Discussion RDNA4 + CDNA3 Architectures Thread

[DisEnchantment]

With the GFX940 patches in full swing since first week of March, it is looking like MI300 is not far in the distant future!
Usually AMD takes around 3Qs to get the support in LLVM and amdgpu. Lately, since RDNA2 the window they push to add support for new devices is much reduced to prevent leaks.
But looking at the flurry of code in LLVM, it is a lot of commits. Maybe because US Govt is starting to prepare the SW environment for El Capitan (Maybe to avoid slow bring up situation like Frontier for example)

See here for the GFX940 specific commits

History for llvm/lib/Target/AMDGPU - llvm/llvm-project

The LLVM Project is a collection of modular and reusable compiler and toolchain technologies. - History for llvm/lib/Target/AMDGPU - llvm/llvm-project

github.com

Or Phoronix

More AMD "GFX940" Enablement Work Landing In LLVM - Phoronix

www.phoronix.com

There is a lot more if you know whom to follow in LLVM review chains (before getting merged to github), but I am not going to link AMD employees.

I am starting to think MI300 will launch around the same time like Hopper probably only a couple of months later!
Although I believe Hopper had problems not having a host CPU capable of doing PCIe 5 in the very near future therefore it might have gotten pushed back a bit until SPR and Genoa arrives later in 2022.
If PVC slips again I believe MI300 could launch before it

This is nuts, MI100/200/300 cadence is impressive.



Previous thread on CDNA2 and RDNA3 here

Question - Speculation: RDNA3 + CDNA2 Architectures Thread

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...

forums.anandtech.com

 

[SolidQ]

Or the leakers are just making assumptions based upon raw specs that aren't true.

Maybe @Kepler_L2 can give answer, WGP is reworked or not

 

[CouncilorIrissa]

So, uh...

https://twitter.com/x/status/1776244864856051785

Navi 48:

32
64
256
693
2770
~240 mm²

Navi 44

16
32
128
288
515
~130 mm²


Navi 44 die size 130mm²... WTF

>7600 performance in a 130mm² die size??? WOW!!!

So, is the 515 a typo and it should 2515 instead?
I assume these are base clocks and not boost (or whatever the hell AMD are calling them now) ones.

 

[branch_suggestion]

So, uh...

https://twitter.com/x/status/1776244864856051785

Navi 48:

32
64
256
693
2770
~240 mm²

Navi 44

16
32
128
288
515
~130 mm²


Navi 44 die size 130mm²... WTF

>7600 performance in a 130mm² die size??? WOW!!!

240/140mm^2 was my estimate.
So if it is 4/2SE instead of 2/1, extremely impressive area eff.
First number is WGP
Second is CU and probably Infinity Cache too
Third is bus width
Fourth is memory bandwidth, 21.7Gbps GDDR6 for N48, 18Gbps GDDR6 for N44, same as N33.
Final number appears to be effective memory bandwidth, basically confirming IC is the same as N32/33 with a clock bump. https://www.amd.com/en/products/graphics/amd-radeon-rx-7800-xt https://www.amd.com/en/products/graphics/amd-radeon-rx-7600-xt

So, is the 515 a typo and it should 2515 instead?
I assume these are base clocks and not boost (or whatever the hell AMD are calling them now) ones.

AMD no longer gives base clocks (or just for RDNA3 as base clocks were weird) Just game (typical median) clocks and boost (typical peak).

 

[CouncilorIrissa]

240/140mm^2 was my estimate.
So if it is 4/2SE instead of 2/1, extremely impressive area eff.
First number is WGP
Second is CU and probably Infinity Cache too
Third is bus width
Fourth is memory bandwidth, 21.7Gbps GDDR6 for N48, 18Gbps GDDR6 for N44, same as N33.
Final number appears to be effective memory bandwidth, basically confirming IC is the same as N32/33 with a clock bump. https://www.amd.com/en/products/graphics/amd-radeon-rx-7800-xt https://www.amd.com/en/products/graphics/amd-radeon-rx-7600-xt

AMD no longer gives base clocks (or just for RDNA3 as base clocks were weird) Just game (typical median) clocks and boost (typical peak).

Yea, the number looks like IC capacity, otherwise it does not really make sense to list that number as 1WGP = 2CU anyway.

2770 MHz sounds slow-ish? I was expecting them to finally meet the 3GHz clk meme (unless there are serious IPC gains compared to RDNA3?)

 

[branch_suggestion]

Yea, the number looks like IC capacity, otherwise it does not really make sense to list that number as 1WGP = 2CU anyway.

2770 MHz sounds slow-ish? I was expecting them to finally meet the 3GHz clk meme (unless there are serious IPC gains compared to RDNA3?)

The last number cannot be clock speeds as AMD has made them non static, plus each SE and the frontend have their own clock domains on RDNA3.
Both numbers line up with effective memory bandwidth.
Oh, and clocks are probably not final yet anyway.

 

[CouncilorIrissa]

The last number cannot be clock speeds as AMD has made them non static, plus each SE and the frontend have their own clock domains on RDNA3.
Both numbers line up with effective memory bandwidth.

Ah, I misread your previous post. Thanks.

 

[SolidQ]

So if it is 4/2SE instead of 2/1, extremely impressive area eff

Here is

 

[adroc_thurston]

It is a driver issues if the adrenaline software package does not function correctly with stable overclocks

It's not a driver issue, AMD has a problem with resets

Or the leakers are just making assumptions based upon raw specs that aren't true.

Oh but it is a new WGP.

 

[leoneazzurro]

>7600 performance in a 130mm² die size??? WOW!!!

Well, 7600 is a N6 part with die size around 200mm^2, N44 should be on N4

 

[Mopetar]

Regardless of whether all of those leaks are spot on or not, they're still missing the most important number of all, which is cost.

We've seen that when AMD has good prices they succeed in the market. Let's hope they've learned from this experience.

 

[branch_suggestion]

Assuming this leak is legit, I guess I need to kneel to Teska.
Oh well, I guess it was a tad too soon for GDDR7, but for AMD to fit so much in such a small space is outstanding, I was expecting a more exotic memory subsystem but it turns out it is basically N32/33 after L2. Looking forward to uncore changes too, and how much of that uncore is being gutted in N44.
The meat is of course the WGP/SA/SE changes.

 

[SolidQ]

AMD has a problem with resets

They can fix it? RDNA4 etc

 

[Timorous]

Regardless of whether all of those leaks are spot on or not, they're still missing the most important number of all, which is cost.

We've seen that when AMD has good prices they succeed in the market. Let's hope they've learned from this experience.

RV770 was not great for AMD. It gave them marketshare but not enough to offset the low margins. They would have been better off selling it far more in line with the perf/% of NVs cards.

If N48 can offer 7900XT performance then only expect a small price advantage despite the cheaper die because AMD would much rather have the margin than get into a price competition with NV. Especially since a lot of buyers don't want to buy AMD they just want to buy NV for less (or get a higher tier at the same price).

 

[SolidQ]

Fourth is memory bandwidth, 21.7Gbps GDDR6 for N48, 18Gbps GDDR6 for N44,

Leaker replying to this

 

[Mahboi]

>7600 performance in a 130mm² die size??? WOW!!!

Is it? I thought they aimed higher, closer to a 7700 xt than a 7600, but that seems excessive. 128 bit bus is not surprising but a bit disappointing if it's not on GDDR7 and doesn't get a 12/10Go version.

The current (botched ofc) N33 is 202mm² I believe.
So...we're looking at a possible ~50% improvement, plus a 20% general improvement from the lost perf in RDNA 3.
Seems consistent with earlier RDNA gens, no? 50% per gen? Although that's perf not area.

 

[SolidQ]

doesn't get a 12/10Go version.

at least should be 8/16, If they going 16GB version for 300$ with around 7700XT perf, gonna be nice card for 1080p

anyway leakers still hiding from us RT perfomance

 

[Ghostsonplanets]

https://twitter.com/x/status/1776266835002077590

The Navi 4 follows the Navi 3 by removing the older design to improve density.

Is it? I thought they aimed higher, closer to a 7700 xt than a 7600, but that seems excessive. 128 bit bus is not surprising but a bit disappointing if it's not on GDDR7 and doesn't get a 12/10Go version.

The current (botched ofc) N33 is 202mm² I believe.
So...we're looking at a possible ~50% improvement, plus a 20% general improvement from the lost perf in RDNA 3.
Seems consistent with earlier RDNA gens, no? 50% per gen? Although that's perf not area.

Navi 44 should be rough in-line with 4060 Ti/Radeon 6750XT according to rumors iirc. Maybe even the 7700XT.

130mm² size might be in part due to N4P manufacturing node, but it's still very impressive. We need to remember that this is a monolithic die with a huge amount of cache onboard, which doesn't shrink much from N6 -> N5. That AMD still was able to get such level of reduction in are is quite impressive imo.

I really wonder how AMD will price N44. 130mm² is so small that they would be able to release a very cheap part. Maybe the return of $199 budget GPUs? Would be amazing.

 

[Mahboi]

AMD would much rather have the margin than get into a price competition with NV. Especially since a lot of buyers don't want to buy AMD they just want to buy NV for less (or get a higher tier at the same price).

This market is never getting healed with this situation. The NV mind control is so damn prevalent.

at least should be 8/16, If they going 16GB version for 300$ with around 7700XT perf, gonna be nice card for 1080p

anyway leakers still hiding from us RT perfomance

I was more thinking of a GDDR7 version with 2Go and later 3Go chips, getting us 8/12 and a cut down to 10. I don't think 16Go on a card this small is serious, it's quite a waste.

I really wonder how AMD will price N44. 130mm² is so small that they would be able to release a very cheap part. Maybe the return of $199 budget GPUs? Would be amazing.

I don't buy it at all.
Their current low end offerings are overpriced and just bad. If they have a really good part, they're certainly not doing it for cheap, at least not for DIY.
It may be very very very interesting to see it in laptops though, wonder what price they'll offer for OEMs.

 

[branch_suggestion]

https://twitter.com/x/status/1775919013106790743

Hmm

 

[igor_kavinski]

The NV mind control is so damn prevalent.

We can't do anything in laptops but in desktop, I personally have shunned Geforce 4000 series until I find something used at a throwaway price. If I don't, I won't care coz I already got a cheap 3090 for $600 right after the mining market crashed. Maybe this is what people should do to knock some sense into Jensen. Keep buying their older stuff used until they panic due to lack of sales and drop their current gen prices.

 

[igor_kavinski]

It's not a driver issue, AMD has a problem with resets

Is that specifically a GPU silicon problem or something to do with some component on the PCB that AMD has chosen for cost reasons?

 

[igor_kavinski]

130mm² size might be in part due to N4P manufacturing node, but it's still very impressive. We need to remember that this is a monolithic die with a huge amount of cache onboard, which doesn't shrink much from N6 -> N5. That AMD still was able to get such level of reduction in are is quite impressive imo.

Small size, higher logic density and lots of cache? Sounds like a recipe for extreme heat output. How are they gonna cool this scorching hot piece of silicon beauty?

 

[Ghostsonplanets]

Small size, higher logic density and lots of cache? Sounds like a recipe for extreme heat output. How are they gonna cool this scorching hot piece of silicon beauty?

Navi 24 is ~107mm² and clocks at 2.8GHz. Ryzen 7545U iGPU has 2 WGPs clocking at 2.8GHz too. I'm not worried.

Navi 44 will probably clock at 3GHz+. And I don't doubt, by the time Navi 5 comes, we'll be seeing GPUs being clocked at ~4GHz.

AMD cross-pollination between CPU and GPU divisions will flourish in a beautiful way.

 

[CouncilorIrissa]

This market is never getting healed with this situation. The NV mind control is so damn prevalent.

I look at the "8/12 GB VRAM not being" enough thread and feel that the consumer GPU market deserves whatever's being done to it by Jensen.

 

[SolidQ]

Now we need to know, last part, it's RT/PT Ampere or Ada level.