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

 

[maddie]

SS8 definitely limited the performance of Ampere. Power consumption worse, etc.

Yes, but the final product has that included in the design parameters. You don't need to say it as if it's an independent property. Using SS8, a certain design is possible and using N4 another one is. The final design is what matters. Shaders, memory, power, frequency & price . SS8, N4, N3, N6, N7, SS5, who cares?

 

[beginner99]

Outside of latency (and I suppose multiple PCI bus connections) though what is the difference between coordinating two chiplets in the same PCB or two chiplets on different cards?

What would prevent the same issues with SLI/Xfire implementations from occurring or not needing the additional work from developers to extract the performance gain.

I men the connection is a huge part of that. If designed that way the chiplets can be right next to each other possibly on an interposer which will give much higher bandwidth and lower latency than via pcie.
SLI/Xfire always has been an ugly hack. If you design your gpu as multi-compute chiplet and only that, doesn't need to also work as a single card, then "all you need" (not saying it is easy, not at all but from a high level view it is true) is an IO die that completely hides the fact that there are multiple chiplets from the driver and the OS. It works in hardware and hence no need to expose it and no need for driver fixes for each new game.

In essence like Ryzen just a bit more complex. Do developers need to employ any special tricks to get maximum multi-threaded performance out of the 2 chiplets? No. I do admit it is benefical when the OS is aware of the topology for thread scheduling, eg. bouncing threads from one chiplet to the other is bad but that is handeled by the OS and the average programmer doesn't need to worry about it. In case of the GPU it is howeber not the OS that does the scheduling but the driver (NV) or I think in case of AMD it has been hardware since GCN.

Can someone summarize what is the current "consensus" for N31 vs N21 regarding gaming performance jump based on the unofficial AMD leakers?
Before one year or around that time we had 2.5X-3X depending the leaker (specifically mentioning gaming performance (not FP32) difference...)

The numbers have always been in floating point performance and never gaming performance. Even the 3x ray-tracing figure doesn't mean you get 3x the FPS. From past history I would not expect more than 1.8x for normal games an then a bit more if RT is enabled. it will not be some magical jump.

 

[Timorous]

A 50% perf/watt improvement and a 50% increase in power gets you 2.25x

Edit to add.

6900XT is 2x 5700X in 4k. It has 2.36x more Tflops. Similar scaling for N31s 3x more Tflops vs N21 is 2.5x more perf.

Some obvious assumptions there but it is same ballpark as the perf/watt and TDP increase

 

[ModEl4]

The numbers have always been in floating point performance and never gaming performance. Even the 3x ray-tracing figure doesn't mean you get 3x the FPS. From past history I would not expect more than 1.8x for normal games an then a bit more if RT is enabled. it will not be some magical jump.

No the numbers were not about TF difference (maybe you don't even know how to calculate FP or is it something else?), FP difference became when they realized that this won't fly.
At start they were talking about 2.5GHz with 15360SP (3.33X Terraflop difference not 2.5X) and they were saying 2.5X gaming, deducting -25% from the theoretical flop difference, after a while the frequency became 3GHz in their "predictions" (hence the articles with 4X the Navi 21 TF rating that it was so short-lived since the 12288SP info came so soon) so 4X FP difference with supposedly 3X gaming performance.
Now although we had the SP reduction Graymon55 said that gaming performance difference remained the same as his original prediction, go figure, while others started slowly downscaling (but I didn't see anything specific hence the question) and in order to try save face some of them started the "no we were saying about TF difference" even if the maths easily disprove them!

 

[xpea]

Speaking of yields... if AMD's approach lets them use 400mm2 of TSMC N5 to compete with Nvidia's 600mm2 of N5, the improved yields appear to be in the ~12% range. Not a big amount, but considering that N5 is like 1.8x the cost of N7 (and N6 is supposed to be slightly cheaper than N7), every bit counts.

22mm x 28mm (616 mm2):
View attachment 61641


vs.
17mm x 24mm (408 mm2):

View attachment 61642

Current TSMC D0 on N4 is 0.07 and you get more than 90% yields on AD102-300 (the rumored cut down 126SM out of 144) due to dark silicon, redundancy measures and natural disabling of unused SMs
To say that the price in a mood point on AD102 vs Navi31 due to the high cost of 3D stacking, MCM packaging and interposer on AMD side

 

[Frenetic Pony]

Current TSMC D0 on N4 is 0.07 and you get more than 90% yields on AD102-300 (the rumored cut down 126SM out of 144) due to dark silicon, redundancy measures and natural disabling of unused SMs
To say that the price in a mood point on AD102 vs Navi31 due to the high cost of 3D stacking, MCM packaging and interposer on AMD side

If those weren't cost saving measures vs monolithic they wouldn't exist and the only reason for MCM to exist would be going beyond reticle limit. So maybe we can just logically conclude MCM does save a lot of money vs monolithic, either that or AMD, Intel, and Nvidia judging by future plans are all mistaken together about the exact same thing.

 

[biostud]

If those weren't cost saving measures vs monolithic they wouldn't exist and the only reason for MCM to exist would be going beyond reticle limit. So maybe we can just logically conclude MCM does save a lot of money vs monolithic, either that or AMD, Intel, and Nvidia judging by future plans are all mistaken together about the exact same thing.

I'm sure AMD would like to copy the model of Ryzen to their GPU, where they use the same chiplets with different quantity and a different IO chip. But I guess if it was that simple, they would already have done that

 

[Mopetar]

An N31 with 384 MB of Infinity Cache will be interesting, perhaps for different reasons than most are thinking though.

There are probably a few games that will actually be able to fit entirely within that cache, or nearly so, at least at lower resolutions. I'm curious about what kind of crazy numbers we'll see for FPS, or how much this will shift the bottleneck to the CPU which has also become far more contested as of late.

 

[darkswordsman17]

An N31 with 384 MB of Infinity Cache will be interesting, perhaps for different reasons than most are thinking though.

There are probably a few games that will actually be able to fit entirely within that cache, or nearly so, at least at lower resolutions. I'm curious about what kind of crazy numbers we'll see for FPS, or how much this will shift the bottleneck to the CPU which has also become far more contested as of late.

I'm curious if we might actually see AMD pivot somewhere much more interesting. CDNA is moving to APU chiplets it seems and AMD making all CPUs with GPU (and I assume video processing). I think that would actually make sense for GPUs as well. Where they have fewer but much more gaming performance oriented designs, where being that close (on chip, I'd guess even on die but perhaps separate die) reduces latency and enables the shared cache and memory. Add NAND/storage expansion. And when not-gaming can power the whole thing down and let an efficient APU manage the rest. Plus you run games fully sandboxed with its own dedicated OS further getting rid of ancilliary performance drains.

I think this gives AMD an advantage in that they've already been making such hardware (consoles) for years, only now, they target gaming performance (whereas before they were basically helping console companies leverage AMD's current processing designs into something affordable - often notably removing/downsizing things that boost gaming performance like caches) all out, maximizing the CPU and GPU design for it. It'd make rackmounting for game streaming easy as well, where its holistic (i.e. you provision the CPU and GPU for relative performance, plus memory, etc) and modular. Plus, they should just sell gaming cards as standalone eGPU boxes, and where TB4 becomes an input to the laptop as much as an output. Single cable charges the thin and efficient laptop, while inputting into the laptop display (should you be gaming on the go, otherwise you'd be going externally), with the laptop just handling the comms, overlay stuff. In the future, the laptop can move to wireless charging and we switch to fiber optic for data transmission (possibly wireless laser?).

 

[NostaSeronx]

Just another weird post by Nosta...

~~~~
Navi 30 -> Navi 3X

Tuned specifically for 16K Gaming.
Enhanced RDNA3
6x HBM3 with 24 GB each.
TSMC 3nm

"$4000" if Mass Production
"$6000" if Limited Edition Run
~~~~

Radeon Rage Extreme Max Fury Ultra Deluxe Edition soon?

 

[Karnak]

Chiplets now officially confirmed, also >50% Perf/Watt improvement over RDNA2.

 

[gdansk]

Chiplets now officially confirmed, also >50% Perf/Watt improvement over RDNA2.

So between that and higher TDP it does seem to support the 2x-2.5x performance everyone and their mother has been claiming for some time.

 

[moinmoin]

For CDNA3 the stuff mentioned (if I got it correctly) are bottom stacked V-Cache, 4x perf of CDNA2, 5x perf per watt, 3x perf of Nvidia at half the cost, >8x AI training perf vs MI250X, MI300 combines CDNA3 with Zen 4.

 

[Tuna-Fish]

Chiplets now officially confirmed, also >50% Perf/Watt improvement over RDNA2.

This gets a bit into Kremlinology territory, but I think the precise wording choice between this and the CDNA slide is interesting:

"Advanced Chiplet Packaging" vs "3D Chiplet Packaging". I think CDNA is on an interposer, while RDNA uses the new but much less fancy fanout system.

 

[Ajay]

So between that and higher TDP it does seem to support the 2x-2.5x performance everyone and their mother has been claiming for some time.

Well, 50% more performance per watt *could* translate to 100% more performance at constant watts. Not that it normally works that way. So, maybe 2x is happening, with increased power consumption. 2.5X doesn't seem likely. At least based on the tiny shreds of info we have. AMD is being very coy here, which isn't very helpful to gamers (and I don't think to investors either).

 

[Frenetic Pony]

Well, 50% more performance per watt *could* translate to 100% more performance at constant watts. Not that it normally works that way. So, maybe 2x is happening, with increased power consumption. 2.5X doesn't seem likely. At least based on the tiny shreds of info we have. AMD is being very coy here, which isn't very helpful to gamers (and I don't think to investors either).

This giant press release seems like a "Hey we're competitive so don't purchase that quite yet" type thing, so they'll be pushing as high a set of numbers as they can get away with.

Also I guess they hit that 50% perf per watt improvement over RDNA2, which is, shite, wow. Meaning a hypothetically maxxed out air cooler, currently about 480 watts, could hit over almost triple the performance of a 6900xt. If you pair it with a 512bit bus and 24gbps ram, that could support that as well.

So, uhmmm... yeah.

 

[Ajay]

This giant press release seems like a "Hey we're competitive so don't purchase that quite yet" type thing, so they'll be pushing as high a set of numbers as they can get away with.

Also I guess they hit that 50% perf per watt improvement over RDNA2, which is, shite, wow. Meaning a hypothetically maxxed out air cooler, currently about 480 watts, could hit over almost triple the performance of a 6900xt. If you pair it with a 512bit bus and 24gbps ram, that could support that as well.

So, uhmmm... yeah.

Yes, it's excellent, but it's not a 100% increase in perf/watt - so it will take more perf/clock, or much higher clocks or some combo to get a 2x increase in performance. So, triple the performance of the 6900XT - I'd be surprised; unless it sucked power like a microwave. It the old hockey stick problem - more and more power needs to be pumped in for diminishing gains at some inflection point. Anyway, AMD were very quiet about RDNA3 - so we'll just have to wait and see.

 

[Timorous]

AMD have beaten their perf/watt claims in these sorts of presentations for the last few releases so I expect 50% will be the low end.

RDNA 1 to RDNA 2 saw a 2x performance gain with a >50% perf / watt increase and 33% more TDP.

If RDNA 3 is going up to 450W then that is a 50% higher TDP so a similar >50% perf / watt increase can easily land in the 2 - 2.5x region that has been talked about for a long time.

 

[Heartbreaker]

Well, 50% more performance per watt *could* translate to 100% more performance at constant watts. Not that it normally works that way. So, maybe 2x is happening, with increased power consumption. 2.5X doesn't seem likely. At least based on the tiny shreds of info we have. AMD is being very coy here, which isn't very helpful to gamers (and I don't think to investors either).

No. That's not the math on this at all.

At best it translates int 50% more performance at constant watts.

If it was 100% more performance at constant watts, then it would also be 100% increase in perf/watt.

Generally we are lucky if these claims actually deliver the claimed perf/watt increase at target wattage.

It could deliver 100% more performance if it combines 50% increase in perf/watt, along with 33.3% in power usage, and there is no drop in linearity at peak power levels (which there will be). 1.5*1.333... = 2.0.

 

[Ajay]

No. That's not the math on this at all.

At best it translates int 50% more performance at constant watts.

If it was 100% more performance at constant watts, then it would also be 100% increase in perf/watt.

Generally we are lucky if these claims actually deliver the claimed perf/watt increase at target wattage.

It could deliver 100% more performance if it combines 50% increase in perf/watt, along with 33.3% in power usage, and there is no drop in linearity at peak power levels (which there will be). 1.5*1.333... = 2.0.

Yeah, I was on Vicodin when I posted. That’s all I’m going to say 😬.

 

[Timorous]

No. That's not the math on this at all.

At best it translates int 50% more performance at constant watts.

If it was 100% more performance at constant watts, then it would also be 100% increase in perf/watt.

Generally we are lucky if these claims actually deliver the claimed perf/watt increase at target wattage.

It could deliver 100% more performance if it combines 50% increase in perf/watt, along with 33.3% in power usage, and there is no drop in linearity at peak power levels (which there will be). 1.5*1.333... = 2.0.

AMD claimed RDNA2 was 50% perf/watt. In reality for 30% more TDP than the 5700XT the 6900XT was 2x faster in 4k making it a 54% perf/watt increase.

I suspect this will be similar so based on the numbers AMD just announced I expect ~2.25x performance at 4K, 1.5x perf / watt and 1.5x more power (so 450W).

If you take the 6800 vs 5700XT instead you get 85% more performance at 4K for 9% more TDP which is actually a 70% perf/watt increase. Misread the chart, was comparing 6800 to 5700 performance not 5700XT. vs 5700XT it is still a 50% perf/watt increase.

AMD has historically compared the perf/watt of similar tiers of product where possible and I expect this perf/watt showing is N31 vs the 6900XT or 6950XT. Would love to see the full slide deck with the foot notes included as it may say what the RDNA 2 comparator is.

 

[Heartbreaker]

AMD claimed RDNA2 was 50% perf/watt. In reality for 30% more TDP than the 5700XT the 6900XT was 2x faster in 4k making it a 54% perf/watt increase.

I'm really not concerned about small difference one way or the other. It's clear they will deliver a very nice performance and efficiency gain.

But what I really want is perf/$ gain at the low-midrange area...

 

[jpiniero]

But what I really want is perf/$ gain at the low-midrange area...

Good luck with that. Best bet is used RDNA2 from a miner. Ethereum still needs to fall further to get The Flood though.

 

[Paul98]

This is exciting to see the new packaging! Though I don't need a GPU upgrade as I got a reference 6800xt from AMD, I am excited to see what this generation brings, and what it looks like moving forward.

Nice to see tech moving at such a rapid pace again

 

[Heartbreaker]

Good luck with that. Best bet is used RDNA2 from a miner. Ethereum still needs to fall further to get The Flood though.

Yeah, I know.