Discussion RDNA4 + CDNA3 Architectures Thread

[DisEnchantment]

With the GFX940 patches in full swing since first week of March, it is looking like MI300 is not far in the distant future!
Usually AMD takes around 3Qs to get the support in LLVM and amdgpu. Lately, since RDNA2 the window they push to add support for new devices is much reduced to prevent leaks.
But looking at the flurry of code in LLVM, it is a lot of commits. Maybe because US Govt is starting to prepare the SW environment for El Capitan (Maybe to avoid slow bring up situation like Frontier for example)

See here for the GFX940 specific commits

History for llvm/lib/Target/AMDGPU - llvm/llvm-project

The LLVM Project is a collection of modular and reusable compiler and toolchain technologies. - History for llvm/lib/Target/AMDGPU - llvm/llvm-project

github.com

Or Phoronix

More AMD "GFX940" Enablement Work Landing In LLVM - Phoronix

www.phoronix.com

There is a lot more if you know whom to follow in LLVM review chains (before getting merged to github), but I am not going to link AMD employees.

I am starting to think MI300 will launch around the same time like Hopper probably only a couple of months later!
Although I believe Hopper had problems not having a host CPU capable of doing PCIe 5 in the very near future therefore it might have gotten pushed back a bit until SPR and Genoa arrives later in 2022.
If PVC slips again I believe MI300 could launch before it

This is nuts, MI100/200/300 cadence is impressive.



Previous thread on CDNA2 and RDNA3 here

Question - Speculation: RDNA3 + CDNA2 Architectures Thread

Man I have been dying to make this one for a while now. First rumours for RDNA3 are here so new thread time! Just going to start off with this one for now: kopite7kimi on Twitter: "@VideoCardz Ah, I mean a simple mcm design with 10240 cores is not enough. Because the lift from RDNA2 to RDNA3...

forums.anandtech.com

 

[adroc_thurston]

cause now it's ridiculous.

It's fine.

4060ti is smaller and more silent because they consume 50-100w lower than their counterparts

Anyone can clock N5 die low enough.
even Intel.

 

[SolidQ]

Do you mean MLID was wrong and N31 isn't +80% to +120% (or optimistically +130-140%) over N21?

originally he was saying 50-60% over N21

 

[MrTeal]

originally he was saying 50-60% over N21

 

[SolidQ]

MrTeal​

Two years ago. Even AMD on presentation was up 1.7x

 

[blckgrffn]

This might be silly, but it would be nice if there was a 1 year anniversary re-review of some of these cards. Up to date drivers, current games, and then updated conclusions.

Like NFL draft grades after the year of football is played. Sometimes the tech matures, features are widely seen in games, and its better than thought. Other times there's no uplift and some touted, unresolved widespread bugs and features are MIA IRL and that can be addressed. Or bad power connectors, I dunno but with the cycles seemingly being lengthened I would appreciate this.

Where is the 7900 XTX vs the 6950XT now? Its probably not great, but has the needle moved since launch? (if this exists I'd love a link )

 

[GodisanAtheist]

This might be silly, but it would be nice if there was a 1 year anniversary re-review of some of these cards. Up to date drivers, current games, and then updated conclusions.

Like NFL draft grades after the year of football is played. Sometimes the tech matures, features are widely seen in games, and its better than thought. Other times there's no uplift and some touted, unresolved widespread bugs and features are MIA IRL and that can be addressed. Or bad power connectors, I dunno but with the cycles seemingly being lengthened I would appreciate this.

Where is the 7900 XTX vs the 6950XT now? Its probably not great, but has the needle moved since launch?

-TPU tends to pretty frequently re-run their benches for their whole suite of cards (which is why chart positions occasionally shift with them).

From a recent 4070Ti review:

Looks like their AMD driver revision is the November WHQL release, so roughly a year after launch.



Looks like the 7900XTX is roughly ~40% faster than the 6900XT at 1440p, which is almost exactly the same difference as the launch review.

 

[blckgrffn]

-TPU tends to pretty frequently re-run their benches for their whole suite of cards (which is why chart positions occasionally shift with them).

From a recent 4070Ti review:

Looks like their AMD driver revision is the November WHQL release, so roughly a year after launch.

View attachment 92774

Looks like the 7900XTX is roughly ~40% faster than the 6900XT at 1440p, which is almost exactly the same difference as the launch review.

View attachment 92773

View attachment 92775

Having different cards be the 100% and comparing those graphs is reminding me how much I need coffee in the afternoon. It's nice that they re-run the tests instead of just keeping the data canned, thanks for the heads up on that!

Are folks who paid $1K for the 7900XTX in ~December 22/Jan 23 feeling OK about it? I would assume that they would be, given the value has stayed strong for at least 12 months. I do realize that its more of "only" a 3090 type performer when there is RT involved, but in my book that's enough to be useful.

If RDNA 4 can be about the same performance with a small monolithic die and $600 I think that's a win for most PC enthusiasts. ¯\_(ツ)_/¯ Especially if it can do it with less power usage...

 

[Aapje]

The Super releases are also a great opportunity to compare, since they were released pretty much a year later. Looking at HBU, AMD seems to have made solid gains with the driver:

A year ago:

and a week ago with a more demanding set of games, but you can still look at the relative numbers:

 

[Aapje]

So the 7900 XTX went from being pretty much equal to the 4080 to being quite a bit ahead.

I think it's smart to keep this 'fine wine' aspect in mind when shopping for GPUs. The AMD cards tend to gain more as AMD seems to need more time to sort the drivers.

 

[MrTeal]

MrTeal
Two years ago. Even AMD on presentation was up 1.7x

Hey look at that, we're both right. Shocking. The benefit of being MLID is that you can say all the numbers for all the things whenever you feel like it, and you always kinda get there at some point.

 

[SolidQ]

The benefit of being MLID

alot people missread him. He's even in latest video explain, he give info about current situation, not final, because before product out, there can be some changes

So the 7900 XTX went from being pretty much equal to the 4080 to being quite a bit ahead.

I can give alot pictures, but 3 is enough, People who bought 7900XT are real winners. 4070ti pretty meh card

Spoiler

Personally waiting for N48 and will decide, buy it or no

 

[MrTeal]

alot people missread him. He's even in latest video explain, he give info about current situation, not final, because before product out, there can be some changes

Product naming and price, sure.

RDNA3 being an IO die + one or more compute dies? I don't think that was the situation a couple years ago less than a year from when the 7900XTX launched.

 

[S'renne]

By the way is there no news of AMD improving their media engine? Don't they still have worse encoders compared to NVIDIA and Intel rn

 

[Heartbreaker]

By the way is there no news of AMD improving their media engine? Don't they still have worse encoders compared to NVIDIA and Intel rn

They have great encoders in their Xilinx unit, that they could migrate to GPUs:

AMD Announces Alveo MA35D Media Accelerator: AV1 Video Encode at 1W Per Stream

www.anandtech.com

 

[S'renne]

They have great encoders in their Xilinx unit, that they could migrate to GPUs:

AMD Announces Alveo MA35D Media Accelerator: AV1 Video Encode at 1W Per Stream

www.anandtech.com

Hopefully, since GPUs isn't only used for gaming/ML/AI tasks and AMD being also competitive with their media engine would expand options

 

[adroc_thurston]

By the way is there no news of AMD improving their media engine? Don't they still have worse encoders compared to NVIDIA and Intel rn

VCN is made for a bit different reasons, lowest power decode and maximum amount of encode streams.
But maybe.

 

[Aapje]

VCN is made for a bit different reasons, lowest power decode and maximum amount of encode streams.
But maybe.

But it is worse than the solutions from Nvidia and Intel. They need to step up their game.

 

[adroc_thurston]

But it is worse than the solutions from Nvidia and Intel

Wrt to decode power or the number of 265 streams per instance?
Naaa, solid.
They still need lower decode power since all the phone AP vendors are outright better

 

[Aapje]

Wrt to decode power or the number of 265 streams per instance?

The encoder quality is worse on the same resolution & bitrate:

BTW, you seem to be moving the goalposts. This discussion started off purely about encoders and you are now only talking about decoders and the number of streams.

 

[adroc_thurston]

The encoder quality is worse on the same resolution & bitrate:

Again that's not a priority.

This discussion started off purely about encoders and you are now only talking about decoders and the number of streams.

Yes and it's not a priority.
AMD media engine is primarily for sticking it into APUs where decode power, and overall media block power are the #1 factors.
It took a while for VCN as you know it to even reach dGPUs.

 

[S'renne]

The encoder quality is worse on the same resolution & bitrate:

BTW, you seem to be moving the goalposts. This discussion started off purely about encoders and you are now only talking about decoders and the number of streams.

Welp they arent even on topic anymore, atm the main reason AMD GPUs are bought for clients are just gaming since they cant really do much else and apparently competing on media encoders arent a priority according to that guy. The same market where people buys the a380 for AV1 encoding lol

 

[SolidQ]

The same market where people buys the a380 for AV1 encoding lol

because most people not pixel hunters, for them all same quality amd/nvidia/intel. For pixel hunters there is difference

 

[Aapje]

Again that's not a priority.

Yes and it's not a priority.
AMD media engine is primarily for sticking it into APUs where decode power, and overall media block power are the #1 factors.
It took a while for VCN as you know it to even reach dGPUs.

You claim that it's not a priority, but they've been actually improving it in RDNA3, so clearly it was a sufficient priority to work on.

And fact is that they are worse than the other vendors, so it is something where they are lagging behind.

 

[moinmoin]

You claim that it's not a priority, but they've been actually improving it in RDNA3, so clearly it was a sufficient priority to work on.

It's one of many different IP blocks that are being independently improved over time by separate teams. That by itself is not saying much wrt priority. That said I do hope this team combines with the one at Xilinx to eventually offer improved specs. Seems like an obvious symbiosis enabled by the merger.

 

[eek2121]

This might be silly, but it would be nice if there was a 1 year anniversary re-review of some of these cards. Up to date drivers, current games, and then updated conclusions.

Like NFL draft grades after the year of football is played. Sometimes the tech matures, features are widely seen in games, and its better than thought. Other times there's no uplift and some touted, unresolved widespread bugs and features are MIA IRL and that can be addressed. Or bad power connectors, I dunno but with the cycles seemingly being lengthened I would appreciate this.

Where is the 7900 XTX vs the 6950XT now? Its probably not great, but has the needle moved since launch? (if this exists I'd love a link )

Gamers Nexus routinely reruns their benchmarks.