AMD Announces Radeon RX 7900 XTX / RX 7900 XT Graphics Cards - Linux Driver Support Expectations

[WannaBeOCer]

AMD said 2.7x, which I'm pretty sure is relative to the RX 6950XT. Given that fp32 is like 2.6x as much, that's not very impressive.


I'm sure they can do more than FSR3. Nvidia has a nice secondary business for their gaming GPUs, selling them as Tesla cards (although the Tesla name has been dropped) intended largely for general-purpose inference workloads, and I'll bet AMD wants to do the same.


Are the ray tracing units of their gaming GPUs also cut back, or you're saying this is why someone would pay more for a Nvidia card?

I agree that if someone really wanted a good ray tracing experience, they should get a Nvidia card. It certainly does have some allure for me, but luckily I don't currently have time to dabble with such things.

Not sure if Nvidia limits their RT cores performance in their GeForce lineup like Tensor cores. Nvidia set an artificial bandwidth cap on their GeForce cards Tensor cores bandwidth. Similar to how AMD/Nvidia would cap FP64 on consumer architectures that had the same of FP64 units as their server cards.

Nvidia separated their gaming architecture and server cards years ago. AMD already does the same with RDNA, CDNA. The difference is Nvidia still targets deep learning/content creation on their consumer cards. Again it sounds like the AI accelerators in RDNA3 are aimed at inference not training neural networks.

GP100, GP102
GV100, TU102
GA100, GA102
GH100, AD102

[coder]

Nvidia separated their gaming architecture and server cards years ago. AMD already does the same with RDNA, CDNA.

That's not accurate. Nvidia still sells plenty of the gaming GPUs for use in servers. What they did was to create the 100-tier of dies that's HPC-oriented. And then, the Titan cards were (up until Titan RTX) where they would cross over and sell it into the gaming market. However, that ended with the A100, which lacks the silicon to make a decent gaming GPU (but you could still allegedly put it on a 3D graphics card, unlike AMD's CDNA processors).

https://resources.nvidia.com/l/en-us-gpu

Basically, all of their products not ending in 100 are gaming GPUs just repurposed for server use.

The difference is Nvidia still targets deep learning/content creation on their consumer cards. Again it sounds like the AI accelerators in RDNA3 are aimed at inference not training neural networks.

I wouldn't equate training with content creation. AMD sells workstation cards that are usable for content creation.

https://www.amd.com/en/graphics/workstations

[WannaBeOCer]

That's not accurate. Nvidia still sells plenty of the gaming GPUs for use in servers. What they did was to create the 100-tier of dies that's HPC-oriented. And then, the Titan cards were (up until Titan RTX) where they would cross over and sell it into the gaming market. However, that ended with the A100, which lacks the silicon to make a decent gaming GPU (but you could still allegedly put it on a 3D graphics card, unlike AMD's CDNA processors).

https://resources.nvidia.com/l/en-us-gpu

Basically, all of their products not ending in 100 are gaming GPUs just repurposed for server use.


I wouldn't equate training with content creation. AMD sells workstation cards that are usable for content creation.

https://www.amd.com/en/graphics/workstations

The Titans used the workstation(Quadro )dies which lack FP64 units and HBM. They haven’t used a 100-die in their workstation(Quadro)/gaming GPUs since Maxwell except the Titan V which was special.

AMD’s can be used for training and content creation but perform slower due to lack of fixed function hardware. RT/Tensor cores accelerate OptiX which is the reason why lately there has been a massive gap.

https://github.com/ROCmSoftwarePlatf...ment-991679054

https://www.phoronix.com/review/blender-33-nvidia-amd

[Jabberwocky]

When I saw that the Ryzen 7000 series has a small GPU in the I/O die I had the thought, that their higher performance APUs will have a GPU die next to the CCD.
With the introduction of the RDNA 3 cards I am more certain about this:

The new AM5 APU will be a GCD and a CCD and an I/O die.
- the separation of the functionality follows that of their CPUs
- they don't have to create a different CCD to integrate the graphics
- APUs usually don't have 2 CCD-s anyway.
- Infinity fabric is for this purpose.

It seems logical to me.
What are your thoughts?

This seems like a no-brainer to me. For some reason AMD is releasing APUs loooooong after the discrete cards. Perhaps they are not making so much money on it, and or just not marketing it properly.

With the crazy GPU prices we saw (and still are seeing in some places) and inflation that many countries are seeing I would expect DDR5 APUs to sell really well. That's just my opinion.

[Melcar]

This seems like a no-brainer to me. For some reason AMD is releasing APUs loooooong after the discrete cards. Perhaps they are not making so much money on it, and or just not marketing it properly.

With the crazy GPU prices we saw (and still are seeing in some places) and inflation that many countries are seeing I would expect DDR5 APUs to sell really well. That's just my opinion.

It confuses me too. I get that bandwidth is the main bottleneck (which now should be fixed with DDR5), but with things like FSR they should be marketing the hell out of the APUs. And given that they had already made it clear that APUs will replace budget discrete cards it just seems like something obvious.

[coder]

This seems like a no-brainer to me.

The key question is how much additional cost the chiplet approach adds vs. a monolithic die. Another consideration is power-efficiency.

For some reason AMD is releasing APUs loooooong after the discrete cards.

Yeah, I wonder why the 5000-series didn't get RDNA. Maybe the timing just didn't work out for it, yet they had RDNA far enough along to integrate it into the Zen2-based consoles.

With the crazy GPU prices we saw (and still are seeing in some places) and inflation that many countries are seeing I would expect DDR5 APUs to sell really well.

Infinity Cache could make APUs really interesting.

[brucethemoose]

When I saw that the Ryzen 7000 series has a small GPU in the I/O die I had the thought, that their higher performance APUs will have a GPU die next to the CCD.
With the introduction of the RDNA 3 cards I am more certain about this:

The new AM5 APU will be a GCD and a CCD and an I/O die.
- the separation of the functionality follows that of their CPUs
- they don't have to create a different CCD to integrate the graphics
- APUs usually don't have 2 CCD-s anyway.
- Infinity fabric is for this purpose.

It seems logical to me.
What are your thoughts?

I'm late to this question, but monolithic dies are generally more efficient in laptops, as the inter die connections burn power (especially at low loads or idle). I think a GPU die would further exacerbate this issue since it eats so much bandwidth.

Desktops APUs are basically repurposed laptop parts.


TBH, I would never really want a "large" APU in a desktop, as older GPUs are so much cheaper. Now, if AMD made it truly massive like an XSX or something, that's different, but its never going to happen.

[parityboy]

Nvidia still sells plenty of the gaming GPUs for use in servers.

I wouldn't say that's accurate either. I think what NVIDIA did was to split their silicon into two basic lines: Acceleration and Visualization. Acceleration is the x100 class silicon used for pure server workloads like AI training & inference, database acceleration etc. Visualization uses traditional GPU designs (i.e. capable of doing actual 3D work and (possibly) driving a display). Of the two lines, I think visualization is the more diverse: server (VDI, cloud gaming), workstation (professional & scientific applications) and (con/pro)sumer (gaming, content creation, limited scientific applications). Through silicon cut-down, firmware and drivers each segment has features either disabled or limited in performance.

So I would say that where NVIDIA did indeed start out with gaming GPUs and repurposed them for more professional applications, nowadays the tables have turned: NVIDIA's visualization GPUs are multi-purpose and gaming is merely one capability - the architecture is much broader than that. Most of the silicon in AD102 or GA102 isn't needed for games.

[coder]

TBH, I would never really want a "large" APU in a desktop, as older GPUs are so much cheaper. Now, if AMD made it truly massive like an XSX or something, that's different, but its never going to happen.

Four years ago was the launch of an APU-based Chinese console that doubled as a desktop PC. Specs-wise, it fit right in between PS4 and PS5. It had GDDR5, which I was interested to see benchmarked and compared with a normal Zen APU. More details here:

• https://www.anandtech.com/show/13381...rid-fireflight
• https://www.anandtech.com/show/13163...-amd-ryzen-soc

[coder]

I wouldn't say that's accurate either. I think what NVIDIA did was to split their silicon into two basic lines: Acceleration and Visualization.

No... the workstation boards are probably the visualization products you have in mind, but they also sell rebadged gaming GPUs for servers. The non 100-numbered products here are all powered by gaming GPUs, but they have only passive cooling and no display outputs:

https://resources.nvidia.com/l/en-us-gpu

The way I'd characterize the split is that the 100-series are for training and HPC, while the lower-numbered products are for inferencing, video transcoding, desktop hosting, cloud gaming - basically, everything else you'd do with a GPU in a cloud server.

So I would say that where NVIDIA did indeed start out with gaming GPUs and repurposed them for more professional applications, nowadays the tables have turned: NVIDIA's visualization GPUs are multi-purpose and gaming is merely one capability - the architecture is much broader than that.

I'm using "gaming" as a short-hand. It does seem like gaming prowess is still their first and foremost concern, however.

Most of the silicon in AD102 or GA102 isn't needed for games.

Rubbish. If you include DLSS, then probably the only silicon underutilized by games is the NVDEC/NVENC block.