Discussion RDNA4 + CDNA3 Architectures Thread

DisEnchantment · Mar 23, 2022

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

adroc_thurston · Dec 28, 2023

Glo. said:
I'd like to remind that RDNA3 was supposed to be that Chungus that will give Nvidia run for its money.

The chungus 3 (36) was canned, champ, because the IP went off the rails.

Tigerick · Dec 30, 2023

Hoho, news of RDNA4 N44 has started appearing: -

Gigabyte leak lists unreleased AMD Radeon RX 7600 XT with 16GB of memory - VideoCardz.com

Gigabyte expects Radeon RX 7600 XT to debut with 16GB memory Not 10 or 12, but 16 GB. In contrast to the EEC leak that indicated PowerColor’s Radeon RX 7600 XT GPU with a 10/12GB memory configuration, Gigabyte has provided updated information suggesting a different scenario. The regulatory...

videocardz.com

We have multiple leaks about 7600XT and this one regarding Gigabyte 7600XT with standard and OC versions pretty much confirm my theory about RDNA4. As I speculated, RDNA4s are designed to be low power monolithic dies to slot between N32 and N33 cards, don't expect much faster rasterization cause AMD won't make RDNA4 faster than N32 cards. OTOH, we should expect better ray tracing performance than 7600 cause 7600's ray tracing performance has been crippled as explained here.

I remembered asking @TESKATLIPOKA about AMD's naming before:-

1. Why did AMD name full die of N33 as RX7600, not 7600XT?

2. Why did AMD name cut-down version of N32 as 7700XT, not 7800 or 7700?

3. Why did AMD launch the high power mobile N31 GPU as 7900M rather than mobile N32?

As for N43, it seems delay a bit...anyhow we should be hearing more leaks in the coming days cause AMD's CES event happen on January 9th and yes, I still think AMD will announce RDNA4 cards in that event....

Tuna-Fish · Dec 30, 2023

Tigerick said:
Hoho, news of RDNA4 N44 has started appearing:

What makes you think that is N44?

Tigerick · Dec 30, 2023

Tuna-Fish said:
What makes you think that is N44?

Cause of two versions available, if AMD use N32/N33, the power requirements are much higher for OC version. That's why we should be expecting OC versions of N43 and N44 due to N4P process and smaller die size....

TESKATLIPOKA · Dec 30, 2023

Tuna-Fish said:
What makes you think that is N44?

That is not N44. I see R76XT written there.

If @Tigerick wants to believe that AMD will call a new generation GPU with higher performance as RX 7600XT, then let him do so. It's not like you can change his mind.

Tigerick · Dec 30, 2023

TESKATLIPOKA said:
That is not N44. I see R76XT written there.
View attachment 91198
If @Tigerick wants to believe that AMD will call a new generation GPU with higher performance as RX 7600XT, then let him do so.

Ho, let's see what the die RX7600XT will be using...

TESKATLIPOKA · Dec 30, 2023

Tigerick said:
Ho, let's see what the die RX7600XT will be using...

RDNA3 based. I personally think It will be N33 + 128-bit 16GB Vram.

Timorous · Dec 30, 2023

Tigerick said:
Ho, let's see what the die RX7600XT will be using...

It really ought to be N32 with 40CUs and 2SEs or 42CUs and 3 SEs, guess that depends on what has been coming off of the line.

Anything else barely makes it better than the current 7600 so a totally pointless release.

Tigerick · Dec 30, 2023

TESKATLIPOKA said:
RDNA3 based. I personally think It will be N33 + 128-bit 16GB Vram.

Yeap, that's most people think.......except 7600 already using full die of N33, that's mean AMD has to pump the clock to justify XT naming. Maybe AMD could stay same speed with double memory size, how about OC version?

And why did AMD launch XT version of same GPU eight months later??

Tigerick · Dec 30, 2023

Timorous said:
It really ought to be N32 with 40CUs and 2SEs or 42CUs and 3 SEs, guess that depends on what has been coming off of the line.

Anything else barely makes it better than the current 7600 so a totally pointless release.

Yeap, that could be potential for RX7700 from another leaks.

But my money is on N43, we shall see...

adroc_thurston · Dec 30, 2023

Tigerick said:
But my money is on N43

What's n43?

Tigerick · Dec 31, 2023

Timorous said:
It really ought to be N32 with 40CUs and 2SEs or 42CUs and 3 SEs, guess that depends on what has been coming off of the line.

Anything else barely makes it better than the current 7600 so a totally pointless release.

You are not the only guess like that, the leaker of Gigabyte 7600XT also speculate so:

They even makeup image with such configuration

God, is that possible? With total die size of 273mm2 (200mm2 build by N5) plus 16GB RAM, the BOM sure would be much higher than N33.....not to mention the OC version will consume more power

I will let you guys think about that...

Tigerick · Dec 31, 2023

TESKATLIPOKA said:
That is not N44. I see R76XT written there.
View attachment 91198
It's not like you can change his mind.

Just noticed that, same words back to you....It is 2024, I am feeling lucky, I still bet RDNA4 dGPU belongs to RX7000 series, since you are so against my theory, you can bet against me, do you want to?

@SteinFG and others are welcome to bet against me if you guys think RDNA4 dGPU NOT belong to RX7000 series

Hans Gruber · Dec 31, 2023

AMD is really disappointing in the GPU sector. They use the cheapest and lease efficient 5nm silicon, while Nvidia used the best 5nm silicon available. People can argue with me. Zen 5 will come with the N4 silicon which is equal to what the 40 series GPU's are using.

AMD is not trying to gain market share in the GPU sector. Before Intel entered the GPU arena, gamers said no to AMD GPU's. Now people act like AMD and Nvidia GPU's are great and their respective drivers are great. The AMD drivers have been their weak point, they always optimize and fix them. It just takes them a year or two to have solid drivers. By the time they are optimized, a lot of people move on to a newer card.

AMD could have easily raised the bar by making 16GB Vram the baseline in their RDNA3 GPU's. Memory prices are cheap. They could have had 32GB Vram cards at the 7900xtx level. The other big criticism of AMD GPU's is their lack of encoding prowess. They are good gaming cards but that is it.

AMD is following the healthy margin concept vs. increased sales, revenue and market share. All of this is at the expense of the consumer.

Now all three GPU manufacturers are using TSMC silicon. That means it's the silicon stupid. Not all TSMC silicon is created equal. It's not propaganda when Nvidia says they have the best TSMC silicon money can buy. Nvidia wanted the 40 series GPU's on 3nm TSMC silicon but it wasn't ready and Apple bought the entire first run of 3nm.

When Zen 2 was released, AMD was questioned about using 7nm. Instead of the cutting edge 5nm silicon that was available at the same time. They didn't think they needed 5nm and went cheap. This was based on market conditions at the time. Zen 3 is based on 7nm silicon. It wasn't until Zen 4 that AMD moved to the standard TSMC 5nm silicon.

The RX7600 is on 6nm which is really still 7nm. I do not have much hope for AMD GPU's moving forward with RDNA4.

adroc_thurston · Dec 31, 2023

Hans Gruber said:
They use the cheapest and lease efficient 5nm silicon, while Nvidia used the best 5nm silicon available.

It's all the same thing.

Hans Gruber said:
Zen 5 will come with the N4 silicon which is equal to what the 40 series GPU's are using.

It's all the same thing.
Nodelets are sleepworthy things.

Hans Gruber said:
AMD is not trying to gain market share in the GPU sector.

They can't until the chainsaw is ready.

Hans Gruber said:
It just takes them a year or two to have solid drivers

2016 called wants its memes back.

Hans Gruber said:
All of this is at the expense of the consumer.

Yes your job is to pump my share prices now do the needful and pump it.

Hans Gruber said:
When Zen 2 was released, AMD was questioned about using 7nm. Instead of the cutting edge 5nm silicon that was available at the same time.

N5 wasn't available mid-2019 what are you on?

Hans Gruber said:
Not all TSMC silicon is created equal

All bleeding edge stuff is boringly beyond incremental.

Hans Gruber said:
Nvidia wanted the 40 series GPU's on 3nm TSMC silicon but it wasn't ready and Apple bought the entire first run of 3nm.

N3b isn't a usable node at all.

branch_suggestion · Dec 31, 2023

Hans Gruber said:
AMD is really disappointing in the GPU sector. They use the cheapest and lease efficient 5nm silicon, while Nvidia used the best 5nm silicon available. People can argue with me. Zen 5 will come with the N4 silicon which is equal to what the 40 series GPU's are using.

Stop falling for propaganda, both AMD/NV used the best DTCO N5 class process for their needs.

Hans Gruber said:
AMD is not trying to gain market share in the GPU sector. Before Intel entered the GPU arena, gamers said no to AMD GPU's. Now people act like AMD and Nvidia GPU's are great and their respective drivers are great. The AMD drivers have been their weak point, they always optimize and fix them. It just takes them a year or two to have solid drivers. By the time they are optimized, a lot of people move on to a newer card.

No halo effect and chasing feature parity don't help. Feature parity is close, halo will be a while longer.

Hans Gruber said:
AMD could have easily raised the bar by making 16GB Vram the baseline in their RDNA3 GPU's. Memory prices are cheap. They could have had 32GB Vram cards at the 7900xtx level. The other big criticism of AMD GPU's is their lack of encoding prowess. They are good gaming cards but that is it.

If you're lucky the baseline will be 12GB, but could easily be 9...

Hans Gruber said:
AMD is following the healthy margin concept vs. increased sales, revenue and market share. All of this is at the expense of the consumer.

The former can be gained by PPA maxxing, RDNA3 tried and has actually been salvaged by N32's pricing in spite of the thing being broken. RDNA4 hopefully works as intended. The latter requires the biggest stick to trickle sales down.

Hans Gruber said:
Now all three GPU manufacturers are using TSMC silicon. That means it's the silicon stupid. Not all TSMC silicon is created equal. It's not propaganda when Nvidia says they have the best TSMC silicon money can buy. Nvidia wanted the 40 series GPU's on 3nm TSMC silicon but it wasn't ready and Apple bought the entire first run of 3nm.

This is basically all wrong. Especially N3 availability and N3B sucks big time.

Hans Gruber said:
When Zen 2 was released, AMD was questioned about using 7nm. Instead of the cutting edge 5nm silicon that was available at the same time. They didn't think they needed 5nm and went cheap. This was based on market conditions at the time. Zen 3 is based on 7nm silicon. It wasn't until Zen 4 that AMD moved to the standard TSMC 5nm silicon.

N5 also wasn't ready at the time, maybe for mobile risk but that isn't HPC.

Hans Gruber said:
The RX7600 is on 6nm which is really still 7nm. I do not have much hope for AMD GPU's moving forward with RDNA4.

Wafers are expensive. Cheap products get cheap Si, a tale as old as time.
N44/N48 are N4P, take it or leave it. So is Zen5 classic, and you have to take it as there is nothing remotely competitive with it.

TESKATLIPOKA · Jan 1, 2024

Tigerick said:
Just noticed that, same words back to you....It is 2024, I am feeling lucky, I still bet RDNA4 dGPU belongs to RX7000 series, since you are so against my theory, you can bet against me, do you want to?

@SteinFG and others are welcome to bet against me if you guys think RDNA4 dGPU NOT belong to RX7000 series

You didn't only say that RDNA4 belongs to RX 7000 series, you actually said that 7600XT 16GB is based on RDNA4.
You didn't give a good reason why any of this should be true, so you shouldn't be surprised that others don't agree with your opinion.

OEMs want new generation of products or at least rebrand, so they can sell It like It's something new.
Naming a new GPU based on a newer architecture as the old generation just doesn't make sense.
With this logic of yours RDNA3.5 based IGP in Strix Point and Strix Halo would be called as Radeon 7**M.

BTW, you didn't even think about how a cutdown version of this RDNA4 based RX 7600XT would be called. It can't be called as RX 7600, because that's used by N33.
Then let's not forget CES2024 is almost here, where OEMs announce new laptops, so AMD would rather use RDNA4 for mobile, especially considering It's a monolith.

Timorous · Jan 1, 2024

Tigerick said:
You are not the only guess like that, the leaker of Gigabyte 7600XT also speculate so:

View attachment 91255
View attachment 91256

They even makeup image with such configuration

God, is that possible? With total die size of 273mm2 (200mm2 build by N5) plus 16GB RAM, the BOM sure would be much higher than N33.....not to mention the OC version will consume more power

I will let you guys think about that...

I think 12GB is more likely but guess it depends on bom cost difference between 8 ram chips and 2 mcds Vs 6 ram chips and 3 mcds. Probably very little in it. 12GB would mean 192 bit bus which would be far better imo.

As for power, at lower clocks it will be fine.

Tarkin77 · Jan 1, 2024

adroc_thurston said:
They can't until the chainsaw is ready.

is RDNA5 the "chainsaw"?

TESKATLIPOKA · Jan 1, 2024

Tarkin77 said:
is RDNA5 the "chainsaw"?

Yes, until most of It won't be canned like RDNA4, then we will have to wait for RDNA6.

PJVol · Jan 1, 2024

Tarkin77 said:
is RDNA5 the "chainsaw"?

He mean Wunderwaffe

Tigerick · Jan 1, 2024

TESKATLIPOKA said:
You didn't only say that RDNA4 belongs to RX 7000 series, you actually said that 7600XT 16GB is based on RDNA4.
You didn't give a good reason why any of this should be true, so you shouldn't be surprised that others don't agree with your opinion.

You seem forgotten of my questions about RDNA4 that I asked three months ago: - Link with table

I could see why you don't believe at that time; hey the whole Youtube community is hyping RDNA4 as well. Also thanks to adroc giving some misinformation about RDNA4 also won't help. Anyhow, you should be able to see clearer picture of positioning of RDNA4 now... Actually if you throw GDDR7 out of equation, you can calculate the how many CU N43 is having....

In fact, I can estimate rasterization performance from the table I linked. RX7600XT (N44) should perform 5-15% better than RX7600 with upper range to be OC version.

RX7700 (I will stick with N43 for now) should perform 25-35% better than RX7600, the OC version's performance will approach RX7700XT but not RT performance.

TESKATLIPOKA said:
OEMs want new generation of products or at least rebrand, so they can sell It like It's something new.
Naming a new GPU based on a newer architecture as the old generation just doesn't make sense.
With this logic of yours RDNA3.5 based IGP in Strix Point and Strix Halo would be called as Radeon 7**M.

Yeah, as I explained here, mobile RDNA4 will most likely to be model as 8000M series due to different market conditions. OTOH, Unless AMD renamed whole dGPU to RX8000 series, then RDNA4 dGPU will be model RX7600XT and RX7700...cause as I said, RDNA4 is a die shrink of RDNA3, the changes on micro-archicteture is due to lower clock speed of GPU. As clock speed go lower so does TF, AMD has to pump up some changes on front end scheduler to make it slightly powerful to compensate of lower clock speed. So that, RX7600XT will perform better than RX7600 with lower clock speed and same CU...

TESKATLIPOKA said:
BTW, you didn't even think about how a cutdown version of this RDNA4 based RX 7600XT would be called. It can't be called as RX 7600, because that's used by N33.
Then let's not forget CES2024 is almost here, where OEMs announce new laptops, so AMD would rather use RDNA4 for mobile, especially considering It's a monolith.

Well, at least you start thinking of mobile RDNA4 but clearly you miss the reason why RX7600 won't have cut down version on desktop lineup. It is cause of cut down version is available on mobile 7600M series. The same also apply to RX7600XT. I hope you won't ask me why there is no cut down version of RX7700 in the future

Tigerick · Jan 1, 2024

Timorous said:
I think 12GB is more likely but guess it depends on bom cost difference between 8 ram chips and 2 mcds Vs 6 ram chips and 3 mcds. Probably very little in it. 12GB would mean 192 bit bus which would be far better imo.

As for power, at lower clocks it will be fine.

No man, Gigabyte 7600XT is clearly having 16GB GDDR6....please read above thread to have better understanding of RDNA4 then you should know why 7600XT is using N44 GPU. There is no information of die area of N44 but I could estimate to be around 150mm2. With lower clock speed, AMD able to fit in 16GB of GDDR6 at around $300 price point...that's the reason of existence of RDNA4

Timorous · Jan 1, 2024

Tigerick said:
No man, Gigabyte 7600XT is clearly having 16GB GDDR6....please read above thread to have better understanding of RDNA4 then you should know why 7600XT is using N44 GPU. There is no information of die area of N44 but I could estimate to be around 150mm2. With lower clock speed, AMD able to fit in 16GB of GDDR6 at around $300 price point...that's the reason of existence of RDNA4

There are entries for 10GB, 12GB and 16GB 7600XTs. Looks like all bases are being covered.

I sincerely doubt it will be N44. We have seen small products on a new node before as a pipe cleaner but N4 is not exactly new so seems pointless in this instance.

Tigerick · Jan 1, 2024

Timorous said:
There are entries for 10GB, 12GB and 16GB 7600XTs. Looks like all bases are being covered.

I sincerely doubt it will be N44. We have seen small products on a new node before as a pipe cleaner but N4 is not exactly new so seems pointless in this instance.

Yeah, we should have clearer pictures in the coming days including N43

Discussion RDNA4 + CDNA3 Architectures Thread

Golden Member

Platinum Member

Senior member

Golden Member

Senior member

Platinum Member

Senior member

Platinum Member

Golden Member

Senior member

Senior member

Platinum Member

Senior member

Senior member

Platinum Member

Platinum Member

Senior member

Platinum Member

Golden Member

Member

Platinum Member

Senior member

Senior member

Senior member

Golden Member

Senior member