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

 

[PJVol]

It seems like if the hardware was better utilized more power would be used. If it was some game specific optimization that reduced wasted effort by the hardware then perhaps power usage would remain the same.

I am just curious, not really expecting an answer.

That's a good question. While both usually contribute to performance gains, I believe it's more of the former in case of TLOU.

 

[menhera]

According to AMD, a cache request shown in its profiling tool is 64 Bytes or 128 Bytes. 128 Bytes for vector memory and 64 Bytes for Instruction/Scalar. So why not calculate data transferred in real world scenarios?



Wrote a simple program that calculates bandwidth with given values.



I tested these games with my 6900 XT at 4K simply because I have them installed on my PC.



Unfortunately AMD doesn't expose hitrate for L3, but this explains why AMD added L3 to RDNA 2. VRAM would be a serious bottleneck without L3.

RDNA 2's L1 cache is really bad. It spills too many requests. It's still acceptable in rasterization with other caches mitigating misses from L1. However it becomes problematic in ray tracing. Ray tracing hurts L0 quite a bit as shown in RE: V, and L1 can't mitigate the misses. L2 and (presumably L3) do all the work instead. They're just too slow for ray tracing. This also happens in almost every title with ray tracing support.

Although everyone is belittling RDNA 3, I wonder how RDNA 3's increased register files, L0, and L1 behave during entire frames. Chips and Cheese covered a bit, but only during specific workloads.

 

[menhera]

Forgot to mention that the graphs are average bandwidth required per frame. A single frame consists of thousands of workloads, and their characteristics are all seem to be different. While some workloads mostly rely on L0, there's certainly a number of workloads where L0 and L1 can't do anything. If not for L3, my 6900 XT with 512GB/s bandwidth would be choked in those memory-intensive workloads.

 

[SteinFG]

Then you should mention that in your table, It's not like I know where you got those effective BW data except the one for N21. Where did you find them?
Also what you calculated vs data from graph are quite different.
N21 58% vs 62%. This isn't that different
N22 60% vs 69%. This is very different
N23 and N24 looks weird, but you already know It.

N22: AMD's website, 6700 XT specs page
N23: Review guide for RX 7600, where they compare it to 6600
N24: AMD's website, 6500 XT specs page
N31: AMD's website, 7900 XTX specs page
As for why the data is different, i have no idea why. But if you look at my calculated L3 hitrate for N21 and N22, they are very close to the nearest whole, which suggests that those were exact numbers used in their calculations. The rest doesn't look good though ¯\_(ツ)_/¯ I'm too lazy to do any more digging

switched to RDNA3 thread because there was nothing about RDNA4

 

[TESKATLIPOKA]

N22: AMD's website, 6700 XT specs page
N23: Review guide for RX 7600, where they compare it to 6600
N24: AMD's website, 6500 XT specs page
N31: AMD's website, 7900 XTX specs page
As for why the data is different, i have no idea why. But if you look at my calculated L3 hitrate for N21 and N22, they are very close to the nearest whole, which suggests that those were exact numbers used in their calculations. The rest doesn't look good though ¯\_(ツ)_/¯ I'm too lazy to do any more digging

switched to RDNA3 thread because there was nothing about RDNA4

Thanks for the info about the data. I will check It out. I wouldn't be surprised If AMD marketing made a mistake in calculation.

edit:
I think N21 hitrate was at 4K, N23 at 1440p and N24 at 1080p, then It would be comparable to that graph I provided.
N22 is still different, but let's just say they made a mistake.

What I find interesting is that AMD managed to increase IC channels 3x per MB. Meaning, that with 128MB and the same clock It would provide 3x higher BW. BTW It looks like N31 IC works at 2.3GHz.

 

[SteinFG]

N31 on N32 package spotted in the wild. 84 CUs, 16GB:

Sapphire's NITRO Radeon RX 7900 GRE graphics card has been pictured - VideoCardz.com

Sapphire’s RX 7900 GRE with Navi 31 and 16GB memory A leak featuring Radeon RX 7900 GRE from @wnxod surfaced today, revealing not only a logo but also a photograph of a new card. Sapphire Radeon RX 7900 GRE NITRO, Source: @wnxod The showcased GPU appears to be the Sapphire NITRO+ graphics card...

videocardz.com

top image is Nitro+ 7900XTX, bottom is 7900 GRE

 

[Timorous]

N31 on N32 package spotted in the wild. 84 CUs, 16GB:

Sapphire's NITRO Radeon RX 7900 GRE graphics card has been pictured - VideoCardz.com

Sapphire’s RX 7900 GRE with Navi 31 and 16GB memory A leak featuring Radeon RX 7900 GRE from @wnxod surfaced today, revealing not only a logo but also a photograph of a new card. Sapphire Radeon RX 7900 GRE NITRO, Source: @wnxod The showcased GPU appears to be the Sapphire NITRO+ graphics card...

videocardz.com

top image is Nitro+ 7900XTX, bottom is 7900 GRE
View attachment 83438

So a card that will sometimes match the 7900XT and other time fall behind due to bandwidth. Reminds me of the 5600XT.

If this launches globally and is priced in the $600 - $650 range it would be a good card IMO. Similar performance to a 4070Ti in pure raster but more VRAM and bandwidth to give it longer legs and priced a tad above the standard 4070.

 

[Timorous]

Here is the full spec of the 7900 GRE from This chinese review.

model	AMD Radeon RX 7900 GRE
Core code	Navi 31
Number of Stream Processor Units	5120
Base Frequency(MHz)	1287
Boost Frequency(MHz)	2245
Memory Speed(Gbps)	18
memory type	GDDR6
Memory capacity (GB)	16
Memory interface width (bit)	256
video interface	HDMI 2.1×1 DisplayPort 2.1×2 USB Type-C×1
power supply interface	2*8 Pin
TBP (W)	260
Size(L*W*H)	268mm*51mm*97mm

So slower RAM as well. 4K performance puts it quite a bit behind the 7900XT but quite a way ahead of the 4070. Taking the delta to the 7900XT it looks like this is about on par with the 6950XT which with the above spec is hardly surprising tbh.

Package certainly looks smaller than the one on the 7900 XTX / XT

Compared to N32 we have

5SE vs 3SE
80CU vs 60CU.

To overcome the raw hardware deficit you need 66% higher clocks to match in terms of ROPS or 33% higher clocks to match in terms of shader performance. Given the performance I don't think 160 ROPS is really required. Based on the spec boost clock of 2245 that would mean a 3Ghz N32 based part with 16GB of 18gbps ram should perform in a similar ballpark to this thing.

 

[Tup3x]

AMD's naming is horrible... Maybe NVIDIA should have named their cards as RTX 4090 24GB, RTX 4090 16 GB and RTX 4090 12 GB.

 

[GodisanAtheist]

Let's just sit for a moment and admire what a flaming trash fire AMD's execution has been this generation.

It really is a sight to behold.

How does AMD so deftly, and seemingly skillfully, go from tightly executed and we'll oiled launches like RDNA2 to... Whatever the hell is happening right now?

And off in the distance is NV chugging away, selling crappy chips one tier up for absurd prices across a full product stack.

 

[Timorous]

Let's just sit for a moment and admire what a flaming trash fire AMD's execution has been this generation.

It really is a sight to behold.

How does AMD so deftly, and seemingly skillfully, go from tightly executed and we'll oiled launches like RDNA2 to... Whatever the hell is happening right now?

And off in the distance is NV chugging away, selling crappy chips one tier up for absurd prices across a full product stack.

I think it stems from the fact RDNA3 missed the 1.5x perf/watt increase they were expecting.

Had they hit that then the 7900XTX would be closer to the 4090 and the 7900XT would probably match the 4080. That would make the initial $900 price of the 7900XT far more palatable and they probably would have charged $1200 for the XTX.

From that a 3Ghz + N32 is up there ahead of AIB 6950XTs and has pricing room to be sold for $700 ish and the cut down 7700XT could come in at $450 or so with 6800XT tier performance.

7600 would also perform better in its power envelope pushing it on to perform around 6700XT tier.

Missing this has meant AMD have had to scramble to get their stack launched on a world where top N32 is either massively under performing or one where it has a price ceiling due to the poor 7900XT performance and the crashing prices.

 

[Aapje]

@Timorous

The real issue was not that they missed their targets, but that they thought that they could fix it in the drivers. So they didn't want to price it too low and then offer too much price/performance once the drivers were fixed.

But with the driver never being able to unlock that supposed potential, the positioning was just wrong.

 

[PJVol]

The real issue was not that they missed their targets, but that they thought that they could fix it in the drivers.

They might have made weird decisions before, but not so much to think they could.

 

[jpiniero]

AMD Radeon RX 7900 GRE officially launches with 5120 cores, 16GB memory and 260W TBP, costs $649 - VideoCardz.com

AMD Radeon RX 7900 GRE features fewer cores and less memory than XT model, available globally soon AMD introduces ‘RX 7800 XT’ under a different name. Radeon RX 7900 GRE, Source: AMD During AMD’s presentation at China Joy in Shanghai, China, the much-anticipated Radeon RX 7900 GRE “Golden...

videocardz.com

Videocardz makes it sound like the 7900GRE will launch globally but only through OEMs outside of China.

 

[SteinFG]

So, with some reviews out already, 80CU 7900GRE = 80CU 6950XT ...

https://twitter.com/x/status/1684785180031258625

To get Navi32 60CU RX7800 to perform on par with 4070, it must have at least 2700 mhz boost.

 

[igor_kavinski]

Faster than a 4070? Meh. They have such low ambitions. Could have targeted the 4070 Ti at least at $600. Morons.

I feel like going out in the desert and launching a few rockets into the sand dunes to cool off.

 

[eek2121]

They should have named it the 7900 and launched it globally. I hate AMD’s naming scheme, however. Their scheme has folks mentally comparing 7s, 8s, and 9s with the competition, which makes them look bad. The 7900XTX is a 7800XT at best.

Maybe next gen they will do better. “Hey AMD, give me a reason to upgrade my 4090 to an AMD product.”

 

[SteinFG]

Faster than a 4070? Meh. They have such low ambitions. Could have targeted the 4070 Ti at least at $600. Morons.

I feel like going out in the desert and launching a few rockets into the sand dunes to cool off.

Radeon: "Best I can do is ~10% better than 4070 for ~10% more. Take it and leave me alone."

 

[jpiniero]

They should have named it the 7900 and launched it globally.

This seems to be mainly about dumping really crappy quality bin GCD. The number of chips that don't bin well enough for the 7900 XT probably isn't that high.

 

[SolidQ]

Could have targeted the 4070 Ti at least at $600

Classic Amd set high MSRP, then in half+ year dropping 100-150$

 

[Heartbreaker]

7900GRE should just be 7800XT and globally launched. Even at the same performance it's significantly more desirable than 6950XT. Lower Power usage, AV1 encoder, significantly better FP/ML and RT performance.

 

[jpiniero]

7900GRE should just be 7800XT and globally launched. Even at the same performance it's significantly more desirable than 6950XT. Lower Power usage, AV1 encoder, significantly better FP/ML and RT performance.

It costs a lot more than N21, especially since it's likely the price gap between N5/N4 and N7/N6 has widened.

This is just a dud dump.

 

[Ranulf]

It costs a lot more than N21, especially since it's likely the price gap between N5/N4 and N7/N6 has widened.

This is just a dud dump.

Sure seems that way.

 

[Heartbreaker]

It costs a lot more than N21, especially since it's likely the price gap between N5/N4 and N7/N6 has widened.

This is just a dud dump.

If N32 is as bad as the rest of the lineup, AMD will absolutely need an N31 16GB to fill that gap in the lineup.

Only if AMD N32 is going to be MUCH better (as I keep hoping) than expected (significant better perf/CU than other RDNA, and maybe more features (RDNA 3.5), can they not have a 16 GB N31.

If AMD eventually releases (very late) N32 cards, and they are no better than the rest of RDNA3 lineup, it just adds to the picture of AMD GPU incompetence.

 

[jpiniero]

If N32 is as bad as the rest of the lineup, AMD will absolutely need an N31 16GB to fill that gap in the lineup.

They could do RDNA2 rebrands instead if they feel like they need something inbetween.