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

 

[Glo.]

If only there was a way to double memory bandwidth...

LPDDR5X 8533 MHz on a 128 bit bus has 136 GB/s bandwidth.

Thats nearly as much as Navi 24 has on a 64 bit GDDR6 memory bus(144 GB/s). N24 obviously has 16 MB of IC and is bound to 4 GB VRAM, which unified memory architecture will not be bound by .

Also - caches on the die are also going to be larger, like L2 has to double from 4 to 8 MB, again. If the IC will be 32 MB - it should be fine. If, however, AMD will go for something more, like 64 MB IC - it will have more than enough bandwidth to feed both: CPUs and iGPU even at the same time.

 

[Kaluan]

ComputerBase tested Adrenaline 22.12.2
Playing youtube reduced consumption:
7900XT: 71W -> 46W
7900XTX: 80 -> 54W
But It's still higher than N21, If you also enable HDR, then It's comparable.

Power consumption didn't change while playing, but limiting performance to 144 FPS decreases power draw much more than before.
Maybe some of you still remember that debate, where N31 ended up less efficient than N21 by limiting FPS to 144.
View attachment 74524View attachment 74525

Well now, how cute of them... only 3 weeks late lol

Ancient Gameplays already investigated 22.12.2 (7900) improvements mere days after the driver was out tho, I've posted about it in the RX 7900 review thread awhile back:

With what drivers? The ones released 5-6 days ago (22.12.2) seem to overhaul clocking behavior. Better temps, power draw behavior and even some performance uplifts (CP2077 1% lows are a massive 22% higher). Going by Fabio/Ancient Gameplay's testing:

View attachment 73500View attachment 73497View attachment 73498View attachment 73499

Results are of from his tuned/OC 400W PL (Sapphire) reference 7900 XTX, unlocked framerate first, 100fps cap second. CP2077 raster only, maxed.

Fabio (Ancient Gameplays) also has a short video up on the newer 23.1.1 (7900), on the subject of power efficiency improvements, here's what he found:


Full video here:

 

[Kaluan]

Also - caches on the die are also going to be larger, like L2 has to double from 4 to 8 MB, again. If the IC will be 32 MB - it should be fine. If, however, AMD will go for something more, like 64 MB IC - it will have more than enough bandwidth to feed both: CPUs and iGPU even at the same time.

Smart Access Cache incoming?

A.M.D... it's in the BAG SAC!

 

[Mopetar]

I hope the team working on it doesn't completely miss expectations. I'd hate to see them all get . . . fired.

No, that's not quite right. Terminated!

Hmm. . . . Made redundant?

I swear I had something for this.

 

[TESKATLIPOKA]

As @moinmoin said, 24CU is a lot. That's 2x more than what Phoenix has and at 2.5GHz It would provide the same TFlops as 7600S, at 3.3GHz It would compare to 7700S.
Even If Strix Point used 3nm process, that's only ~25-30% lower power consumption compared to 4nm. 35-45W TDP is clearly not enough, you would need to lower clocks or increase TDP compared to PHX.

Then the largest problem is BW. 128-bit LPDDR5 won't give you enough bandwidth to feed this.
32MB IC is not enough to compensate in my opinion, you would need 64-72MB, that's >70% hitrate at Full HD.

I made a table of Strix Point models in the correct thread. I even calculated the needed amount of IC as part of LLC for GPU. It looks interesting.

 

[igor_kavinski]

I swear I had something for this.

Demoted and then they resign to go work under Raja Koduri, which is worse than demotion but Raja promises them VegaNext in 2025 that is finally shipped in 2030!

7 years of VIP gainful employment. What's not to like?

 

[TESKATLIPOKA]

I made a BW table for Infinity cache 2.

Infinity cache 2	16 MB	32 MB	48 MB	64 MB	96 MB	128 MB (doesn't exist)
B/CLK	384	768	1152	1536	2304	3072
Theoretical BW at 2.5GHz	960 GB/s	1920 GB/s	2880 GB/s	3840 GB/s	5760 GB/s	7680 GB/s
Hit rate at 1080p (BW)	37 % (355 GB/s)	55 % (1056 GB/s)	65 % (1872 GB/s)	72 % (2765 GB/s)	78 % (4493 GB/s)	80 % (6144 GB/s)
Hit Rate at 1440p (BW)	23% (221 GB/s)	38 % (730 GB/s)	48 % (1382 GB/s)	57 % (2189 GB/s)	69 % (3974 GB/s)	74 % (5683 GB/s)
Hit rate at 4K (BW)	19 % (182 GB/s)	27 % (518 GB/s)	34 % (979 GB/s)	42 % (1613 GB/s)	53 % (3053 GB/s)	62 % (4762 GB/s)

BTW 6MB L2 in RDNA3 at 2.5GHz has 7680 GB/s.
AMD did a good job with this.

P.S. I think Infinity cache 2 is working at ~1.9GHz based on this slide.

5300-960 = 4340 GB/s / 2304 B/clock => 1.88GHz
Similar speed as the first Infinity cache.
5300/2304 = 2.3GHz is the correct value.

 

[TESKATLIPOKA]

I just love this.

 

[Glo.]

So if your math is correct, 32 MB IC2 for Strix Point would have 1.65 TB/s bandwidth.

Very interesting.

Edit: ah heck. Only now I see that I missclicked, and wrote 1.65 instead of 1.75 TB/s.

 

[TESKATLIPOKA]

So if your math is correct, 32 MB IC2 for Strix Point would have 1.65 TB/s bandwidth.

Very interesting.

How did you get that number?

 

[Glo.]

How did you get that number?

I divided 96 MB and 5.3 TB/s by 3, which resulted in 32 MB IC and 1.75 TB/s.

IC on Strix Point should be executed the same way as it is on Navi 31 chips.

 

[TESKATLIPOKA]

I divided 96 MB and 5.3 TB/s by 3, which resulted in 32 MB IC and 1.75 TB/s.

IC on Strix Point should be executed the same way as it is on Navi 31 chips.

That's wrong, because It includes BW from GDDR6.
5300-960 = 4340 GB/s is BW for 96MB infinity cache.
I just don't know If this is the theoretical max or not. I need to find endnote RX-818

edit: I found It.

It looks like It's just max BW of Infinity cache without GDDR6.
So what you wrote is correct for 32MB, but effective BW will be lower because It depends on actual hitrate.

 

[Glo.]

That's wrong, because It includes BW from GDDR6.
5300-960 = 4340 GB/s is BW for 96MB infinity cache.
I just don't know If this is the theoretical max or not. I need to find endnote RX-818

edit: I found It.
View attachment 74955
It looks like It's just max BW of Infinity cache without GDDR6.
So what you wrote is correct for 32MB, but effective BW will depend on actual hitrate.

I mean, the Bandwidth was always separate for IC and for GDDR memory. We knew this since Navi 21/22/23/24.

 

[TESKATLIPOKA]

I mean, the Bandwidth was always separate for IC and for GDDR memory. We knew this since Navi 21/22/23/24.

AMD used both for effective bandwidth. 1152 GB/s + 512 GB/s for 1664 GB/s.

 

[Glo.]

You still don't believe that Strix Point iGPU, if it has 24 CUs and 32 MB IC is a game changer for iGPUs?

 

[TESKATLIPOKA]

You still don't believe that Strix Point iGPU, if it has 24 CUs and 32 MB IC is a game changer for iGPUs?

It's only a rumour, we know how wrong such rumours were about RDNA3 just a month or two before release, Strix Point is still far away.
The rumours about N4 process and being monolithic would make this unrealistic paired with bigger Zen5 cores. Size would be clearly >200mm2.

TDP is also a problem. N4 Phoenix with 12CU boosting to 3GHz has 45W TDP. Game frequency is unknown.
2x bigger IGP at N4 would be impossible unless you significantly reduce clocks.
Even N3(E) with 30-34% reduction in power compared to N5 won't be enough to fit It in 45W If the power draw ratio during gaming is 1:2(15W:30W).
If you can't keep the frequency, then It's kinda pointless to put so much CU in It, 20CU would be a better option.

I still question If 32MB with 55% hitrate is enough in combination with LPDDR5(DDR5). There is no problem with BW of IC.
I would rather have a bigger dynamic SLC(LLC) for both CPU and GPU.

 

[insertcarehere]

You still don't believe that Strix Point iGPU, if it has 24 CUs and 32 MB IC is a game changer for iGPUs?

Given Phoenix is already 180mm^2 on N4 with Zen 4 cores, half the CUs and no infinity cache. Doubling the CUs with Zen 5 + IC would make SP very large on N4 and therefore relatively expensive to produce. Such size needs to be compensated for with a price premium, especially for AMD when they can happily pump out CPU chiplets on such silicon instead.

 

[beginner99]

You still don't believe that Strix Point iGPU, if it has 24 CUs and 32 MB IC is a game changer for iGPUs?

Only because there is basically no reasonable dGPU in the low end. That market was completely abandoned by NV and AMD. Look at the crap the 6500 and 3050 are. 6500 is barely better than a 290x (which I own) and cost on release the same I paid for said 290x about 8 years ago. yeah it's this bad. so easy for anything affordable to shine in that performance bracket.

 

[moinmoin]

You still don't believe that Strix Point iGPU, if it has 24 CUs and 32 MB IC is a game changer for iGPUs?

I'm already seeing it coming, it even having as many models and as good availability as Rembrandt.

 

[IntelUser2000]

It's only a rumour, we know how wrong such rumours were about RDNA3 just a month or two before release, Strix Point is still far away.

And besides RGT is the same "source" that claimed 2.5x performance of RDNA3 over the predecessor! In reality, sometimes it's half that at 1.25x.

I don't believe in the practicality of big iGPUs because you pretty much lose the big reason you get them in the first place - practically free! And you often don't even get the power advantage either, that's a hit and miss, just toss a coin to decide which.

Actually even current "iGPUs" arguably cost way more than the ones that were bundled with chipsets, back when memory controllers were off die. Because now they bundle it with higher CPUs. You are kinda forced to pay extra if you want the best iGPU.

And it only takes about a single generation before the iGPUs go back down to the bottom of the barrel in terms of dGPUs.

 

[TESKATLIPOKA]

Only because there is basically no reasonable dGPU in the low end. That market was completely abandoned by NV and AMD. Look at the crap the 6500 and 3050 are. 6500 is barely better than a 290x (which I own) and cost on release the same I paid for said 290x about 8 years ago. yeah it's this bad. so easy for anything affordable to shine in that performance bracket.

Phoenix with 12CU IGP at 3GHz could be pretty close to even 6500XT, If It wasn't bottlenecked.
It's not that phoenix IGP is so good, but that N24 is too weak because It's too small.
It is only 107mm2 on 6nm, The production cost of that chip is not more than $15. AMD could have made a 24CU, 24MB IC, 96bit GDDR6, which should still be ~150mm2 and cost of die would be only $5 more. With this performance they could have kept MSRP at $199 while still being profitable and everyone would sing odes how great a company AMD is.
But no, AMD had to save a few bucks on cost and we ended up with a pretty bad product.

P.S. RX 6500XT 8GB is a lot better than the 4GB version.

 

[Kronos1996]

Phoenix with 12CU IGP at 3GHz could be pretty close to even 6500XT, If It wasn't bottlenecked.
It's not that phoenix IGP is so good, but that N24 is too weak because It's too small.
It is only 107mm2 on 6nm, The production cost of that chip is not more than $15. AMD could have made a 24CU, 24MB IC, 96bit GDDR6, which should still be ~150mm2 and cost of die would be only $5 more. With this performance they could have kept MSRP at $199 while still being profitable and everyone would sing odes how great a company AMD is.
But no, AMD had to save a few bucks on cost and we ended up with a pretty bad product.

P.S. RX 6500XT 8GB is a lot better than the 4GB version.

A 150mm2 die makes no sense when they knew they had a 200mm2 Navi 33 coming that blows it away in performance. A stripped down cheap Navi 24 die that can be produced in massive quantities is perfect budget laptop fodder and will be around for years. Their only mistake was releasing it on desktop. 8GB of GDDR6 costs 2-3X more then the silicon, never-mind the PCB/cooler, so that’s not really a major cost factor here.

 

[coercitiv]

In reality, sometimes it's half that at 1.25x.

They only need to fix the thing to get 2X up and running. I'm surprised you don't know about the thing.

Seriously now, there's some spectacular commitment to performance estimates in this thread, the kind of napkin math I really appreciate. The only problem we have is feeding these wonderful human calculators with "leaked" info that varies from optimistic speculation to pure fabrication. It doesn't matter if folks around here are dedicated, feeding crap at one end results in crap the other end.

If we don't stop echoing "leakers" like RGT and MLID when it has become painfully obvious their signal/noise ratio is poor, threads like this will continue outputting highly distorted estimates when compared with launched products & technologies.

 

[TESKATLIPOKA]

And besides RGT is the same "source" that claimed 2.5x performance of RDNA3 over the predecessor! In reality, sometimes it's half that at 1.25x.

I don't believe in the practicality of big iGPUs because you pretty much lose the big reason you get them in the first place - practically free! And you often don't even get the power advantage either, that's a hit and miss, just toss a coin to decide which.

Actually even current "iGPUs" arguably cost way more than the ones that were bundled with chipsets, back when memory controllers were off die. Because now they bundle it with higher CPUs. You are kinda forced to pay extra if you want the best iGPU.

And it only takes about a single generation before the iGPUs go back down to the bottom of the barrel in terms of dGPUs.

Any iGPU in an APU increases the cost, so It's never really free, but It's still pretty useful. Especially when your dGPU dies.

I can see a niche market for bigger IGPs, but not for ridiculously big ones.
24CU is still reasonable in my opinion, but It's very questionable If It happens.
TDP will have to increase, but It should still offer better perf/W.
The problem is price as you said.
Production cost would increase only by $20-30, but It will be paired with the fastest CPUs in premium laptops, the cost will be much higher.
Where It could be relatively cheap is desktop. 5600-5700G were sold for a great price.

 

[Kronos1996]

Any iGPU in an APU increases the cost, so It's never really free, but It's still pretty useful. Especially when your dGPU dies.

I can see a niche market for bigger IGPs, but not for ridiculously big ones.
24CU is still reasonable in my opinion, but It's very questionable If It happens.
TDP will have to increase, but It should still offer better perf/W.
The problem is price as you said.
Production cost would increase only by $20-30, but It will be paired with the fastest CPUs in premium laptops, the cost will be much higher.
Where It could be relatively cheap is desktop. 5600-5700G were sold for a great price.

If a 24 CU APU exists it’s probably not Strix Point. It’s likely a semi-custom design for some sort of handheld console or a dedicated design to service that market and perhaps thin and light gaming laptops. It wouldn’t make sense from a cost/performance perspective anywhere else. Hell it might just be an Xbox Series S refresh that also goes into Surface Books.