Question Speculation: RDNA3 + CDNA2 Architectures Thread

[uzzi38]

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 is much bigger than from RDNA1 to RDNA2. We should expect many big improvements in GFX11. 🤔" / Twitter

 

[Yosar]

I am perfectly aware of the product mixes.
The point is about "who is going to buy them?"
There are people buying them. Even if it is only 30 per store, the top parts will sell out quickly.

I can buy 4090 in my country, no problem. From the first launch day to current day. You know just for over 2000 EUR. Not many eager to pay.
Anecdotal? Yes. True? Either.
If this is nVidia's plan for business they are doomed (joking).

 

[beginner99]

So where are all the RTX 4090s?
OOS everywhere.
I guess those people buying it didn't hear about the recession.
The cards at the top will sell well if the performance delivers

As far as I remember NV reduced or stopped production entirely and at some point there were rumors of them selling some of their wafer allocation. In essence the "shortage" is fake and intentional to make 3000 series look good and get rid of stock.

 

[Tigerick]

	GeForce RTX 4080	Radeon RX 7900XT	GeForce RTX 4090	Radeon RX 7900XTX
SRP	$1,199	$899	$1,599	$999
CUDA / SP	9728	5376	16384	6144
Boost Clock	2.51 GHz	2.4 GHz	2.52 GHz	2.5 GHz
Memory Size	256-bit 16GB GDDR6X 23 Gbps	320-bit 20GB GDDR6 20 Gbps	384-bit 24GB GDDR6X 21 Gbps	384-bit 24GB GDDR6 20 Gbps
Memory BW	736 GB/sec	800 GB/sec	1008 GB/sec	960 GB/sec
Power	320 W	300W	450 W	355 W
RT Cores	76	84	128	96
Tensor / AI Cores	304	168	512	192
Cache	64 MB	80 MB	72 MB	96 MB
Extra Features	DLSS3	FSR3	DLSS3	FSR3

Above is comparison chart between AMD and nVidia's flapship GPU for 2023. Will be interesting to watch how much AMD going to price RX7900XT since some specs are better than RTX4080. We may see nVidia adjusting price of RTX 4080 before they start selling it; Happy to see another slap on Jensen's face

 

[KompuKare]

Single player games.

Have to say that I only play single player games but I have never though to myself:
"If only this added more fake frames between each real one, things would look smoother."
I really see zero or less value in DLSS 3.0 - certainly at the moment with the artifacts and in the future too unless the DL AI becomes truly sentient and plays the game for me. And even the, if I wanted to watch gaming footage I could go to youtube!

 

[Stuka87]

	GeForce RTX 4080	Radeon RX 7900 ?	GeForce RTX 4090	Radeon RX 7900XT ?
SRP	$1,199	$999 ?	$1,599	$1,299 ?
CUDA / SP	9728	10240 - 10752 ?	16384	10752 - 12288 ?
Boost Clock	2.51 GHz	?	2.52 GHz	?
Memory Size	256-bit 16GB GDDR6X 23 Gbps	320-BIT 20GB GDDR6 20 Gbps	384-bit 24GB GDDR6X 21 Gbps	384-bit 24GB GDDR6X 20 Gbps
Memory BW	736 GB/sec	800 GB/sec	1008 GB/sec	960 GB/sec
Power	320 W	?	450 W	?
RT Cores	76	?	128	?
Tensor Cores	304	?	512	?
Cache	64 MB	80 MB	72 MB	96 MB
Extra Features	DLSS3	FSR3 ?	DLSS3	FSR3 ?
PCIe Interface	Gen 4 - 31.5 GB/s	Gen 5 - 63 GB/s	Gen 4 - 31.5 GB/s	Gen 5 - 63 GB/sec

It would be odd for two cards with the same GPU to have different memory bandwidths (320 vs 384).

And if the XT is using GDDR6x, shouldn't it have more than 20Gbps of bandwidth? Otherwise, why waste the price/power on 6x if it has the same bandwidth as the standard GDDR6 on the non-XT?

 

[Timorous]

It would be odd for two cards with the same GPU to have different memory bandwidths (320 vs 384).

And if the XT is using GDDR6x, shouldn't it have more than 20Gbps of bandwidth? Otherwise, why waste the price/power on 6x if it has the same bandwidth as the standard GDDR6 on the non-XT?

For N31 it makes sense. 5x 64bit MCDs is 320 and 6x 64bit MCDs is 384. That is an advantage of the chiplet design. if you are going to cut the memory bus you use less silicon.

I presume the GDDR6X is just a typo.

 

[Stuka87]

For N31 it makes sense. 5x 64bit MCDs is 320 and 6x 64bit MCDs is 384. That is an advantage of the chiplet design. if you are going to cut the memory bus you use less silicon.

I presume the GDDR6X is just a typo.

Oh, of course. I totally forgot about the chiplet setup. Its still dark out, and I need some coffee

 

[Tigerick]

For N31 it makes sense. 5x 64bit MCDs is 320 and 6x 64bit MCDs is 384. That is an advantage of the chiplet design. if you are going to cut the memory bus you use less silicon.

I presume the GDDR6X is just a typo.

Yeah, typo. Corrected

 

[maddie]

For N31 it makes sense. 5x 64bit MCDs is 320 and 6x 64bit MCDs is 384. That is an advantage of the chiplet design. if you are going to cut the memory bus you use less silicon.

I presume the GDDR6X is just a typo.

Do you think it's a planned 5 MCD or a failed 6 MCD bond? If the latter, then no saving. We have seen a 2 core die 7600X exists. which is very strange. The present bond failure rate might be higher than assumed.

 

[Saylick]

Ah dang, how did I miss this one in reading the Twitter thread. After 10 years of using a hardware scheduler (AMD transitioned to it with GCN), it looks like software-based scheduling is back on the menu for RDNA 3, likely for the sake of increasing perf/W and perf/A. Of course, this approach requires more software development work on the compiler but perhaps it will pan out well for AMD given they have far more resources today than before. Also, the side bonus is that compilers can get optimized over time, so FineWine is more of a possibility.

https://twitter.com/x/status/1587201577017917442

 

[Aapje]

@maddie

They don't need huge bonding failures to cut the card this way. One less MCD is still a saving, compared to zero savings for lasering a perfectly good part of the die.

@Saylick

A lot of 'fine wine' was AMD simply lacking the development capacity to use their hardware fully at release and only unlocking the actual power of the card much later. So less fine wine is generally better.

 

[Saylick]

@Saylick

A lot of 'fine wine' was AMD simply lacking the development capacity to use their hardware fully at release and only unlocking the actual power of the card much later. So less fine wine is generally better.

Yeah, that's a good point. Given what I've read, AMD have been putting in the time and effort to make sure the launch drivers are as polished as they can be, so the wine may already be fine. I suppose one negative of a SW based scheduler is that if AMD ever change the WGP too much, development of the optimizations for older architectures may fall by the wayside, similar to what happened with Kepler after Maxwell launched.

 

[Tuna-Fish]

Do you think it's a planned 5 MCD or a failed 6 MCD bond?

Both, obviously. How large proportion is "planned" of course depends on their yields for the step. I'd assume it would start with a lot higher failure rate and then over time the ratio of "true" 5-MCD parts would creep up.

We will know that failure is high if they bring products with the larger GCD and 4 MCDs to market.

If the latter, then no saving.

Being able to sell harvested products is absolutely going to save money.

 

[maddie]

Being able to sell harvested products is absolutely going to save money.

Read the post I replied to please. This, in isolation reply, is frustrating.

 

[Timorous]

Do you think it's a planned 5 MCD or a failed 6 MCD bond? If the latter, then no saving. We have seen a 2 core die 7600X exists. which is very strange. The present bond failure rate might be higher than assumed.

Both as Tuna says but i expect the majority will be 5 MCDs.

 

[Kaluan]

We will know that failure is high if they bring products with the larger GCD and 4 MCDs to market.

Hm, I've already speculated a mid-gen refresh consisting of a N31-based RX 7850 XT may have 4-5 MCDs (and maybe 24Gbps GDDR6).

IDK, it's way too early to speculate about tertiary failure/salvages on advanced packaging, "next gen" is barely off the ground. But considering this is the general trend for chip industries, it will eventually be thing product planners and analysts have to openly account for.

 

[Hans Gruber]

This could be the time when AMD does better in the GPU market than it does in the CPU market.

 

[RnR_au]

Have to say that I only play single player games but I have never though to myself:
"If only this added more fake frames between each real one, things would look smoother."

The only single player game I touch nowadays is from 2005, so the frames are already high enough not to matter for me and it sounds like its the same for you

But I can see the value that some people would attach to smoother game play which still preserving high detail and resolution.

I don't understand the labeling of 'fake' frames as if they are somehow inherently bad. Are we going to call out 'fake' pixels too? Ridicule usage of FSR perhaps?

 

[Shmee]

This could be the time when AMD does better in the GPU market than it does in the CPU market.

It is certainly possible, as it has happened before. It was like this for most of GCN before Zen was out, IMO. The difference now, is that the AMD CPUs are already doing quite well, whether talking AM4 or AM5.

 

[Joe NYC]

This could be the time when AMD does better in the GPU market than it does in the CPU market.

I think AMD will have a lead in performance / power in both CPU and GPU.

I think the era of products with unlimited power consumption will fizzle out.

 

[eek2121]

This could be the time when AMD does better in the GPU market than it does in the CPU market.

Doubt it. Too many folks equate AMD with having slow GPUs with buggy drivers, which is a shame, because next gen sounds like it is going to lay the smackdown on NVIDIA.

AMD could charge similar prices to last year and still wouldn’t sell as many units as NVIDIA.

 

[Kepler_L2]

Do you think it's a planned 5 MCD or a failed 6 MCD bond? If the latter, then no saving. We have seen a 2 core die 7600X exists. which is very strange. The present bond failure rate might be higher than assumed.

Most will probably be 5 MCD, but no doubt there will be some 6 MCD with a failed bond, as well as 4 full MCD + 2 half disabled MCD.

 

[utahraptor]

Any predictions on how AMD will compare on ray tracing this generation?

 

[Stuka87]

Any predictions on how AMD will compare on ray tracing this generation?

Most were thinking below Ada, maybe a bit above Ampere.

But that was before some of this more recent info came out in regards to ray coherence sorting. So, not sure where the current guesses are. Guess we will find out in a few days.

 

[Saylick]

Any predictions on how AMD will compare on ray tracing this generation?

I think anywhere between 2x to 3x hybrid RT performance over N21 is fair game. Who knows what the average might be across 50 titles. Won't know until it gets benched.

Most were thinking below Ada, maybe a bit above Ampere.

But that was before some of this more recent info came out in regards to ray coherence sorting. So, not sure where the current guesses are. Guess we will find out in a few days.

AMD patent for Compute Unit Sorting for Reduced Divergence: https://www.freepatentsonline.com/20200409695.pdf