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

 

[Mopetar]

If N32 is getting cut to 40 CU it's an absolute salvage-tier part. Limiting it to 8 GB means they'd be needlessly crippling a lot of silicon as well.

I don't think they'd use it for a 7600X either because there's too much of a gap. 7600X is a refresh part in early 2024 with higher clocks (maybe 16 GB of VRAM) if we ever get one at all.

 

[Ranulf]

Do you mean a 7600Xt with 16GB?

Because the 7600 is apparently already the full 32 CUs...

I thought he was talking about the 7600, not the XT. Do we have any rumors on it? Honestly I might have missed it and I thought we just knew the 7600 was this month.

 

[Timorous]

60 cu + 16GB for 7800XT
48 cu + 16GB for 7700XT
40 cu + 12GB for 7600XT

I can see that with perhaps an intermediate 54CU 7800 or alternatively they do a 7800XTX based on that 70CU N31.

Seems obvious a lot of fully working dies will be cut to make the lower tier parts but AMD could go aggressive on the top v/f curve bins.

 

[jpiniero]

The 7600 XT could be a 40 CU, 8 GB version of N32.

There would be a better case to be made to just rebrand N22 if they were going to do something like that. Especially if N7/6 prices tank.

 

[Heartbreaker]

60 cu + 16GB for 7800XT
48 cu + 16GB for 7700XT
40 cu + 12GB for 7600XT

I can see that with perhaps an intermediate 54CU 7800 or alternatively they do a 7800XTX based on that 70CU N31.

60 CU N32 won't have close to enough performance vs 72 CU 6800 XT to call it a 7800 XT. Lets face it, 72 CU N31 is barely going to make that leap. But 72 CU seems like the next reasonable cut of N31.

72 cu + 16GB for 7800XT (N31) - Though this is most likely a smaller improvement than NVidia 4080 over 3080.
60 cu + 16GB for 7800 (N32) - again even using the full part, it's going to be small gain over 6800.
48 cu + 12GB for 7700XT - This one would be a nice improvement over 6700 XT.

Maybe one more cut, or maybe not. The lower end cuts are just leaving money on the table. They aren't cheaper to produce, they just have lower margin.

 

[Ajay]

60 CU N32 won't have close to enough performance vs 72 CU 6800 XT to call it a 7800 XT. Lets face it, 72 CU N31 is barely going to make that leap. But 72 CU seems like the next reasonable cut of N31.

72 cu + 16GB for 7800XT (N31) - Though this is most likely a smaller improvement than NVidia 4080 over 3080.
60 cu + 16GB for 7800 (N32) - again even using the full part, it's going to be small gain over 6800.
48 cu + 12GB for 7700XT - This one would be a nice improvement over 6700 XT.

Maybe one more cut, or maybe not. The lower end cuts are just leaving money on the table. They aren't cheaper to produce, they just have lower margin.

Yeah. As has been discussed, the new RDNA3 architecture and chipset system didn’t move the performance needle enough. Perf/watt is good. So I think you are offering a reasonable solution to the product stack problem AMD's has created for itself (the 7900XT should have been the new 7800XT). RDNA4 can’t come soon enough for AMD, but it still ~ a year away.

 

[linkgoron]

Yeah. As has been discussed, the new RDNA3 architecture and chipset system didn’t move the performance needle enough. Perf/watt is good. So I think you are offering a reasonable solution to the product stack problem AMD's has created for itself (the 7900XT should have been the new 7800XT). RDNA4 can’t come soon enough for AMD, but it still ~ a year away.

The perf/watt isn't good really - that's why it didn't move the performance enough. I'd assume that RDNA 4 is closer to a year and a half than a year, given the recent release history.

 

[PJVol]

The perf/watt isn't good really - that's why it didn't move the performance enough

Does anyone have an idea of how far doubling ALU count in CU (i.e. ILP) was really justified? It'd also be nice to know if these ALUs can be temporarily disabled and how it will affect performance.

Or, to put it another way, does these ALUs still contribute to the SIMD logic IC's total switching activity, issuing single instruction per cycle?

Anyone can shed some light?

 

[Mopetar]

This was something NVidia went through a while ago and the drivers need to be able to leverage the hardware in this way. Since AMD never previously had the capability of executing two floating point operations in the same cycle, the software was never built to accommodate this and even if they had the support for it, it still relies on whatever program is running to take advantage of this.

If it's mainly a driver side issue from AMD then expect RDNA 3 to be more fine wine since the untapped potential is there. However, there are limits. NVidia never saw a magical doubling in performance because most games can't leverage that extra capacity.

There were said to be some hardware bugs with RDNA3 as well, but I don't know if those are related to this particular aspect of the hardware.

 

[Timorous]

60 CU N32 won't have close to enough performance vs 72 CU 6800 XT to call it a 7800 XT. Lets face it, 72 CU N31 is barely going to make that leap. But 72 CU seems like the next reasonable cut of N31.

72 cu + 16GB for 7800XT (N31) - Though this is most likely a smaller improvement than NVidia 4080 over 3080.
60 cu + 16GB for 7800 (N32) - again even using the full part, it's going to be small gain over 6800.
48 cu + 12GB for 7700XT - This one would be a nice improvement over 6700 XT.

Maybe one more cut, or maybe not. The lower end cuts are just leaving money on the table. They aren't cheaper to produce, they just have lower margin.

At 20% higher clocks N32 should be a small amount ahead of the 6800XT. Based on a stock 6800XT that would be around 2.7Ghz sustained.

That is something N22 can already achieve and N31 clocks better than N21 as does N33 Vs N23 so I see no reason why N32 could not reasonably hit 3Ghz. This would be a 10% gain over N22 and is about what the 7900XTX reference manages Vs the 6950XT reference.

3Ghz sustained would be a 33% increase in pure clock Vs the 6800XT and with an approximate 9% IPC gain as well that should be enough to match the AIB 6950XTs which are around 18% faster than the 6800XT at 4K.

As for using cut N31 we know if AMD release such a thing to consumers it will probably be a 70CU 4MCD part like the W7800 is. Which way AMD goes will depend on how well N32 clocks. 3Ghz or more and I don't think AMD will bother with cut N31 for dGPUs. If it clocks lower AMD might go for it to have a reasonable gain over the 6800XT.

 

[jpiniero]

This was something NVidia went through a while ago and the drivers need to be able to leverage the hardware in this way.

IIRC nVidia went from FP32+INT32 to FP32+(FP32 or INT32). So it's not quite the same thing.

 

[PJVol]

There were said to be some hardware bugs with RDNA3 as well, but I don't know if those are related to this particular aspect of the hardware.

Yeah, these could be different, but anyway I'm curious as to what engine's power draw could have been if those "additional" ALUs can be completely disabled, thinking about the extent to which the perf/W improvement of TSMC's 5nm over 7nm node was actually utilized.

 

[Ajay]

IIRC nVidia went from FP32+INT32 to FP32+(FP32 or INT32). So it's not quite the same thing.

Yes, that’s what NV did. pretty sure their profiling show that 70% of the compute workload was FP and 30% was INT - so the change made sense.

 

[Heartbreaker]

At 20% higher clocks N32 should be a small amount ahead of the 6800XT. Based on a stock 6800XT that would be around 2.7Ghz sustained.

That is something N22 can already achieve and N31 clocks better than N21 as does N33 Vs N23 so I see no reason why N32 could not reasonably hit 3Ghz. This would be a 10% gain over N22 and is about what the 7900XTX reference manages Vs the 6950XT reference.

3Ghz sustained would be a 33% increase in pure clock Vs the 6800XT and with an approximate 9% IPC gain as well that should be enough to match the AIB 6950XTs which are around 18% faster than the 6800XT at 4K.

As for using cut N31 we know if AMD release such a thing to consumers it will probably be a 70CU 4MCD part like the W7800 is. Which way AMD goes will depend on how well N32 clocks. 3Ghz or more and I don't think AMD will bother with cut N31 for dGPUs. If it clocks lower AMD might go for it to have a reasonable gain over the 6800XT.

From what I can see 7900 XT (84 CU) is about 30% faster than 6800 XT (72 CU) that makes gains about half CU count increase and the rest clock/IPC increases, so not much more than 15% from clocks/IPC.

So since the 6800 XT has 20% more CUs that N32 Max, it would be a stretch N32 to even match 6800XT. Maybe a small gain. That would make it an OK 7800..

The 72 CU N31 will likely best 6800 XT by 15%-20%, which makes for a small upgrade for 7800 XT, but that seems to be the generation we are in.

I wish they would just release the things already. Maybe they will surprise us somewhere.

 

[Kepler_L2]

There would be a better case to be made to just rebrand N22 if they were going to do something like that. Especially if N7/6 prices tank.

AMD is not going to rebrand RDNA2 as RX 7000.

 

[Timorous]

From what I can see 7900 XT (84 CU) is about 30% faster than 6800 XT (72 CU) that makes gains about half CU count increase and the rest clock/IPC increases, so not much more than 15% from clocks/IPC.

So since the 6800 XT has 20% more CUs that N32 Max, it would be a stretch N32 to even match 6800XT. Maybe a small gain. That would make it an OK 7800..

The 72 CU N31 will likely best 6800 XT by 15%-20%, which makes for a small upgrade for 7800 XT, but that seems to be the generation we are in.

I wish they would just release the things already. Maybe they will surprise us somewhere.

At 4K 7900XT is about 34% faster than 6800XT per TPU, Techspot and the meta review. That is from a 19% clockspeed increase and a 16% CU count increase. Practically the clockspeed * CU delta.

3 GHz sustained would be a 33% increase over 6800XT and with 80% of the CUs that would calculate to 6.4% faster than 6800XT which is pretty much 6900XT level.

For me it still comes down to how well N32 clocks. Over 3Ghz sustained and AMD don't need to bother with another cut N31 for the gaming market. Anything less than 3Ghz and chances are N32 will top out at 7800 and they will need that 70CU N31 variant to fill the role of 7800XT.

 

[KompuKare]

The perf/watt isn't good really - that's why it didn't move the performance enough. I'd assume that RDNA 4 is closer to a year and a half than a year, given the recent release history.

I guess both things can be true: perf/watt is pretty good considering it is now chiplet but compared to it's monolithic predecessors it has stayed stagnant.

In terms of risk management, going chiplet for both Navi31 and Navi32 still seems excessively risky to me. Yes, they get to save on the IO die but the R&D costs were higher and AMD have little volume to spread this.

To me, a halo part with chiplets and the main die about 500mm² as a proff of concept would have made more sense. Then below that stick to monoliths. For a halo part perf/watt doesn't matter as much especially if such a huge part would have beaten the 4090.

 

[PJVol]

I guess both things can be true: perf/watt is pretty good considering it is now chiplet but compared to it's monolithic predecessors it has stayed stagnant.

Any indication of this is true for the current generation?

Then below that stick to monoliths. For a halo part perf/watt doesn't matter as much

 

[Heartbreaker]

RX 7600 shows up online store in Canada. Priced above 6650 XT. 443.99 CAD (330.80 USD) for the lowest priced one.

Custom Radeon RX 7600 graphics cards listed in Canada at 444 CAD - VideoCardz.com

AMD Radeon RX 7600 launching soon Custom Radeon RX 7600 graphics cards have been listed by a Canadian retailer. We have received a tip from one of our readers that PC Canada is already listing two custom models from Sapphire and MSI. As per the product description, both cards are confirmed to...

videocardz.com

 

[jpiniero]

RX 7600 shows up online store in Canada. Priced above 6650 XT. 443.99 CAD (330.80 USD) for the lowest priced one.

Custom Radeon RX 7600 graphics cards listed in Canada at 444 CAD - VideoCardz.com

AMD Radeon RX 7600 launching soon Custom Radeon RX 7600 graphics cards have been listed by a Canadian retailer. We have received a tip from one of our readers that PC Canada is already listing two custom models from Sapphire and MSI. As per the product description, both cards are confirmed to...

videocardz.com

$319/$329 would make sense as the MSRP, given that the original MSRP of the 6600 was $329. Only this time you get the full die instead. Wouldn't even look that bad against the 4060, assuming that is $399.

The problem is RDNA 2's current prices.

 

[Heartbreaker]

$319/$329 would make sense as the MSRP, given that the original MSRP of the 6600 was $329. Only this time you get the full die instead. Wouldn't even look that bad against the 4060, assuming that is $399.

The problem is RDNA 2's current prices.

I'd bet on 7600, $299 MSRP in the USA, Canada often ends up paying more than USA.

Though you can buy 10GB 6700 at Newegg for $299 right now, so this isn't going to look great in comparison.

 

[Aapje]

And a 6700 XT for $320, so $330 is DOA.

But these are prelaunch listings which are extremely unreliable. The companies have learned not to tell the real price to shops, because they would inevitably get leaked. So they just get a maximum price. So we now only know that the price is below $330.

 

[Thunder 57]

I said $299 but some people really didn't want to listen and insisted on $249. I expect what will happen is it will be $299 and then everyone who has been holding out will just buy a 6700 and eat up the remaining supply (knowing there won't be any better deals), and suddenly the 7600 looks like a good value.

Of course, we still have to wait a bit longer to see what it really launches at.

 

[Aapje]

I fully expect the card to be mediocre value compared to the old cards and then slowly creep down in price. I don't see much point in holding out for the 7600 anyway. Perhaps for a clamshell 16 GB version.

 

[TESKATLIPOKA]

At 4K 7900XT is about 34% faster than 6800XT per TPU, Techspot and the meta review. That is from a 19% clockspeed increase and a 16% CU count increase. Practically the clockspeed * CU delta.

3 GHz sustained would be a 33% increase over 6800XT and with 80% of the CUs that would calculate to 6.4% faster than 6800XT which is pretty much 6900XT level.

For me it still comes down to how well N32 clocks. Over 3Ghz sustained and AMD don't need to bother with another cut N31 for the gaming market. Anything less than 3Ghz and chances are N32 will top out at 7800 and they will need that 70CU N31 variant to fill the role of 7800XT.

I see 2 potential problems even If N32 can sustain 3GHz.
1.) How much performance increase we will see from 33% higher clocks.
2.) How much power would be needed for that. RX 6500XT showed us how inefficient It can be.