XPost: alt.comp.hardware.pc-homebuilt

On Wed, 2/5/2025 4:01 AM, David Chmelik wrote:

Do any graphics processing unit (GPU) or display/video/graphics card companies admit/market GPUs or cards to be used for serious reasons:
science operating systems (OS, UNIX/GNU/Linux) CUDA & OpenCL (I know they
do for ones more expensive than Nvidia GeForce RTX 5090 OC, AMD Radeon RX 7900 XTX & forthcoming 9000s, but asking about these & low-to-mid-range)?

We have AMD, ASUS, GigaByte, Nvidia hardware. One or more companies won't answer about CUDA & OpenCL nor safe temperature ranges nor configuration software (either ASUS or GigaByte makes their own... the other doesn't nor knows defaults) saying 'graphics cards are for gaming' (or that GPU companies' software shows safe temperatures despite cards differ) which is ridiculous: 'hardcore gamers' probably only use small percentage of (powerful) graphics cards... first of all, they're for graphics (computational geometry/graphics computer programming/science to make
games, (photo)graphics/art/computer-aided-design, video, etc. software)
but they won't say this first. Now most (powerful) graphics cards (other than fair number of average/office/non-gaming or casual/rare gaming usage) may be in (UNIX/GNU/Linux) science (super)computers, whether modelling/ research (like/using BOINC.Berkeley.edu), artificial intelligence (AI) or small-to-medium science (AI) workstation P(S)C usage... and many still cryptocurrency mining. I'll call all this 'number-crunching'.

Seems some these companies outsourced customer support to third-world countries unaware scientific/non-desktop/-gaming usage exists. Some put 'gaming' in card names, despite maybe many/most are for users or number- crunching. I asked one about OpenCL and they referred me to '"our" forum'
on gitlab.freedesktop.org , which isn't their site (it's for graphical
user interfaces (GUI) such as the X Window System (X), etc. for science
OS) let alone a forum, and CUDA & OpenCL don't need desktop: typically
start on command-line/console/shell/terminal (by 'terminal' I mean not graphical user interface (GUI)) to compute in background (screen/tmux).
The customer support agent apparently thought OpenCL misspelled OpenGL.

Lately I asked AMD if GNU/Linux AMDGPU-Pro OpenCL still works on Radeon RX Vega... I tried unofficial close variants (Debian/Devuan, Ubuntu/Neon GNU/ Linuxes) which used to work but now some don't: too altered, so should I
try something listed in amdgpu-install (Debian, Mint, *Ubuntu) rather than variants, or was Vega dropped? I don't think AMDGPU-Pro depends on
systemd (oriented to desktops, not traditional servers) so since Devuan doesn't alter anything else, OpenCL should work, and KDE Neon claims they just add newer X/KDE desktop environment (DE) without altering underlying Kubuntu (an official Ubuntu) though I've found sometimes false (upgrades
to experimental/development/testing KDE graphics libraries too new for AMDGPU-Pro with older cards requiring dkms, despite OpenCL doesn't use
dkms, can completely break display many ways, which we need).

Is it best to ask AMD, ASUS, GigaByte, Nvidia, try other companies (Intel GPUs, who else for cards)? I still like ASUS for having most display-
ports, but lately prefer GigaByte for best price with longer warranty
(than ASUS Tuf).

On AMD, I would start with ROCm. ROCm is available immediately on 7900XTX, 7900XT, and the Radeon VII you mention (and that card is drifting out of support and soon to be deprecated). Apparently there are moves afoot,
to include some of the lesser RDNA cards, so the supported list could
grow in future.

https://en.wikipedia.org/wiki/ROCm

It's hard for me to say whether OpenCL is still a separate development or
not. I don't know if it is packaged separately.

One of the sites that was capable of writing good article was Anandtech.
Future has laid off the staff there. Their server remains running,
there is a user forum, but in terms of writing new technical articles,
there is no one there now to do that. Anand, the original owner of the site, went off to work for Apple, and Ryan Smith took over running the site after that. And he was laid off. They could have written an article explaining
the block diagram of the software product.

But other than that, the marketing efforts are awful, and don't
tell end users anything about why you would buy a product. There is
Quadro with certified drivers for CAD work. The full capabilities of
OpenGL are enabled on a Quadro, whereas if you do OpenGL CAD
work on a gamer card, it "slows down after around 50 objects".
I expect the driver has the ability to consider what the program
name is which just started, and control the level of acceleration
available to that program.

NVidia has TensorCore, but what is that and how is it implemented ?
Is it just a shader program ? Or a logic block ?

(This covers some of it, but is likely a bit game-oriented.)

https://www.techpowerup.com/299130/nvidia-ada-ad102-block-diagram-and-new-architectural-features-detailed

Even readouts like this, do not map to whatever software
stack exists today. The card (ten years old?) on the left,
has no Ray Tracing, and is about to out of support. The device
on the right, is the iGPU in an AMD 5600G processor.

[Picture]

https://i.postimg.cc/59v3Fb8B/Techpowerup-GPUZ-2-62.gif

bullwinkle@TIKTACK:~$ inxi -F
System:
Host: TIKTACK Kernel: 6.8.0-51-generic arch: x86_64 bits: 64
Desktop: Cinnamon v: 6.4.6 Distro: Linux Mint 22.1 Xia
Machine:
Type: Desktop Mobo: Micro-Star model: MPG B550 GAMING PLUS (MS-7C56) v: 1.0
UEFI: AMI date: 07/13/2024
CPU:
Info: 6-core model: AMD Ryzen 5 5600G with Radeon Graphics bits: 64
Graphics:
Device-1: AMD Cezanne [Radeon Vega Series / Radeon Mobile Series]
driver: amdgpu v: kernel
Display: x11 server: X.Org v: 21.1.11 with: Xwayland v: 23.2.6 driver: X:
loaded: amdgpu unloaded: fbdev,modesetting,vesa dri: radeonsi gpu: amdgpu
resolution: 1920x1080~75Hz
API: EGL v: 1.5 drivers: radeonsi,swrast platforms: x11,surfaceless,device
API: OpenGL v: 4.6 compat-v: 4.5 vendor: amd mesa v: 24.0.9-0ubuntu0.3
renderer: AMD Radeon Graphics (radeonsi renoir LLVM 17.0.6 DRM 3.57
6.8.0-51-generic)

Driver Manager says it "needs no drivers", which means AMDGPU is
already running it after the OS was installed. It was the default driver. (Whereas on NVidia, you have Nouveau versus NVidia from Driver Manager.)

Paul

--- SoupGate-Win32 v1.05
* Origin: fsxNet Usenet Gateway (21:1/5)

XPost: alt.comp.hardware.pc-homebuilt

On Wed, 2/5/2025 6:30 AM, Paul wrote:

On Wed, 2/5/2025 4:01 AM, David Chmelik wrote:

Do any graphics processing unit (GPU) or display/video/graphics card
companies admit/market GPUs or cards to be used for serious reasons:
science operating systems (OS, UNIX/GNU/Linux) CUDA & OpenCL (I know they
do for ones more expensive than Nvidia GeForce RTX 5090 OC, AMD Radeon RX
7900 XTX & forthcoming 9000s, but asking about these & low-to-mid-range)?

We have AMD, ASUS, GigaByte, Nvidia hardware. One or more companies won't >> answer about CUDA & OpenCL nor safe temperature ranges nor configuration
software (either ASUS or GigaByte makes their own... the other doesn't nor >> knows defaults) saying 'graphics cards are for gaming' (or that GPU
companies' software shows safe temperatures despite cards differ) which is >> ridiculous: 'hardcore gamers' probably only use small percentage of
(powerful) graphics cards... first of all, they're for graphics
(computational geometry/graphics computer programming/science to make
games, (photo)graphics/art/computer-aided-design, video, etc. software)
but they won't say this first. Now most (powerful) graphics cards (other
than fair number of average/office/non-gaming or casual/rare gaming usage) >> may be in (UNIX/GNU/Linux) science (super)computers, whether modelling/
research (like/using BOINC.Berkeley.edu), artificial intelligence (AI) or
small-to-medium science (AI) workstation P(S)C usage... and many still
cryptocurrency mining. I'll call all this 'number-crunching'.

Seems some these companies outsourced customer support to third-world
countries unaware scientific/non-desktop/-gaming usage exists. Some put
'gaming' in card names, despite maybe many/most are for users or number-
crunching. I asked one about OpenCL and they referred me to '"our" forum' >> on gitlab.freedesktop.org , which isn't their site (it's for graphical
user interfaces (GUI) such as the X Window System (X), etc. for science
OS) let alone a forum, and CUDA & OpenCL don't need desktop: typically
start on command-line/console/shell/terminal (by 'terminal' I mean not
graphical user interface (GUI)) to compute in background (screen/tmux).
The customer support agent apparently thought OpenCL misspelled OpenGL.

Lately I asked AMD if GNU/Linux AMDGPU-Pro OpenCL still works on Radeon RX >> Vega... I tried unofficial close variants (Debian/Devuan, Ubuntu/Neon GNU/ >> Linuxes) which used to work but now some don't: too altered, so should I
try something listed in amdgpu-install (Debian, Mint, *Ubuntu) rather than >> variants, or was Vega dropped? I don't think AMDGPU-Pro depends on
systemd (oriented to desktops, not traditional servers) so since Devuan
doesn't alter anything else, OpenCL should work, and KDE Neon claims they
just add newer X/KDE desktop environment (DE) without altering underlying
Kubuntu (an official Ubuntu) though I've found sometimes false (upgrades
to experimental/development/testing KDE graphics libraries too new for
AMDGPU-Pro with older cards requiring dkms, despite OpenCL doesn't use
dkms, can completely break display many ways, which we need).

Is it best to ask AMD, ASUS, GigaByte, Nvidia, try other companies (Intel
GPUs, who else for cards)? I still like ASUS for having most display-
ports, but lately prefer GigaByte for best price with longer warranty
(than ASUS Tuf).

As for things like FP64, ordinary gamer video cards have low ratios
(1 in 32 cores does FP64). The more expensive cards, the CAD ones,
enable more FP64 units.

The exception to the rule, was a certain NVidia Titan card, which
had a high percentage of FP64, yet it was a consumer card. Those were
prized items at one time, and their used price might have been
higher than their new price.

You can check the build options here, to see what "desirable"
configurations look like.

https://www.pugetsystems.com/workstations/threadripper/t140-xl/

One of the reasons you can't have four of the "top" accelerators,
is the power supply runs out of legs. And it is possible the cooling
becomes an issue (can't vent the hot air fast enough).

I have no idea, how AI launchers treat the accelerator cards,
whether they are clever with the card usage, or, only one card is
used and the other three ignored. In the past, there was an issue
with bus master transfers, where one PCIe card could bus master transfer
right into the VRAM on a second card. Whether anything needs to be
present in the CPU PCIe Hub for that, I don't know. On some old NVidia
designs, a separate switching chip added to the motherboard, seemed
to have the required PCIe properties for that style of operation.

And there exist, even larger power supplies, but they don't
fit in a conventional computer case.

Paul

--- SoupGate-Win32 v1.05
* Origin: fsxNet Usenet Gateway (21:1/5)

XPost: alt.comp.hardware.pc-homebuilt

On Wed, 5 Feb 2025 06:30:54 -0500, Paul wrote:

It's hard for me to say whether OpenCL is still a separate development
or not.

OpenCL seemed like a nice idea for a vendor-independent standard, but
there was never enough competition in the GPU market for it to succeed.

I think it has now been revamped and renamed to “SYCL”, but I don’t think it is achieving much more success.

Market-dominant vendors (*cough* Nvidia *cough*) see no benefit in
allowing competition with their own proprietary products.

--- SoupGate-Win32 v1.05
* Origin: fsxNet Usenet Gateway (21:1/5)

XPost: alt.comp.hardware.pc-homebuilt

On Thu, 2/6/2025 4:05 PM, Lawrence D'Oliveiro wrote:

On Wed, 5 Feb 2025 06:30:54 -0500, Paul wrote:

It's hard for me to say whether OpenCL is still a separate development
or not.

OpenCL seemed like a nice idea for a vendor-independent standard, but
there was never enough competition in the GPU market for it to succeed.

I think it has now been revamped and renamed to “SYCL”, but I don’t think
it is achieving much more success.

Market-dominant vendors (*cough* Nvidia *cough*) see no benefit in
allowing competition with their own proprietary products.

I think OpenCL runs on all three platforms. Now
that Intel makes stuff powerful enough to benefit
from it.

Even NVidia supports OpenCL, so you should be able to
write an application, and have multiple hardware assist
options for it.

But we won't know about the good works of these companies,
unless someone writes technical articles about them.
We need to know what the arch is inside, and why
a certain product is worth buying.

https://www.techpowerup.com/gpu-specs/geforce-rtx-5090.c4216

Shading Units 21760
TMUs 680
ROPs 176
SM Count 170
Tensor Cores 680
RT Cores 170

L1 Cache 128 KB (per SM)
L2 Cache 96 MB

Pixel Rate 423.6 GPixel/s
Texture Rate 1,637 GTexel/s
FP16 (half) 104.8 TFLOPS (1:1)
FP32 (float) 104.8 TFLOPS
FP64 (double) 1.637 TFLOPS (1:64) <=== more enabled, on CAD version of card

As a customer, it's your job to figure out what those
mean, and how those could help you.

For CUDA programs, the dev has to "tune" the resource
settings for their program, so it "runs good" on "your board".
Naturally, some programs offered, are tested on only one board
and it won't be the board you have.

The architecture on the cards, isn't exactly the same, but at
least on this card, it has its first Ray Tracing cores. And as
far as I know, Ray Tracing cores are good for gaming, and not
for BOINC or AI. Tensor Cores might be good for AI.

https://www.techpowerup.com/gpu-specs/radeon-rx-7900-xtx.c3941

Shading Units 6144
TMUs 384
ROPs 192
Compute Units 96
RT Cores 96

L0 Cache 64 KB per WGP
L1 Cache 256 KB per Array
L2 Cache 6 MB
L3 Cache 96 MB

Pixel Rate 479.6 GPixel/s
Texture Rate 959.2 GTexel/s
FP16 (half) 122.8 TFLOPS (2:1)
FP32 (float) 61.39 TFLOPS
FP64 (double) 1.918 TFLOPS (1:32)

and of course we should list that card, because at the
moment, there are none left for sale.

Good luck,

Paul

--- SoupGate-Win32 v1.05
* Origin: fsxNet Usenet Gateway (21:1/5)

XPost: alt.comp.hardware.pc-homebuilt

On Fri, 2/7/2025 6:28 AM, Mr. Man-wai Chang wrote:

On 7/2/2025 5:05 am, Lawrence D'Oliveiro wrote:

On Wed, 5 Feb 2025 06:30:54 -0500, Paul wrote:

It's hard for me to say whether OpenCL is still a separate development
or not.

OpenCL seemed like a nice idea for a vendor-independent standard, but
there was never enough competition in the GPU market for it to succeed.

I think OpemCL was about competing against Micro$oft DirectX. But how many games need to be portable these days? :)

OpenCL C=Computing
OpenGL G=Graphics

Paul

--- SoupGate-Win32 v1.05
* Origin: fsxNet Usenet Gateway (21:1/5)