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

 

[Ajay]

that and rdna3 was the most lukewarm serving of kickass in years from amd. unimpressive. it makes more sense to grab a 6900xt.

6950 XTs are actually cheaper. (generally, though supply is running down).

 

[SolidQ]

AMD CEO already teases next-generation Instinct MI400 accelerator series - VideoCardz.com

AMD CEO hints on Instinct MI400 AMD has recently unveiled the Instinct MI300 series, a combination of CDNA3 and Zen4 architectures. Among the series, the MI300X stands out as a GPU-only accelerator boasting an 192GB of HBM3 memory. This particular model is specifically designed to cater to large...

videocardz.com

 

[Paul98]

Really felt like it should have been a 7900 xt and 7800 xt, with pricing similar pricing had they debuted at 999, and 650 or 700 it would have felt much better. Since we are already seeing price drops or lack of demand at these prices next gen maybe we can see some changes to more reasonable pricing all around.

Hopefully they will have a good bit figured out between the first and second gen GCD/MCD hardware and drivers as well. That's what I am most excited to see, as the first gen is usually has issues.

 

[adroc_thurston]

Certainly not with 2 RDNA4 chips to come and also Strix Halo with up to 40CU IGP.

Those are very different CUs.

 

[TESKATLIPOKA]

Those are very different CUs.

You mean RDNA4 ones compared to RDNA3.5 in Strix Halo?
Still, what does It matter with what I wrote? Because my projected table is off? Doesn't really matter If I am off, It was made just out of boredom.
What's important is the performance of those RDNA4 chips. If the weaker RDNA4 chip offers similar or not much better performance than the top APU, I could understand, but the other one should be a lot faster or It wouldn't make sense to me.

 

[adroc_thurston]

You mean RDNA4 ones compared to RDNA3.5 in Strix Halo?

Yes.

Still, what does It matter with what I wrote?

Yes.

Because my projected table is off?

Of course.
Midrange is midrange, $399 parts tops.

but the other one should be a lot faster or It wouldn't make sense to me.

It'll make sense once you understand what the rest of the lineup looked like.
You're thinking in silly relative product categories and not whole lineups.

 

[A///]

6950 XTs are actually cheaper. (generally, though supply is running down).

Really? How's that work out?

 

[TESKATLIPOKA]

Of course.
Midrange is midrange, $399 parts tops.

It'll make sense once you understand what the rest of the lineup looked like.
You're thinking in silly relative product categories and not whole lineups.

Current midrange doesn't cost only $399. For me midrange is full N32 or RTX 4070.

I don't know anything about RDNA4 lineup, true. On the other hand, I just don't believe that AMD will release the next gen product, which doesn't even perform as the middle chip N32.
Maybe If It was based on a single GCD and the higher models were made of more than one, but were canned, then I could understand It.

 

[adroc_thurston]

Current midrange doesn't cost only $399

4060ti is very much a midrange GPU.

For me midrange is full N32 or RTX 4070.

That's not midrange.

I just don't believe that AMD will release the next gen product, which doesn't even perform as the middle chip N32.

Why?
Tiny mainstream dies are tiny.

Maybe If It was based on a single GCD and the higher models were made of more than one

You're thinking in antiques when the vision of the future is right in front of you.

 

[TESKATLIPOKA]

4060ti is very much a midrange GPU.

RTX 4060Ti is Low-midrange GPU at best.

That's not midrange.

Then what is N32 or RTX 4070? Highend? In this gen I would say only RTX 4090 is enthusiastic, So ADa103 and N31 are Highend.
But this is not very important.

Why?
Tiny mainstream dies are tiny.

If AMD will ask $399 max for RDNA4 based GPUs then performance won't be that much better than RTX 4060Ti. So basically only ~1/2 of the performance offered with RDNA3.
Even If the die is tiny so what? AMD won't sell It to me much cheaper, instead they will have higher profits on It.

You're thinking in antiques when the vision of the future is right in front of you.

Vision of the future? You mean RDNA3? Didn't end up very well.

 

[adroc_thurston]

RTX 4060Ti is Low-midrange GPU at best.

No, $399 is very midrange those days.

Then what is N32 or RTX 4070? Highend?

Yea.

So ADa103 and N31 are Highend.

Those are enthusiast.

4090 is enthusiastic

and this is halo.

If AMD will ask $399 max for RDNA4 based GPUs then performance won't be that much better than RTX 4060Ti

how do you know? have you seen the uArch?

AMD won't sell It to me much cheaper, instead they will have higher profits on It.

Aren't you the smartest guy in the whole wide world.
AMD's no charity.

Vision of the future?

Yes.

You mean RDNA3?

It has elements yes, but the real vision is a bit elsewhere.

 

[SolidQ]

won't be that much better than RTX 4060Ti.

If 7700XT should be like 6800, why 8700XT will be same?

 

[TESKATLIPOKA]

how do you know? have you seen the uArch?

No, I didn't see anything, I have no contacts in the industry. I am just a tech HW fan.
You said It should cost $399 max, and you also said AMD is no charity, so I can only deduce that It won't be that much better performance wise than RTX 4060Ti.

 

[Timorous]

Assuming the design is MCM $400 tops suggests a GCD area of no more than 150mm² so probably a 2SE 40 CU design.

If it manages to clock at 3.5Ghz then without an IPC gain it will be about 30% faster than a 6700XT which is not a million miles away from the 6800XT.

With an IPC gain that might stretch to 6950XT top end as a guess.

That would also suggest a 4 MCD design or perhaps 3 MCDs with 3gb GDDR7. Expect AMD will do the cost estimation and go with the cheaper option.

A 40 CU design would be a great building block for an 80 CU 2 GCD config and a 3 GCD 120CU config so if they no longer exist it suggests to me that AMD could not get chiplet graphics working in time to be properly competitive so ditched it. Perhaps with the intention of bringing rdna5 a bit sooner should yields and supply be there.

So I would expect a 8700XT with 40CUs somewhere around 7800XT to 7900GRE performance and a 8600XT somewhere around 7700XT / 6800 performance.

Obviously this is super rough guessing so let's see where things land.

 

[adroc_thurston]

A 40 CU design would be a great building block for an 80 CU 2 GCD config and a 3 GCD 120CU config so if they no longer exist it suggests to me that AMD could not get chiplet graphics working in time to be properly competitive so ditched it

That's now what they were planning to do.
Just look at MI300.

 

[maddie]

Assuming the design is MCM $400 tops suggests a GCD area of no more than 150mm² so probably a 2SE 40 CU design.

If it manages to clock at 3.5Ghz then without an IPC gain it will be about 30% faster than a 6700XT which is not a million miles away from the 6800XT.

With an IPC gain that might stretch to 6950XT top end as a guess.

That would also suggest a 4 MCD design or perhaps 3 MCDs with 3gb GDDR7. Expect AMD will do the cost estimation and go with the cheaper option.

A 40 CU design would be a great building block for an 80 CU 2 GCD config and a 3 GCD 120CU config so if they no longer exist it suggests to me that AMD could not get chiplet graphics working in time to be properly competitive so ditched it. Perhaps with the intention of bringing rdna5 a bit sooner should yields and supply be there.

So I would expect a 8700XT with 40CUs somewhere around 7800XT to 7900GRE performance and a 8600XT somewhere around 7700XT / 6800 performance.

Obviously this is super rough guessing so let's see where things land.

I was also thinking of a ~150mm² die. Will it be on N3E? Then maybe 50-60CU (1.6X logic vs 5nm) if using SoIC connections to a lower base cache/IO 6nm die of equal area. Hot die on top.

What's the capacity of SoIC packaging in Q3 2024?

 

[adroc_thurston]

What's the capacity of SoIC packaging in Q3 2024?

Midrange won't be SoIC for a whole lot longer.

 

[Timorous]

That's now what they were planning to do.
Just look at MI300.

Is there a typo there? Now = not?

 

[maddie]

Midrange won't be SoIC for a whole lot longer.

If we agree that PPA (really PP$ cost) is a top priority, then logic on more expensive node and rest on 6nm since cache/IO scaling dies after 6, barring some new designs for cache, etc.

If not SoIC then ????? Can't see how unified die = lowest PP$

 

[adroc_thurston]

Is there a typo there? Now = not?

Yea, sorry.

If we agree that PPA (really PP$ cost) is a top priority, then logic on more expensive node and rest on 6nm since cache/IO scaling dies after 6, barring some new designs for cache, etc.

SoIC-X is the only viable option for GPUs and it's too slow and pricey of a flow for mainstream parts.

 

[TESKATLIPOKA]

Assuming the design is MCM $400 tops suggests a GCD area of no more than 150mm² so probably a 2SE 40 CU design.

If it manages to clock at 3.5Ghz then without an IPC gain it will be about 30% faster than a 6700XT which is not a million miles away from the 6800XT.

With an IPC gain that might stretch to 6950XT top end as a guess.

That would also suggest a 4 MCD design or perhaps 3 MCDs with 3gb GDDR7. Expect AMD will do the cost estimation and go with the cheaper option.

A 40 CU design would be a great building block for an 80 CU 2 GCD config and a 3 GCD 120CU config so if they no longer exist it suggests to me that AMD could not get chiplet graphics working in time to be properly competitive so ditched it. Perhaps with the intention of bringing rdna5 a bit sooner should yields and supply be there.

So I would expect a 8700XT with 40CUs somewhere around 7800XT to 7900GRE performance and a 8600XT somewhere around 7700XT / 6800 performance.

Obviously this is super rough guessing so let's see where things land.

RX 6800 is 30% faster than RX6700XT at 4K. TPU
RX 6800XT is 50% faster than RX6700XT at 4K. TPU
A 40CU GPU at 3.5GHz without any IPC gain would be around RX 6800 level of performance.
It would need +15% IPC to be on par with 60CU 7800XT or ~4.2GHz clocks.

There is no need for GDDR7 If we are talking only about BW.
192-bit 21.5gbps would provide 516GB/s, which is a tiny bit more than 6800XT.
192-bit 24gbps would provide 576GB/s, which is the same as RX 6950XT.
The problem would be that 3 MCDs would provide only 48MB IC and 12GB Vram.

 

[Timorous]

RX 6800 is 30% faster than RX6700XT at 4K. TPU
RX 6800XT is 50% faster than RX6700XT at 4K. TPU
A 40CU GPU at 3.5GHz without any IPC gain would be around RX 6800 level of performance.
It would need +15% IPC to be on par with 60CU 7800XT or ~4.2GHz clocks.

There is no need for GDDR7 If we are talking only about BW.
192-bit 21.5gbps would provide 516GB/s, which is a tiny bit more than 6800XT.
192-bit 24gbps would provide 576GB/s, which is the same as RX 6950XT.
The problem would be that 3 MCDs would provide only 48MB IC and 12GB Vram.

I said

If it manages to clock at 3.5Ghz then without an IPC gain it will be about 30% faster than a 6700XT which is not a million miles away from the 6800XT.

If you want more than roundabout numbers which seems utterly daft given how little we know then I will go with +40% because the 6700XT sustains around 2.5Ghz so 3.5Ghz is a 40% increase. In that case it lands between the 6800 and 6800XT and describing that kind of delta as not a million miles away from the 6800XT seems fair to me so why be so nit picky on super estimated numbers giving broad stroke guesses?

If that is the spec then a final performance of 6800XT region to 6950XT region would be well within expectation imo.

As for GDDR6 Vs 7 it will depend on cost. 256bit + 8 2GB chips for 16GB VRAM and decent bandwidth Vs 192bit + 6 3GB chips (if they even get manufactured ofc) for 18GB vram is much of a muchness in terms of performance. There is also no guarantee the MCDs will be the same, could be cost saving to re-use the existing ones but might be better long term to do another iteration which could include more cache or something. Far too many unknowns and both ways look viable to me.

 

[TESKATLIPOKA]

I said
If you want more than roundabout numbers which seems utterly daft given how little we know then I will go with +40% because the 6700XT sustains around 2.5Ghz so 3.5Ghz is a 40% increase. In that case it lands between the 6800 and 6800XT and describing that kind of delta as not a million miles away from the 6800XT seems fair to me so why be so nit picky on super estimated numbers giving broad stroke guesses?

If that is the spec then a final performance of 6800XT region to 6950XT region would be well within expectation imo.

I did read what you wrote.
I just provided a bit more detailed numbers needed to be on par with 6800xt(7800XT). That's all.
It was never about you being correct or not considering we don't know anything about RDNA4.

But I miscalculated in sustained clocks of RX 6700XT.
So yes, you are right that at 3.5GHz It should be in the middle between 6800 and 6800XT.
It would need ~+7% IPC to be on par with 6800XT or ~3.75GHz clocks.

As for GDDR6 Vs 7 it will depend on cost. 256bit + 8 2GB chips for 16GB VRAM and decent bandwidth Vs 192bit + 6 3GB chips (if they even get manufactured ofc) for 18GB vram is much of a muchness in terms of performance. There is also no guarantee the MCDs will be the same, could be cost saving to re-use the existing ones but might be better long term to do another iteration which could include more cache or something. Far too many unknowns and both ways look viable to me.

The question is If GDDR7 will be already available at the time of RDNA4's release, but probably yes.
If GDDR7 will be available, then l can imagine a new MCD with GDDR7 support and more Infinity cache -> 24MB.
3*MCD chips would then provide 72MB IC and up to 768GB/s(32gbps) + 18GB Vram.
If there won't be 24gb chips, then I wouldn't be surprised If AMD used 4 old MCD to have 64MB IC and 16GB Vram and up to 768GB/s(24gbps).

 

[Heartbreaker]

OUCH. GG if true.

https://twitter.com/x/status/1687388515670056960

Judging by the predating @uzzi38 reaction this looks to be true:

https://twitter.com/x/status/1687379360276185088

I can understand the RDNA1 generation as AMD was totally broke for most of the time it was under development, but ...

Frankly, I don't get why they did it. Are the economics really just not there? Or rather is it a sign that chiplet-based GPUs aren't really working out yet?

Maybe it's a nice change to ridiculously under hype this time.

Looking back at the early RDNA 3 rumors it was going to be a massive 2.7X performance jump, and we can see how accurate that was...

I expect these early RDNA 4 rumors to be just as ridiculously wrong...

 

[Timorous]

Maybe it's a nice change to ridiculously under hype this time.

Looking back at the early RDNA 3 rumors it was going to be a massive 2.7X performance jump, and we can see how accurate that was...

I expect these early RDNA 4 rumors to be just as ridiculously wrong...

For compute the Tflops is what 2.5x or something so it technically met that target or close enough.