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

 

[Panino Manino]

It's an uphill battle.
RDNA3 may be 3x better, but it'll compete with Ampere Next. How much better will that competitor be? See the current landscape for example, AMD launched good GPUs that compete well with the Nvidia equivalents, but soon Nvidia will once again launch new skus that will be positioned just ahead of the AMD's. Nvidia always manages to do this, to come with an extra card to remove AMD's advantage.
It's good and all that AMD have a simpler lineup but when the rival is flooding the market with products and mindshare is hard to compete.

 

[fleshconsumed]

All for naught if they can't make enough and if crypto mining is still a thing. Hopefully, both will not be an issue when RDNA3/lovelace launch.

 

[biostud]

It's an uphill battle.
RDNA3 may be 3x better, but it'll compete with Ampere Next. How much better will that competitor be? See the current landscape for example, AMD launched good GPUs that compete well with the Nvidia equivalents, but soon Nvidia will once again launch new skus that will be positioned just ahead of the AMD's. Nvidia always manages to do this, to come with an extra card to remove AMD's advantage.
It's good and all that AMD have a simpler lineup but when the rival is flooding the market with products and mindshare is hard to compete.

For the last couple of generations that is true, the same could be said about Intel vs AMD. But that AMD has closed the gap to nvidia significantly in this generation, with a radically new way of designing a GPU, tells me that AMD is not going to let nvidia win the battle easily. So, Yay for competition

 

[Panino Manino]

For the last couple of generations that is true, the same could be said about Intel vs AMD. But that AMD has closed the gap to nvidia significantly in this generation, with a radically new way of designing a GPU, tells me that AMD is not going to let nvidia win the battle easily. So, Yay for competition

Of course they'll tried, but will failt. Sheng Tsung is strong and always have some trickes on his sleeve.
RTG is still doing what they can do and only what they can do, Nvidia will not need to give it everything and save some ammunition for a long time still.

 

[DiogoDX]

I also think that the 2.5X - 3X will be for RT games.

 

[darkswordsman17]

Of course they'll tried, but will failt. Sheng Tsung is strong and always have some trickes on his sleeve.
RTG is still doing what they can do and only what they can do, Nvidia will not need to give it everything and save some ammunition for a long time still.

Yes yes, and we had people like you swearing that there was absolutely no way AMD could have even closed the gap with RDNA2. And they said the same thing with regards to AMD CPUs against Intel.

There's a lot more at play as well. Heck, Nvidia's own focus isn't even on gaming GPUs, its them trying to embed their GPU IP into other markets (notably they're trying to gain dominance by forcing it via ARM most recently). And who cares what Nvidia's GPU can do if its MSRP is considerably higher than a slightly lower performing AMD one, and you can't buy it for even like twice that price.

 

[darkswordsman17]

I also think that the 2.5X - 3X will be for RT games.

I'd be curious if it might be something similar to the new consoles. Maybe embedded NAND or a PCIe 4 (or 5?) slot on the back of the card, enabling massive asset streaming boost. Plus the 16X slot of PCIe 4 already wasn't being fully utilized by games I don't think, so it should have plenty extra to function to the CPU, but for gaming it could have an even shorter and faster path (lower latency), enabling massive swaps to/from memory.

 

[exquisitechar]

Of course they'll tried, but will failt. Sheng Tsung is strong and always have some trickes on his sleeve.
RTG is still doing what they can do and only what they can do, Nvidia will not need to give it everything and save some ammunition for a long time still.

Sounds like the pessimism surrounding RDNA2 all over again.

I'm pretty sure that AMD will take the performance crown.

 

[Gideon]

Sounds like the pessimism surrounding RDNA2 all over again.

I'm pretty sure that AMD will take the performance crown.

Well yeah, if RDNA3 is indeed chiplet based and released roughly in a year and Lovelace is not, it will almost certainly be faster. At least in rasterisation.

That won't stop Nvidia from spoiling the market with proprietary shenanigans and software locks though ...

Hopefully Intel and AMD can find some common ground to counter Nvidia's vendor lock-in. It's ridiculous how badly UE 4 is optimized for AMD on PC for instance , considering it has Console fast paths.

 

[Olikan]

Theoricaly the 6900xt is 2,6 times faster than the 5700xt from a pure Gflop persperctive, never mind the changes at ROPs, geometry and efective bandwidth... yet at 4K resolution is baraly 1.8 times faster

Diminishing gains is gonna hit even harder past 80CUs, so does clocks speed, Infinity cache hit rates is gonna plateau soon

Imho leakers are just pointing theorical performance increase, real performance is not gonna be near that

 

[Ajay]

Sounds like the pessimism surrounding RDNA2 all over again.

I'm pretty sure that AMD will take the performance crown.

IDK, Nvidia isn't the same clusterfk that Intel is right now. Nvidia also does a great job retaining highly talented engineers. IIRC, part of NV's problem was that Samsung wasn't going to deliver its N7(EUV) node at HVM in time for Ampere. So NV had to go with the 'special' 8nm node. Too bad really; the world could use two top level Fab companies.

 

[eek2121]

Pretty simple to achieve:

• 50% higher perf/watt
• twice the CUs thanks to a chiplet based design
• slightly higher TDP
• large cache combined with a slightly larger bus
• simplified design allowing for 3+ ghz clocks

 

[Timorous]

Theoricaly the 6900xt is 2,6 times faster than the 5700xt from a pure Gflop persperctive, never mind the changes at ROPs, geometry and efective bandwidth... yet at 4K resolution is baraly 1.8 times faster

Diminishing gains is gonna hit even harder past 80CUs, so does clocks speed, Infinity cache hit rates is gonna plateau soon

Imho leakers are just pointing theorical performance increase, real performance is not gonna be near that

In the TPU 4k suite it is 2x faster at 4k and that suite is pretty well balanced.

 

[GodisanAtheist]

Pretty simple to achieve:

• 50% higher perf/watt
• twice the CUs thanks to a chiplet based design
• slightly higher TDP
• large cache combined with a slightly larger bus
• simplified design allowing for 3+ ghz clocks

- Yeah its a piece of cake, I know NV and AMD have been pouring considerable R&D resources into this but really anyone can whip up this kind of improvement in their garage on a lazy Sunday

 

[eek2121]

- Yeah its a piece of cake, I know NV and AMD have been pouring considerable R&D resources into this but really anyone can whip up this kind of improvement in their garage on a lazy Sunday

🤣

Hey, I didn’t say it was easy! I am expecting a non-chiplet design that is 50% faster (IIRC AMD said something about a 50% perf/watt increase, which is likely coming from a new node).

 

[Panino Manino]

Sounds like the pessimism surrounding RDNA2 all over again.

I'm pretty sure that AMD will take the performance crown.

I have no doubts that RDNA3 will be as good as people expect, but that the field will be reversed? No.

 

[eek2121]

The “rumor” also claims late next year. What are we getting before that?

 

[Mopetar]

Sounds like the pessimism surrounding RDNA2 all over again.

I'm pretty sure that AMD will take the performance crown.

RDNA2 showed us that past performance isn't always a good predictor for future products. If you really learned anything from that you should think it would be as silly to expect AMD to have another massive leap on that same level.

 

[Ajay]

The “rumor” also claims late next year. What are we getting before that?

Uhm, more unit volume would be good; which starts happening in the second half of this year, I think.
If AMD can sell all the GPUs it can make - seems like it would be wise to just wait till RDNA3 is ready.

 

[biostud]

The “rumor” also claims late next year. What are we getting before that?

If we are lucky 6xxx at MSRP

 

[GodisanAtheist]

I do wonder if the current supply situation will have both NV and AMD extending the product cycle out a bit further for this gen.

If people are willing to buy your old stuff at MSRP two years down the line, might as well sit on the next gen a bit longer and start working on the next thing.

NV and AMD are acutely aware of that the end is silicon is nigh, no point rushing to get there first when we don't have a good replacement in the pipe.

 

[Timorous]

I do wonder if the current supply situation will have both NV and AMD extending the product cycle out a bit further for this gen.

If people are willing to buy your old stuff at MSRP two years down the line, might as well sit on the next gen a bit longer and start working on the next thing.

NV and AMD are acutely aware of that the end is silicon is nigh, no point rushing to get there first when we don't have a good replacement in the pipe.

Depends on capacity. Sure they could keep pumping out 6k series cards and Zen 3 parts on 7nm but if they can gain more capacity and sell more units by having a 5nm 7k series GPU and Zen 4 it might be better to move to that as fast as possible to get the volume.

I would expect AMD want to increase orders with OEMs and increase those kinds of partnerships so increasing the number of units produced is probably better for AMD right now. If they were in Intels or Nvidias position then holding out a bit longer with the current gen stuff could work fine.

EDIT:

RDNA3 2.5x - 3x perf increase speculation. I am very doubtful for pure raster performance. Even with another 50% perf/watt increase which I see as entirely possible it would still top out at around 2x with a 400W TDP while making the assumption that the perf : power scaling from 6700XT to 6900XT holds true for 7700XT to 7900XT and the 7700XT has similar to 6900XT performance.

I do see it as more possible if this is with RT turned on but I think it still needs a TDP increase to at least 350W to pull off.

I guess the other option is 5nm may give AMD the ability to do much better than the 50% perf/watt target but I only see that as possible if they go wide / slow so you have 3 GPU chiplets + IO die for 240 CUs at conservative clock speeds.

I really don't think AMD can get more than a 2x raster uplift and I think doing that will need a large TDP increase. 2.5x performance with RT on I see as far more doable. This is all subject to change when AMD announce their perf/watt target for RDNA3 because their previous public targets have been on the money and we can do some napkin stuff to get ballpark performance at various TDPs.

 

[eek2121]

Depends on capacity. Sure they could keep pumping out 6k series cards and Zen 3 parts on 7nm but if they can gain more capacity and sell more units by having a 5nm 7k series GPU and Zen 4 it might be better to move to that as fast as possible to get the volume.

I would expect AMD want to increase orders with OEMs and increase those kinds of partnerships so increasing the number of units produced is probably better for AMD right now. If they were in Intels or Nvidias position then holding out a bit longer with the current gen stuff could work fine.

EDIT:

RDNA3 2.5x - 3x perf increase speculation. I am very doubtful for pure raster performance. Even with another 50% perf/watt increase which I see as entirely possible it would still top out at around 2x with a 400W TDP while making the assumption that the perf : power scaling from 6700XT to 6900XT holds true for 7700XT to 7900XT and the 7700XT has similar to 6900XT performance.

I do see it as more possible if this is with RT turned on but I think it still needs a TDP increase to at least 350W to pull off.

I guess the other option is 5nm may give AMD the ability to do much better than the 50% perf/watt target but I only see that as possible if they go wide / slow so you have 3 GPU chiplets + IO die for 240 CUs at conservative clock speeds.

I really don't think AMD can get more than a 2x raster uplift and I think doing that will need a large TDP increase. 2.5x performance with RT on I see as far more doable. This is all subject to change when AMD announce their perf/watt target for RDNA3 because their previous public targets have been on the money and we can do some napkin stuff to get ballpark performance at various TDPs.

Well they are moving from 7nm to 5nm and doubling the CU count, so anything is possible.

We also know they are working with the CPU design teams.

 

[Mopetar]

Well they are moving from 7nm to 5nm and doubling the CU count, so anything is possible.

That will certainly increase the theoretical peak performance, but we've seen time and time again that few titles are capable of utilizing twice as many shaders to produce a similar gain in performance. In some cases there may be other bottlenecks that limit performance first, but in a lot it's just an inability to find some meaningful work for all of those extra shaders to do or to be able to evenly distribute the work across all of them so that there isn't as much time where the others are waiting on a few shaders to finish up their work.

I suppose if 8K gaming takes off it will be more important to have a massive amount of shaders simply because there are that many more pixels, but I wouldn't expect any kind of massive performance jump at 4K or 1440p.

 

[GodisanAtheist]

Depends on capacity. Sure they could keep pumping out 6k series cards and Zen 3 parts on 7nm but if they can gain more capacity and sell more units by having a 5nm 7k series GPU and Zen 4 it might be better to move to that as fast as possible to get the volume.

I would expect AMD want to increase orders with OEMs and increase those kinds of partnerships so increasing the number of units produced is probably better for AMD right now. If they were in Intels or Nvidias position then holding out a bit longer with the current gen stuff could work fine.

EDIT:

RDNA3 2.5x - 3x perf increase speculation. I am very doubtful for pure raster performance. Even with another 50% perf/watt increase which I see as entirely possible it would still top out at around 2x with a 400W TDP while making the assumption that the perf : power scaling from 6700XT to 6900XT holds true for 7700XT to 7900XT and the 7700XT has similar to 6900XT performance.

I do see it as more possible if this is with RT turned on but I think it still needs a TDP increase to at least 350W to pull off.

I guess the other option is 5nm may give AMD the ability to do much better than the 50% perf/watt target but I only see that as possible if they go wide / slow so you have 3 GPU chiplets + IO die for 240 CUs at conservative clock speeds.

I really don't think AMD can get more than a 2x raster uplift and I think doing that will need a large TDP increase. 2.5x performance with RT on I see as far more doable. This is all subject to change when AMD announce their perf/watt target for RDNA3 because their previous public targets have been on the money and we can do some napkin stuff to get ballpark performance at various TDPs.

- Hmmm, we all assume RDNA3 is going to be the first arch on 5nm, but its possible we just get RDNA2 on steroids first, then the "tock" happens later and we get the new arch on the same 5nm node. Maybe keep the lower, less margin oriented end of the series on the old node.

That's one way to really squeeze the most out of your R&D dollar.