now.

I haven't made it through the entire thread but have you see this?


https://www.pugetsystems.com/labs/hpc/How-To-Use-MKL-with-AMD-Ryzen-and-Threadripper-CPU-s-Effectively-for-Python-Numpy-And-Other-Applications-1637/


AVX-512 wont be enabled but you do get some of the AVX2 instructions.

I think the recent AMD chips all fully support AVX2. AVX2 has been around for a long time. It certainly should be supported now.

It has been, but for the longest time if you were using AMD and was using the Intel MKL instead of say openBLAS you saw you huge performance drops because Intel was checking if you were a "Geniune Intel CPU" and then going down a different code path, basically blocking AVX2 instructions and other optimizations.


https://www.extremetech.com/computing/302650-how-to-bypass-matlab-cripple-amd-ryzen-threadripper-cpus

I got friends at The MathWorks who helped pushed the fix for this into core Matlab so AMD users aren't getting screwed out of their money due to TMW using Intel MKL.

[coppice]

It has been, but for the longest time if you were using AMD and was using the Intel MKL instead of say openBLAS you saw you huge performance drops because Intel was checking if you were a "Geniune Intel CPU" and then going down a different code path, basically blocking AVX2 instructions and other optimizations.


https://www.extremetech.com/computing/302650-how-to-bypass-matlab-cripple-amd-ryzen-threadripper-cpus

I got friends at The MathWorks who helped pushed the fix for this into core Matlab so AMD users aren't getting screwed out of their money due to TMW using Intel MKL.

Intel have been playing those tricks for years, not just with AVX2. They did it with some of the older instruction set enhancements too.

[KungFuJosh]

I got the H115i Platinum cooler today. It's dead quiet, and the chip average temperatures are down by at least 10 degrees C.

[NiHaoMike]

It has been, but for the longest time if you were using AMD and was using the Intel MKL instead of say openBLAS you saw you huge performance drops because Intel was checking if you were a "Geniune Intel CPU" and then going down a different code path, basically blocking AVX2 instructions and other optimizations.

https://www.extremetech.com/computing/302650-how-to-bypass-matlab-cripple-amd-ryzen-threadripper-cpus

I got friends at The MathWorks who helped pushed the fix for this into core Matlab so AMD users aren't getting screwed out of their money due to TMW using Intel MKL.

Intel have been playing those tricks for years, not just with AVX2. They did it with some of the older instruction set enhancements too.

As AMD gains market share, I wonder if Intel's "tricks" end up backfiring when benchmarks start to show open source compilers like GCC greatly outperforming Intel CC.

I also wonder if the kernel can be patched to intercept and spoof requests to identify the CPU, such that it would be difficult or impossible for the application to tell that it's being spoofed. I suppose it's not as straightforward as it might seem since I haven't heard of any distros specifically "tuned" for AMD in that way.

[KungFuJosh]

Hmmm. Interesting results so far. AVS Video Converter did benefit from the extra threads. But so far I'm not impressed with Vegas Pro's performance. I upped the max threads to 24 in the settings, and it's still taking about 1 hour and 45 minutes to render an 8 minute 4K video.