AMD Announces Navi 14 Based Radeon RX 5500 Series

[Venemo]

PCIe Atomics are available since Haswell on Intel (2013)

Laptop chips haven't supported PCIe atomics. So if you get an external GPU box with an AMD GPU, and connect that to your laptop, you won't be able to use the compute capability of the AMD card.

[Meteorhead]

bridgman Allow me to quickly outline a case-study / recreate the thought process behind a device purchase:

There were 2 recent product announcements with AMD HW inside: Surface Laptop 3 with custom Ryzen APUs and RX 5500M MSI laptops (again with Ryzen APU). What would I need a machine for? Casual gaming, develop OpenCL, SYCL and HIP applications, preferred OS is Windows, but I can bite my finger and use Linux as my daily driver. Let's break it down:

Surface Laptop 3

• Windows
  • OpenCL
    • Perf needs a lot more tuning
    • No CPU
  • SYCL
    • No ComputeCpp without standard IRs
    • No HipSYCL withuot ROCm on Windows
  • HIP
    • No HIP without ROCm on Windows
• Linux
  • AMDGPU-PRO
    • No HIP
    • No SYCL
  • ROCm
    • Requires 3rd-party package / build from source for APU support
    • Generally have no clue if it will work or not (today or even 2 releases from now)

MSI Alpha15

• Windows
  • Same things apply, but even worse perf for Navi
• Linux
  • AMDGPU-PRO
    • No HIP
    • No SYCL
  • ROCm
    • Still no ROCm support for Navi
    • Requires 3rd-party package / build from source for APU support

It's almost as if code is supposed to magically appear out of thin air. Code is developed on PCs, not mainframes. Even If I abandon all standards and lean into the vision that is HIP, I still have to put up with platforms and HW that is deemed "consumer" centric to have very limited or zero driver support. It HAS to run consumer grade stuff. When premium and brand new still doesn't make the cut for developing SW, there is big trouble.

Segmenting the software stack to pro-cad / hpc-compute / consumer-gaming along with platform support was one of the most painful decisions from an end-user perspective. The problem is, that the support matrix is sparse.

Minimally I need 2 machines, both of them dual-boot, or a thrid machine with triple-boot setup.

Acer Helios 500 (Ryzen 2700 and Vega56)

• Windows boot for gaming and an abandoned compute stack for vanilla OpenCL development
• Ubuntu 16.04.3 boot with outdated 17.40 AMDGPU-PRO for SYCL development
• Ubuntu 18.04 boot with mainline ROCm drivers for HIP development and OpenCL profiling

Or I can buy something else and destin myself for remote development for the coming 3-4 years, primarily due to no HIP with other vendors. This is shaping up to be the most compelling scenario.

[tuxd3v]

Nvidia does have other problems, their hardware generally does not age well. As soon as a card gets into legacy support, things get crappy rather quickly, with no support for newer X11 versions, no support for newer technology past X11, compiler hazzles, kernel update woes...

I disagree with that..
NVidia is proprietary, and that for us( me included ) is bad, but I do need to give them credits for the longevity stacks they have..

[tuxd3v]

It's the only thing that needs an upgrade. E.g. the 2011 2700k paired with 16 gigs of RAM and 1TB SSD is perfectly fine for gaming

Exactly,
Some times because the actual card doesn't have the performance needed, or because it failed..
Some times, its a server machine, or a test machine you have, when is needed new graphics card( for gaming or for Compute work ), its nice to have them available, and supporting the hardware..

[Hibbelharry]

Laptop chips haven't supported PCIe atomics. So if you get an external GPU box with an AMD GPU, and connect that to your laptop, you won't be able to use the compute capability of the AMD card.

That's mostly a niche, isn't it? Bulldozer didn't sell well, but i bet external GPUs used to do OpenCL, attached to now legacy notebooks, are quite uncommon. Like 10 per continent.

I disagree with that..
NVidia is proprietary, and that for us( me included ) is bad, but I do need to give them credits for the longevity stacks they have..

I've had the opposite experience, quite often. I once had a notebook which fell into this non-support case, I did use a nvidia-based HTPC at home, and do try to keep PCs in our company working. We do software development in a mostly linux environment, which is CPU and RAM intensive, but not taxing any GPU. There are quite some Core i5 machines where I did remove that nvidia crap for failing after updates. It's just much more efficient if I swap out those bastard low spec Geforce 500/600/700 cards for an AMD based card than fiddling around with patching dkms packages over and over again.

[artivision]

https://trac.ffmpeg.org/wiki/Hardware/VAAPI

https://wiki.libav.org/Hardware/vaapi

At least with ffmpeg, the key appears to be "-vaapi_device /dev/dri/renderD128 -vf 'format=nv12,hwupload' -c:v h264_vaapi".

I mostly used emby for this so haven't done much of the command line myself.

EDIT: Use vainfo to make sure your card supports the h264 profile you're targeting.

For example my 7770 only supports the following:

vainfo: Supported profile and entrypoints
VAProfileMPEG2Simple : VAEntrypointVLD
VAProfileMPEG2Main : VAEntrypointVLD
VAProfileNone : VAEntrypointVideoProc

Tried with Polaris and its terrible. The default gives me 2.6x speed, big size for selected Mbps and cannot play a half file. I was expecting like a 10x for a 5.5Tflops Gpu.

[fuzz]

Tried with Polaris and its terrible. The default gives me 2.6x speed, big size for selected Mbps and cannot play a half file. I was expecting like a 10x for a 5.5Tflops Gpu.

I agree it's not great, but as far as I know it doesn't use all those Tflops. It uses the video encoding chip on the card -- Video Coding Engine (VCE)

[skeevy420]

Tried with Polaris and its terrible. The default gives me 2.6x speed, big size for selected Mbps and cannot play a half file. I was expecting like a 10x for a 5.5Tflops Gpu.

Can confirm that with my RX 580:

ffmpeg -hwaccel vaapi -hwaccel_device /dev/dri/renderD128 -hwaccel_output_format vaapi -i not_a_downloaded_legion_episode.mkv -c:v h264_vaapi -level 4.2 -profile high -b:v 2M -c:a copy output.mp4

To convert video for the PS4 gives me:

frame= 1711 fps= 61 q=-0.0 Lsize=   19316kB time=00:01:12.15 bitrate=2193.0kbits/s speed=2.55x

[artivision]

Any way to use OpenCL? My attempt with Ffmpeg x264 CL resulted to 1.7x speed and full Cpu load.

[arQon]

It's the only thing that needs an upgrade. E.g. the 2011 2700k paired with 16 gigs of RAM and 1TB SSD is perfectly fine for gaming https://ark.intel.com/content/www/us...-3-90-ghz.html

Jep. I don't have anything QUITE that old - by a whole year :P - but an Ivy-era i7 is still more than enough for almost anything even now, CPU-wise. With the hype of 4K displays in particular over the last few years, there also may be quite a few casual gamers who need to push a lot more pixels than before but don't have any higher geometry demands.