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

In2Photos · Sunday at 6:41 PM

HuesToo4 said:
To be honest AMD is not at a point yet where regular non-tech savvy people are specifically looking for Ryzen stickers, partly because Intel still dominates OEMs.

That's my point. They finally have their foot in the door. You have to start somewhere. I've had several IT departments recommend AMD systems in the last year or so for new workstations. That's never happened before.

adroc_thurston · Sunday at 6:47 PM

In2Photos said:
People wanted the Intel inside sticker too. 😏

no they didn't, Intel is whatever, and PCs are a commodity outside of Apple lands.
GPUs are toys and toys command utter, slavish loyalty to the brand.

In2Photos · Sunday at 9:26 PM

adroc_thurston said:
no they didn't, Intel is whatever, and PCs are a commodity outside of Apple lands.
GPUs are toys and toys command utter, slavish loyalty to the brand.

I literally sold PCs to people in the mid 90s that came in looking for those stickers but OK.

adroc_thurston · Sunday at 9:54 PM

In2Photos said:
I literally sold PCs to people in the mid 90s that came in looking for those stickers but OK.

Sorry bro today is not the mid 90s.
PCs have been commoditized to hell.

marees · 2025-03-31T01:22:30-0400

Win2012R2 said:
To win market share AMD must make enough cards to satisfy demand - especially for the launch period, but that's clearly has not happened, if 200k units sold is true (seems plausible given known stock figures at a bunch of retailers) then they clearly failed to make anywhere near enough to grab even tiny market share, despite their claims that they want to get to 40-50%, they won't if they got no stuff to sell.

In general I agree with your points.

The RDNA 4 hardware was ready last june

They also know that Nvidia releases the xx60 cards pretty late

So they could have planned to sell more cards if launched early

Only hitch is that they have to get the AIBs to buy the hype. AIBs usually give an estimate of how many cards they need. If AMD produces more silicon there is a risk of inventory just sitting in the warehouse (& the shareholders/stakeholders may not be in favor of that)

Ultimately it boils down to: does AMD consider itself as a serious player in the desktop gaming market ...

Vikv1918 · 2025-03-31T01:29:50-0400

Idk its very hard to think of anything other than some kind of market politics being the reason. I dont think consumer preference has anything to do with the current state of Radeon laptops having 0.05% marketshare. At that marketshare the consumer is irrelevant because they dont even have a choice in the first place. Do we have a choice to buy a $700 7600m laptop? Or a $1000 7800m laptop? Or a $1400 7900m laptop? No we don't, so we can't prove "people dont buy AMD".

Furthermore, its not an "all or nothing" scenario either. Its totally viable from a financial perspective for both AMD and OEMs for Radeon to have 20% laptop marketshare just like their desktop GPUs. Somehow AMD was selling laptops pre-Ryzen and it didn't have a 0.01% marketshare like they do in dGPU.

OEMs simply dont offer such products (and don't tell me its not technically or economically feasible, nvidia themselves sell $750 4060 laptops). And its become even more glaring with the 9000 series which would make the absolute best laptop chips at their price and efficiency.

IMO the gaming laptop situation needs to be researched by the tech or investigative journalists. I have never seen anything like this total monopoly before.

beginner99 · 2025-03-31T02:04:55-0400

HuesToo4 said:
To be honest AMD is not at a point yet where regular non-tech savvy people are specifically looking for Ryzen stickers, partly because Intel still dominates OEMs.

true and in a corporate setting you don't have a choice. if you are lucky you can choose from 2-3 different models (from the same OEM) but thats about were it ends. had to get a new one last year. no amd option and the intel options where limited to slower U variants only.

In2Photos · 2025-03-31T06:28:17-0400

adroc_thurston said:
Sorry bro today is not the mid 90s.
PCs have been commoditized to hell.

Yeah cuz, isn't it amazing how things can change? Funny how that works over time.

eek2121 · 2025-03-31T08:05:53-0400

marees said:
In general I agree with your points.

The RDNA 4 hardware was ready last june

I don’t think you get how any of this works.

I want you to look into the effort that goes into designing, developing, manufacturing, and shipping such a complex part along with the associated software stack.

AMD could have launched slightly earlier, but not much. End of December at the earliest. People don’t buy stuff in that time period because Christmas just ended.

They were also still working on drivers.

Kronos1996 · 2025-03-31T08:12:11-0400

It takes 3 months just to make a finished wafer IIRC. So probably 6 months for a finished GPU to hit shelves. If they started manufacturing in June, that lines up pretty well with when they launched.

GTracing · 2025-03-31T11:30:53-0400

I decided to try and put a number to how much better RDNA4 is at getting gaming performance out of a given amount of memory bandwidth. I'm using TPU's 1440p raster numbers.

The issue is it's possible for a GPU to have more memory bandwidth than it needs. The 9070 XT has 42% more raster performance than the 7800 XT with only 4% more memory bandwidth. However, I don't think the 7800 XT is not being constrained by memory bandwidth, so that wouldn't be a fair comparison. To try and solve this problem, I'm going to use the 7900 GRE. It seems to be the most bandwidth constrained RDNA3 GPU. Both the 7900 GRE and 9070 XT each have much higher flops than the 7800 XT and 9070 respectively, but that doesn't translate to FPS.

The 7900 GRE has ~23% higher flops than the 7800 XT, but only 10% more gaming performance.

	mem bandwidth	flops	gaming perf vs GRE
7800 XT	624.1 GB/s	37 TFLOPS	90%
7900 GRE	576.0 GB/s	46 TFLOPS	100%

The 9070 XT has 35% higher flops than the 9070, but only 13% more gaming performance.

	mem bandwidth	flops	gaming perf vs GRE
9070	644.6 GB/s	36 TFLOPS	115%
9070 XT	644.6 GB/s	49 TFTFLOPS	128%

The 9070 XT has 28% more gaming performance than the 7900 GRE, but only 12% more memory bandwidth. To normalize that number, the 9070 XT has 14% higher FPS per GB/s.

caveats:

memory bandwidth constraints are not a hard dropoff. It's possible that one of the 7900 GRE or 9070 XT is more bandwidth constrained than the other.
This comparison does not take into account ray tracing performance. From TPU's numbers, the 9070 XT has 53% more RT performance, or 37% higher RT FPS per GB/s.

jpiniero · 2025-03-31T12:34:48-0400

eek2121 said:
I don’t think you get how any of this works.

I want you to look into the effort that goes into designing, developing, manufacturing, and shipping such a complex part along with the associated software stack.

AMD could have launched slightly earlier, but not much. End of December at the earliest. People don’t buy stuff in that time period because Christmas just ended.

They were also still working on drivers.

Yeah. If it was easy to meet demand properly, you wouldn't have the Fire Sales that has happened in the past.

reaperrr3 · 2025-03-31T12:40:37-0400

GTracing said:
(....)

Some more caveats to consider:

- 9070 and 9070 XT have identical Frontend (CP, SEs/Primitive Pipes) and Backend (same number of ROPs).

Scaling of performance has traditionally NEVER been linear with more SM/CUs, regardless of mem bandwidth.
Rule of thumb is that additional perf/clk per SM/CU is rarely more than 60-70% of the additional SM/CUs even in best-case scenarios.

- Computerbase tested average ref card clocks of ~2640mhz for the 9070 and ~2890mhz for the XT under load, meaning the "true" TFLOPs advantage of the XT is "only" ~26%.

marees · 2025-03-31T12:56:14-0400

eek2121 said:
I don’t think you get how any of this works.

I want you to look into the effort that goes into designing, developing, manufacturing, and shipping such a complex part along with the associated software stack.

AMD could have launched slightly earlier, but not much. End of December at the earliest. People don’t buy stuff in that time period because Christmas just ended.

They were also still working on drivers.

Kronos1996 said:
It takes 3 months just to make a finished wafer IIRC. So probably 6 months for a finished GPU to hit shelves. If they started manufacturing in June, that lines up pretty well with when they launched.

jpiniero said:
Yeah. If it was easy to meet demand properly, you wouldn't have the Fire Sales that has happened in the past.

Didn't MLID claim AMD could launch RDNA 4 in october ??

MrTeal · 2025-03-31T13:58:12-0400

marees said:
Didn't MLID claim AMD could launch RDNA 4 in october ??

That's how you know it was impossible.

gaav87 · 2025-03-31T15:15:44-0400

What is even the point of navi48 with 32GB when 9070xt can't even AI in windows ?
You need zluda hack and slash with rdna3.5 apu dll's to get 6800xt perf. What, a joke.
Rocm on windows still dead linux is what 3-4 versions ahead ?

carrotmania · 2025-03-31T15:45:34-0400

marees said:
Didn't MLID claim AMD could launch RDNA 4 in october ??

You went from sounding like you knew what you talking about, even though, not really, to scrambling for a MLID quote? That's some thick skin you have, to carry on showing up in this thread. I applaud your persistence.

To go over your first post though: just because the hardware is done, we have no idea where not just the drivers were up to, but also FSR4. Maybe they got wind of DLSS4 Transformer early enough to start an update of their own. And it paid off. AMD of the past would have been raked hard for their software no matter when they released. But there's been little to no complaints about AMD drivers recently, and not just in comparison to the s#%show that NV is going through. And FSR4 is a flippin triumph. Hanging with NVs new model for the most part, definitely being better than D3.8.

Other than fanb... ooops... whinging about timescales an developments they are not privy to, this has been one of AMDs very best releases... ever.

Sure, we wish there had been a higher card, and more of the ones they did release, but that doesn't give you license to a) demand and 2) make up your own bunk timescales.

marees · 2025-03-31T16:22:54-0400

carrotmania said:
scrambling for a MLID quote?

MLID was one (final) data point

It started with this

AMD Talks RDNA 4, GPU-Based AI Accelerators, Next-Gen Graphics Pipeline: Promises To Evolve To RDNA 4 With Even Higher Performance In Near Future

AMD's Radeon head, David Wang, has talked about RDNA 4 GPUs, AI Accelerators within GPUs and next-gen graphics pipeline in an interview.

wccftech.com

marees · 2025-03-31T16:27:57-0400

The hype started here

https://twitter.com/x/status/1695020416736141670

blackangus · 2025-03-31T17:07:58-0400

marees said:
MLID was one (final) data point

It started with this

AMD Talks RDNA 4, GPU-Based AI Accelerators, Next-Gen Graphics Pipeline: Promises To Evolve To RDNA 4 With Even Higher Performance In Near Future

AMD's Radeon head, David Wang, has talked about RDNA 4 GPUs, AI Accelerators within GPUs and next-gen graphics pipeline in an interview.

wccftech.com

View attachment 121025

Im pretty sure that slide is years old, I believe I saw that in the RDNA 2 launch.
If that was recent it would not have said "Advanced Node" for the delivery as that was in stone quite a while ago.

GTracing · 2025-03-31T17:14:46-0400

reaperrr3 said:
Some more caveats to consider:

- 9070 and 9070 XT have identical Frontend (CP, SEs/Primitive Pipes) and Backend (same number of ROPs).

Scaling of performance has traditionally NEVER been linear with more SM/CUs, regardless of mem bandwidth.
Rule of thumb is that additional perf/clk per SM/CU is rarely more than 60-70% of the additional SM/CUs even in best-case scenarios.

- Computerbase tested average ref card clocks of ~2640mhz for the 9070 and ~2890mhz for the XT under load, meaning the "true" TFLOPs advantage of the XT is "only" ~26%.

Fair points. Though it's easy to find exceptions to that 60-70% rule.

7700 XT is a cut down 7800 XT. It has the same number of ROPs, 90% as many TMUs, and 90% as many CUs. And yet the 7800 XT is 15% faster. The gaming performance actually scales more than linearly with CUs between the two GPUs.

It seems to me the 7700 XT must be held back by memory and cache.

DavidC1 · 2025-03-31T19:11:41-0400

reaperrr3 said:
Scaling of performance has traditionally NEVER been linear with more SM/CUs, regardless of mem bandwidth.
Rule of thumb is that additional perf/clk per SM/CU is rarely more than 60-70% of the additional SM/CUs even in best-case scenarios.

Something that puts out performance, a GPU, CPU, or even a car for that matter can't have 100% scaling from scaling up a portion of the unit - it just doesn't make sense.

Even if you are extremely bound by something, let's say memory banwidth, you'll still see small amount of gains by keeping memory same and beefing up everything else.

A device that scales nearly 100% from scaling up just one part, memory, SM, texture unit, is an imbalanced architecture and is thus a subpar design.

memory bandwidth constraints are not a hard dropoff. It's possible that one of the 7900 GRE or 9070 XT is more bandwidth constrained than the other.

It cannot, and will never be a hard drop off. In order to get 100% gains, everything needs to be boosted 100%, because otherwise the part that isn't boosted 100% would end up being a bottleneck.

In a simplistic manner, the split in performance is roughly 1/3rd between each of the following:
-Shader throughput
-Fillrate
-Memory Bandwidth

If you want 100% faster GPU, it needs to have 2x the Flops, 2x the fillrate, and 2x the memory BW, regardless of whether it's "bottlenecked" in memory, or flops, or fillrate, because inevitably there are games where it needs more of some than the other.

The entire struggle that GPU designers have is getting power and thermals controlled, getting up to date with the latest APIs, while having the right balance of all the resources(Flops, Fillrate, BW), in a multi-billion transistor microchip.

eek2121 · 2025-03-31T19:34:05-0400

marees said:
Didn't MLID claim AMD could launch RDNA 4 in october ??

MLID knows nothing. He isn't even a technical guy. I've been in this industry since the 80s and I tell you that he is an embarrassment to YouTube.

Picture someone with a gun and a target. They have infinite ammo and claim to have a perfect bullseye score. They post lots of videos of them shooting at the target, then delete the ones where they miss the bullseye. That is MLID. It has been proven and their are archives of his deleted videos. Also, in his case he claims to be able to hit every part of the target as well, so he covers everyone between those looking for bullseyes and those wanting to see a bullet hitting the outer ring of the target. His videos are made to ensure everyone is wrong except him. He's been called out by some of the top folks in the tech community, and he doesn't care, because people still watch him, and he still makes money. There is a group of folks out there archiving videos that have been deleted from him and others like him. They plan to do something with the videos, but they haven't said what. If they follow through, it should be interesting regardless.

General rule of thumb. If a YouTuber or social media person is selling you a rumor via ads or some other nonsense, and offers nothing else of value, ignore the rumor. If their whole MO and/or identity is based upon that, they will feel the need to constantly sell you on something all the time. WCCFTech is the same way. I don't look at their site anymore, but they used to be terrible and write/delete articles based upon what is wrong or right.

Gideon · 2025-04-01T03:31:55-0400

eek2121 said:
MLID knows nothing. He isn't even a technical guy. I've been in this industry since the 80s and I tell you that he is an embarrassment to YouTube.

...
General rule of thumb. If a YouTuber or social media person is selling you a rumor via ads or some other nonsense, and offers nothing else of value, ignore the rumor. If their whole MO and/or identity is based upon that, they will feel the need to constantly sell you on something all the time. WCCFTech is the same way. I don't look at their site anymore, but they used to be terrible and write/delete articles based upon what is wrong or right.

Yeah, the telltale is the regular release of long videos just to keep the engagement up. That means that there constantly needs to be content. Some has turned out to be actual leaks (like these zen5 slides) - biut those are usually easily identifiable by having actual consistent technical data MLID himself is never able to come up with:

But most is just fluff for engagement.

I don't blame people watching him. But please consider the content in those videos as your daily Horoscope -they are pretty similar in substance. Any actual info bits in them will be on Videocardz, reddit or this very forum in no time. No need to watch 10+ minute fluff pieces.

Aapje · 2025-04-01T03:58:56-0400

GTracing said:
7700 XT is a cut down 7800 XT. It has the same number of ROPs, 90% as many TMUs, and 90% as many CUs. And yet the 7800 XT is 15% faster. The gaming performance actually scales more than linearly with CUs between the two GPUs.

It seems to me the 7700 XT must be held back by memory and cache.

An issue is that the full chip is designed to be a balanced package, and it's not necessarily easy to cut it down in a way that makes financial sense and is still balanced.

For example, judging by the mere 10% cut of the TMU/CUs, the yield on those parts of the chip is very good, and lowering the cut-off even further probably results in very little extra chips that can be sold. But cache seems more prone to error, and the bus size cannot really be cut to such a small extent. So the 7700 XT is probably unbalanced.

Discussion RDNA4 + CDNA3 Architectures Thread

Golden Member

Platinum Member

Diamond Member

Platinum Member

Diamond Member

Senior member

Junior Member

Diamond Member

Platinum Member

Diamond Member

Member

Senior member

Lifer

Member

Senior member

Diamond Member

Senior member

Member

Senior member

Senior member

Senior member

Senior member

Golden Member

Diamond Member

Golden Member

Golden Member