GPU Architecture Deep Dive: Nvidia Ada Lovelace, AMD RDNA 3 and Intel Arc Alchemist

[neeyik]

Three vendors. Three architectures. Three approaches to GPU design. Join us as we dive into their semiconductor hearts and see how Ada Lovelace, Arc Alchemist, and RDNA 3 contrast and compare.

Read the full article here.

https://www.techspot.com/article/2570-gpu-architectures-nvidia-intel-amd/

 

[scavengerspc]

Nick, I can't even imagine the work that went into this. All I can envision, is all the stuff I would have to learn to understand it.... And all the stuff I would have to learn just to understand that stuff. Etc, etc.

So let me say kudos for the work.

Now, on with the tech adept folks with something to say that is actually intelligent!

 

[neeyik]

So let me say kudos for the work.

Many thanks! This one took a lot longer than the last one I did, partly because of now having to cover Intel too, but mostly because none of the three vendors did a great job of saying much about their new architectures.

Compared to the RDNA 2 launch, AMD provided relatively little detail, so rather than wait for the next set of developer docs to be released, I had to sift through a whole raft of other material, doing a fair amount of cross-referencing, to fully understand all of the changes. Even then, there's still quite a bit that's relatively unknown about RDNA 3.

Intel actually provides quite a lot of detail but it's a mammoth task hunting it all down and comparing the varying sources to make sure one isn't outdated or the like. The full architectural details are still somewhat murky, such as the load/store system.

For Ada, Nvidia actually said remarkably little, compared to Ampere and Turing -- most of the focus was on DLSS 3 and its related technology, which isn't surprising, of course. Mind you, they still don't say a lot about the insides compared to their much older architectures; even with a developer account, it's like trying to find a needle in a planet-sized haystack.

 

[kinetix]

Many thanks! This one took a lot longer than the last one I did, partly because of now having to cover Intel too, but mostly because none of the three vendors did a great job of saying much about their new architectures.

Compared to the RDNA 2 launch, AMD provided relatively little detail, so rather than wait for the next set of developer docs to be released, I had to sift through a whole raft of other material, doing a fair amount of cross-referencing, to fully understand all of the changes. Even then, there's still quite a bit that's relatively unknown about RDNA 3.

Intel actually provides quite a lot of detail but it's a mammoth task hunting it all down and comparing the varying sources to make sure one isn't outdated or the like. The full architectural details are still somewhat murky, such as the load/store system.

For Ada, Nvidia actually said remarkably little, compared to Ampere and Turing -- most of the focus was on DLSS 3 and its related technology, which isn't surprising, of course. Mind you, they still don't say a lot about the insides compared to their much older architectures; even with a developer account, it's like trying to find a needle in a planet-sized haystack.

Even for RDNA2 AMD did not publish their whitepaper. Much, if not everything, was left to what they showed in their slides, or the document on the updates to the ISA. Their last whitepaper on architecture was RDNA (I think for CDNA they did publish one, but it is nothing more than a pure computer update of GCN5, and from this WP you can extract information and ideas about the matrix capabilities that have been included in the RDNA2 CUs.).

 

[DSirius]

What a great article. Congratulations again and keep posting.
The future looks brighter for all Gpu's RT view points

 

[m3tavision]

This is a great article from neeyik @ Techspot.

It is, things like this^ in-depth layman's review, the reason why I always recommend Techspot to co-workers & friends & acquaintances when inquiring to me, about something. The headspace for much of this can get out of hand and Nick did a great job of consolidating the tech around the consumer.

kudos..

 

[Lounds]

It'll be interesting to see how efficient RDNA3 happens to be or if it's all to do with reducing manufacturing costs. RDNA3 seems efficient for the amount of compute per mm2 of die but how well that translates into gaming performance is yet to be seen.

 

[msroadkill612]

I find it hard to believe that Nvidia & Intel can legally devise a competitive chiplet architecture.

AMD have spent over a decade evolving their cpu & gpuS in the lab & marketplace.

Its not as easy as it seems & both intel & nvidia are missing a leg of a 3 leg stool - nvidia lack a serious cpu & Intel a gpu til very recently.

 

[neeyik]

It'll be interesting to see how efficient RDNA3 happens to be or if it's all to do with reducing manufacturing costs.

As always, any performance improvements seen in games and the like will be dependent on where the processing bottleneck lies within the GPU, in that given application. If it's geometry bound, for example, then a like-for-like RDNA 3 top-end card should be 50% 'better' than an RDNA 2 one, as there are 50% more Shader Engines in Navi 31 than Navi 21.

It will take a bit of driver maturing and programming experience for the compiler and devs to take full advantage of wider SIMD units, but at least they're sensibly backed up with larger register files and low-level caches.

 

[Lionvibez]

This was a fantastic write up thank you TS.

 

[OortCloud]

Great article and not one mention of Elon Musk or any suggestion that the latest Windows Update is causing peoples PCs to catch fire.

 

[msroadkill612]

As always, any performance improvements seen in games and the like will be dependent on where the processing bottleneck lies within the GPU, in that given application. If it's geometry bound, for example, then a like-for-like RDNA 3 top-end card should be 50% 'better' than an RDNA 2 one, as there are 50% more Shader Engines in Navi 31 than Navi 21.

It will take a bit of driver maturing and programming experience for the compiler and devs to take full advantage of wider SIMD units, but at least they're sensibly backed up with larger register files and low-level caches.

Yep - rdna3 is radically new gen& I suspect a lot of fine wine can be harvested over the years from improved drivers & apps code.

 

[enemys]

>They also announced another RTX 4080, one using the smaller-still AD104, but back-tracked on that launch due to the weight of criticism they received. It's expected that this GPU will now be used to launch the RTX 4070 lineup.
It's launched a while already.