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

 

[Mahboi]

It is truly amusing that you think AMD is in dire straits.

Reminder that at this stage:
RDNA 3 was borked due to a dated ROP/final raster pipeline stage that made the cards need 700W to reach the perf it should've reached with 350W
RDNA 3.5 was researched and produced (it's in STX Point)No respin on RDNA 3, despite terrible sales and the R&D already paid for the fix
RDNA 4 is rumoredly a tiny gen whose biggest die is 240mm². So a 6600 xt's size. The excuse for it? "Chiplets didn't return as expected so they """"have"""" to abandon all mid and high range SKUs"

Meanwhile, Nvidia is just fine with crapping out giant, costly dies and selling them for a crazy price. I get that they ship a lot more volume than AMD so amortization is much better, but how do you suppose that AMD's ever going to get on their sales level if all they do is tiny, least cost investments? Just wait and hope that NV screws up?

The RDNA 3.5 situation just baffles me, they did the R&D, the product works great...and it's in iGPUs. Could have an 7950 XTX with 15% better perf out. Or you could have just ignored RDNA 3 and moved on to an RDNA 4 replacement.
No replacement. RDNA 4's biggest will get around a 7900 xt's raster perf and 4070 Ti's RT perf, it doesn't even equal an XTX's raster. Baby GPU instead of the borked but mighty XTX.
No fixes. Just for tiny iGPUs.
Two more years of "wait for RDNA 5 saving the world".What is your game @ amdradeon? It's getting difficult to understand at this point. It just feels like the bare minimum except for CDNA/MI300.

At least having something better than half NV's performance in Blender would be great. It's not like RDNA 3 was bad at compute, it crushes the competition in most other software.

These interviews are realpolitik, nothing more.

The "real" part of this is what I just outlined above.
AMD is not making any headroom or advancement with their current penny pinching strategy.

 

[Glo.]

It appears that Hyuhn could be suggesting move towards APUs and prioritizing such products over highest end Graphics.

 

[marees]

It appears that Hyuhn could be suggesting move towards APUs and prioritizing such products over highest end Graphics.

Atleast with RDNA 4 it seems AMD has focussed on laptop parts such as the 7900 GRE

Still if 7900xtx production is already stopped 2 months ago, then that leaves a large gap to fill until Q1 2026 at the least

Plus the 6800 is now out of stock everywhere & the sales spike during nov/dec in winter

So what is Radeon's game plan now 🤔 🤔🤔

 

[linkgoron]

AMD deprioritizing flagship gaming GPUs: Jack Huynh talks new strategy against Nvidia in gaming market

The battle seems to be over before it starts.

www.tomshardware.com

A couple of years later you will get a repeat of this same headline.

Only then the word 'flagship' will be missing and 'Intel' will get added to it.

Amusing interview. He's talking as if having a flagship means that you can't also build products that have a large TAM. AMD's architectures have failed to deliver, so they need to say something about TAM or whatever. RDNA3 and probably RDNA4 have been a failure. They also told the same stories with Polaris:

In an interview with Ars, AMD's Roy Taylor also confirmed that Polaris would target mainstream users, particularly those interested in creating a VR-ready system.

"The reason Polaris is a big deal, is because I believe we will be able to grow that TAM [total addressable market] significantly," said Taylor. "I don't think Nvidia is going to do anything to increase the TAM, because according to everything we've seen around Pascal, it's a high-end part. I don't know what the price is gonna be, but let's say it's as low as £500/$600 and as high as £800/$1000. That price range is not going to expand the TAM for VR. We're going on the record right now to say Polaris will expand the TAM. Full stop."

We all know how that worked out. Nvidia just released the GTX 1060, which was the best selling card of its generation.

It is truly amusing that you think AMD is in dire straits. These interviews are realpolitik, nothing more.

AMD failed to execute on RDNA3 and scrapped all of the high end of RDNA4. Their best performing card will compete with Nvidia's 5070 lineup. You tell me what the situation is.

NV could price AMD out of the market completely right now, but they know leaving openings for scraps will keep the FTC off their case.

Why would they? Nvidia has been going for larger margins and more profit, not for killing off AMD.

What happened is simple, RDNA3 was ~20% weaker than anticipated, AMD lost the momentum they had built with RDNA2, and now they have to start the ball rolling all over again.

RDNA3 was bad. If RDNA4 is such a success where's the high end RDNA4?

 

[Mahboi]

Amusing interview. He's talking as if having a flagship means that you can't also build products that have a large TAM. AMD's architectures have failed to deliver, so they need to say something about TAM or whatever.

Yeah that was the same ridiculous cope that MLID came out with: "AMD is focusing on midrange cause see...midrange makes sense! See! A lot of people buy midrange" as if somehow making a 600mm² die put all the other machines at TSMC to such shame that they'd all die rather than make a 240mm² die.

RDNA3 and probably RDNA4 have been a failure. They also told the same stories with Polaris:
We all know how that worked out. Nvidia just released the GTX 1060, which was the best selling card of its generation.

According to Kepler it doesn't seem like RDNA 4 is a big failure or anything, but it was meant to be on a node with somewhat poor yield. Again AMD gives up when it's costly and NV doesn't care and just pays.

AMD failed to execute on RDNA3 and scrapped all of the high end of RDNA4. Their best performing card will compete with Nvidia's 5070 lineup. You tell me what the situation is.

Very bad. It's 2 years too, there's absolutely no indication that we'll gain anything except maybe 3 months. No competition anywhere on top until 2026. Most likely end of it.

It's ridiculous, we don't even have a 24Go VRAM card this time. The biggest is 16, and while that's just fine for gaming, for any kind of demanding productivity...just don't demand.
I do have some hopium for decent CDNA cards or RDNA Pro that'll push that productivity perf, but almost nobody buys those as they're mad expensive vs the consumer product.

RDNA3 was bad. If RDNA4 is such a success where's the high end RDNA4?

Yet another case of "let's do chiplets to save costs" and the costs aren't saved enough, so just cancel and put out some castrated gen.
Is this going to be a cycle from now on?

Castrated	Success	Epic Misstep
RDNA 1	RDNA 2	RDNA 3
RDNA 4	RDNA 5	RDNA 6

 

[Glo.]

Atleast with RDNA 4 it seems AMD has focussed on laptop parts such as the 7900 GRE

Still if 7900xtx production is already stopped 2 months ago, then that leaves a large gap to fill until Q1 2026 at the least

Plus the 6800 is now out of stock everywhere & the sales spike during nov/dec in winter

So what is Radeon's game plan now 🤔 🤔🤔

The paint is on the wall. dGPUs will exist. But AMD (and Intel) do have other, more important, much more lucrative segments to develop tech for.

For consumers - we should expect steady ramping up of designs of powerful APUs. Both from Intel and AMD.

We all know how that worked out. Nvidia just released the GTX 1060, which was the best selling card of its generation.

RX 6600 cost 40$ less than RTX 3050, used the same amount of power, had equally bad RT performance, had the same 8 GB VRAM, was 30% faster in rasterization, was just below RTX 3060 in raster performance and 3050 still outsold RX 6600 3:1.

Thats why its impossible to justify for AMD developing, and burning potentially hundreds of millions of dollars for development of highest end GPUs if they will not sell them. Its that simple.

You won't win if consumers in this space do not care about AMD products, regardless of how good they are against their direct competition.

 

[marees]

Yet another case of "let's do chiplets to save costs" and the costs aren't saved enough, so just cancel and put out some castrated gen.
Is this going to be a cycle from now on?

Castrated	Success	Epic Misstep
RDNA 1	RDNA 2	RDNA 3
RDNA 4	RDNA 5	RDNA 6

My take

RDNA 1 - castrated (+ partial PS5)
RDNA 2 - success (+ xbox x + partial PS5)
RDNA 3 - unmitigated failure
RDNA 3.5 — all future IGP (low power lessons learnt from Samsung mobile)

RDNA 4 - castrated (plus originally meant for laptops like 7900 GRE. Also RT of PS5 pro)
RDNA 5 - full stack including software crossfire? (Hopefully xbox next)
RDNA 6 - ??? (PS6 ??, Xbox hardware stopped in favor of xbox o/s)

Also AMD need an answer to CUDA that just works on all gaming GPUs & on windows

 

[SolidQ]

3050 still outsold RX 6600 3:1

Even now NVidia doesn't have any good 150-500$ cards, people still buy trash cards from NV in those range. That problem in users head.

AMD need an answer to CUDA

Isn't that rocm target with RDNA5?

 

[Mahboi]

AMD deprioritizing flagship gaming GPUs: Jack Huynh talks new strategy against Nvidia in gaming market

The battle seems to be over before it starts.

www.tomshardware.com

"So, I ask, is it fun to go King of the Hill?"
YES.
YES IT'S FUN. IT'S WHAT PEOPLE TALK ABOUT FOR LITERALLY YEARS.
YOU ABSOLUTE *REDACTED*

 

[Mahboi]

RX 6600 cost 40$ less than RTX 3050, used the same amount of power, had equally bad RT performance, had the same 8 GB VRAM, was 30% faster in rasterization, was just below RTX 3060 in raster performance and 3050 still outsold RX 6600 3:1.

That's a very different story. People who buy these kinds of weak GPUs typically don't care much to look for HW. Please remember that NV is the default, AMD is the outlier.
People buy NV because NV = GPU. They buy AMD because they bother to look into GPUs enough to actually try to find a good one for themselves.
If for example the 6700 xt sold 3 times less than the 3060 Ti, then I'd actually flip the damn table over it. But really low tier GPUs aren't market representative.

Thats why its impossible to justify for AMD developing, and burning potentially hundreds of millions of dollars for development of highest end GPUs if they will not sell them. Its that simple.

Couldn't disagree more, although it's true that the gaming market is far from impressive right now, it still commands a solid amount of public image. On the contrary developing for it and getting ahead of NV would put a lot of eyes on AMD.

You won't win if consumers in this space do not care about AMD products, regardless of how good they are against their direct competition.

They're not going to care if you don't develop so that's kind of a pointless statement. You have to launch something strong for people to care.
Mostly I still feel like 90% of the problem is software though.

Look at this. The guy focuses solely on video editing, photoshop and Blender. It's basically what you assume AMD to do worse than Nvidia.

The XTX actually trounces the 4090. Yes the 4090, not the 4080 Super.
It wins handily in Photoshop.

Wins at everything except the GPU effects and RAW in Premiere Pro.

Crushes it in Davinci Resolve except in 8K.

Then you get to Blender.


That's the software difference. I don't see how a 3D gaming card somehow yields 1/3rd of the performance in a 3D modeling software when it has proven that it has the chops to do everything else better than the competition.

Also, another thing: when RDNA 3 came out, this guy actually bothered to try the 7900 XT and dropped it after a week. Reason? Crashes and random reboots. I never got those with my XT, maybe a freak reboot once in 2 years, and all my crashes recovered gracefully (except when my PSU wasn't strong enough but that was obviously the PSU). But I only bought the 7900 XT 6 months after it came out.
You just need the software to be more all around trustworthy. Blender needs to work full throttle, the cards can't afford to crash for 3-4 months after they come out and drivers can't be fully mature when the goddang generation is on the way out.

 

[SolidQ]

I don't know people need from AMD. Bad RT? or they want 4090 lvl RT on 800$ GPU?
Where those people which praised 3090 RT?

That's the software difference

AMD said, become software company, we gonna see how it turns out.

Then you get to Blender

I'm now learning making game models in Blender. I don't see any difference between AMD or NV, because doesn't need render there.
Even saw video lesson, where he's working with AMD IGPU, and working fine.

 

[Glo.]

That's a very different story. People who buy these kinds of weak GPUs typically don't care much to look for HW. Please remember that NV is the default, AMD is the outlier.
People buy NV because NV = GPU. They buy AMD because they bother to look into GPUs enough to actually try to find a good one for themselves.
If for example the 6700 xt sold 3 times less than the 3060 Ti, then I'd actually flip the damn table over it. But really low tier GPUs aren't market representative.

Couldn't disagree more, although it's true that the gaming market is far from impressive right now, it still commands a solid amount of public image. On the contrary developing for it and getting ahead of NV would put a lot of eyes on AMD.

They're not going to care if you don't develop so that's kind of a pointless statement. You have to launch something strong for people to care.
Mostly I still feel like 90% of the problem is software though.

Look at this. The guy focuses solely on video editing, photoshop and Blender. It's basically what you assume AMD to do worse than Nvidia.

The XTX actually trounces the 4090. Yes the 4090, not the 4080 Super.
It wins handily in Photoshop.
View attachment 107078
Wins at everything except the GPU effects and RAW in Premiere Pro.
View attachment 107079
Crushes it in Davinci Resolve except in 8K.
View attachment 107077
Then you get to Blender.
View attachment 107076

That's the software difference. I don't see how a 3D gaming card somehow yields 1/3rd of the performance in a 3D modeling software when it has proven that it has the chops to do everything else better than the competition.

Also, another thing: when RDNA 3 came out, this guy actually bothered to try the 7900 XT and dropped it after a week. Reason? Crashes and random reboots. I never got those with my XT, maybe a freak reboot once in 2 years, and all my crashes recovered gracefully (except when my PSU wasn't strong enough but that was obviously the PSU). But I only bought the 7900 XT 6 months after it came out.
You just need the software to be more all around trustworthy. Blender needs to work full throttle, the cards can't afford to crash for 3-4 months after they come out and drivers can't be fully mature when the goddang generation is on the way out.

You disagree with me, and yet you have literal proof that its actually exactly what I have written. In benchmarks you have a GPU that wins with Nvidia, yet nobody, compared to Nvidia GPUs buys it.

Let me ask you again. What is the incentive for AMD to develop Highest end GPUs if they WILL NOT sell compared to Nvidia, despite how powerful they are?

Because market thinks simply: GPU = Nvidia. And nothing else.

 

[marees]

Because market thinks simply: GPU = Nvidia. And nothing else.

There is a bit of nuance here

As recently as a year ago the $500 7800xt was flying of the shelves

Even today GPUs in $350 to $750 range are selling well.

Which is why I am hopeful that navi 48 xtx would launch soon.

But still puzzling that there is no revision of 7900xt / 7900xtx in sight

 

[adroc_thurston]

What is the incentive for AMD to develop Highest end GPUs if they WILL NOT sell compared to Nvidia, despite how powerful they are?

Marketing.
But MCP people are enslaved to yearly GPGPU cadence so no halo client goodies. For now, anyway.

 

[Mahboi]

Let me ask you again. What is the incentive for AMD to develop Highest end GPUs if they WILL NOT sell compared to Nvidia, despite how powerful they are?

Already answered and you just rejected the answer because it doesn't fit in the massive copium narrative that people have for AMD.

 

[Glo.]

Already answered and you just rejected the answer because it doesn't fit in the massive copium narrative that people have for AMD.

And YOU are not rejecting the other side of the perspective, because it doesn't fit your narrative?

 

[dr1337]

Thats why its impossible to justify for AMD developing, and burning potentially hundreds of millions of dollars for development of highest end GPUs if they will not sell them. Its that simple.

I wish we actually had any real sales numbers. Even if AMD only has 10% of the flagship market that could still be hundreds of millions of dollars in sales, and that's not counting the fact they rebrand some consumer dies into prosumer.

I do not deny that mindshare matters, but its not like there isn't an opportunity for companies to compete with Nvidia at various tiers of the market. Even when the 3050 outsells a 6600, AMD is still shipping millions of dollars of units.

Why give up on money if customers want to give it to you? If a unit is profitable then its profitable, QED.

 

[Mahboi]

I wish we actually had any real sales numbers. Even if AMD only has 10% of the flagship market that could still be hundreds of millions of dollars in sales, and that's not counting the fact they rebrand some consumer dies into prosumer.

Not really the problem. As always R&D is a net loss and the margins don't pay just for the cards but also for the R&D.
That's why Nvidia can afford to waste so much more money: if NV gets an effective 30% net margin per card, and sells 10M, while AMD sells just 2M, even if they have 50% per card, Nvidia just gets a lot more money, assuming that AMD is spending about as much in R&D (they're not).

The problem with this kind of vicious circle is that if you can't climb in market share, you can't put enough resources to better yourself, so you fall off more and more. AMD has no fundamental problem there since they have a hefty CPU income (and AI obv.) but it's still a general problem.

I do not deny that mindshare matters, but its not like there isn't an opportunity for companies to compete with Nvidia at various tiers of the market. Even when the 3050 outsells a 6600, AMD is still shipping millions of dollars of units.

It doesn't work that way either.
Sales fundamentally come from three things:
- access
- response to client needs
- price

AMD is a global company selling small electronics, access isn't a physical problem. But most people know about Nvidia first. When the non electronics inclined people go into an electronics store, Nvidia will be front and center since that's what sells. AMD will be "that other company" in case NV somehow doesn't work out for you.

Response to client needs is why we have segmentation. It's not about selling a good SKU or a few good SKUs, it's about getting the product people want.
This is where AMD really suffers actually, because NV managed to stick them with that rep of "bad drivers, bad support, bad quality", which AMD doesn't put enough effort into beating back.
When RDNA 3 come out, VR wasn't properly functional (stuttering) for 7 months.
People doing video editing work complained about crashes.
I even remember LTT complaining about the cards being not worth buying for 2 months or so (leave them in the oven longer).

That inability to get a day one flawless product is what's keeping this reputation going easy for Nvidia. Most of these problems are solved within the first 2 months (that VR bit was real bad though), but it's during these 2 months that most reviewers would try the cards.
As long as people believe that AMD will bring them crashes, trouble, or won't be able to do something or another that they might want to do, they'll go to Nvidia. Price barely even matters there. If you told me "you can possibly get this done on Radeon, possibly not, but on Geforce always, however it's 30% more expensive"...I'd buy Nvidia. I just am interested enough to know the truth of it, but most normies will simply want to pay, get the GPU, go home and play/model/video edit.

This is why mindshare matters so much. As long as AMD is seen as N°2, it's an uphill battle to sell vs NV, the "natural" seller. If AMD could get their software to about the same level as NV, already the rep of "poor man's GPU" or "bad drivers" would disappear over some time. If they can get their hardware to really smash NV's face in once, they'd go from being the outsider to being the new top dog.

I think that if AMD comes out with a monster GPU but has crappy SW support, it won't amount to that much. Lots of people will try, be disappointed, and that's going to be another bust. If they come out with strong software, then they'll be judged on price, not on client needs. Either works in the end, but I think SW is their real door to success. And it costs so much $$$$...

 

[Mahboi]

Everyone's reacting like that lol, but you still see the copium of "buT tHiS Is tHE rIgHT StRaTEGy"...
"Don't take chances and keep running away when things get difficult" - Sun Tzu

 

[DiogoDX]

Amd does this every time they release a bad uarch. Maybe this guys are too young to remember R600 or even VEGA.


Hoping for RDNA5 to be good like RDNA2 but I think Intel will surpass them to be the 2nd vendor in 2 or maybe 3 generations.

 

[adroc_thurston]

but I think Intel will surpass them to be the 2nd vendor in 2 or maybe 3 generations.

Intel has no money to keep client dGFX going.
Literally a broke company.

 

[Mopetar]

As long as RDNA4 works well on release AMD will be fine. Compared to Vega, this is hardly concerning. Let them fix their issues and get back on stride so they can deliver a high-end card worth owning or that will make NV customers consider buying AMD.

 

[marees]

As long as RDNA4 works well on release AMD will be fine. Compared to Vega, this is hardly concerning. Let them fix their issues and get back on stride so they can deliver a high-end card worth owning or that will make NV customers consider buying AMD.

Going by Daniel Owen's video, it seems a flood of 50 class & 60 class consumers are waiting to upgrade. If AMD can grab a significant portion of them then they would do well

View attachment 107096
Everyone's reacting like that lol, but you still see the copium of "buT tHiS Is tHE rIgHT StRaTEGy"...
"Don't take chances and keep running away when things get difficult" - Sun Tzu

 

[SolidQ]

Let them fix their issues

Their issue is upscaling. No competition to DLSS/XESS, because every youtuber screaming about it, DLSS/XESS is better, so people choose NV.
That one bigger reason why people choose NV, not RT, RT just bonus to marketing.
Now games out with default upscaling options

Just looking that video(timestamp)
He's saying same things, AMD just need true competitor to DLSS

 

[branch_suggestion]

CDNA requires all hands on deck as far as chiplets are concerned.
That is why high end RDNA4 was canned, RDNA5 will have the thing as was hinted at but it might be a while to wait.