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

jpiniero · Nov 13, 2024

AMD Confirms Laying Off 4% Of Its Employees To Align Resources With "Largest Growth Opportunities"

Chip designer Advanced Micro Devices, Inc. (AMD) is purportedly laying off 4% of its workforce or 1,000 workers, according to online reports.

wccftech.com

AMD is doing some layoffs. Might this be the beginning of the end of the dGPUs at AMD if they can't get prices up?

marees · Nov 13, 2024

jpiniero said:
AMD Confirms Laying Off 4% Of Its Employees To Align Resources With "Largest Growth Opportunities"

Chip designer Advanced Micro Devices, Inc. (AMD) is purportedly laying off 4% of its workforce or 1,000 workers, according to online reports.

wccftech.com

AMD is doing some layoffs. Might this be the beginning of the end of the dGPUs at AMD if they can't get prices up?

Depends on RDNA 5 goes, I guess

ToTTenTranz · Nov 13, 2024

I feel for the people personally affected, but in general 4% is practically nothing, compared to all the layoffs happening this year in tech. Sounds like it's very few people considering a shift in the company's priorities (AI AI AI).

I'm not sure how this correlates to discrete GPUs at all.

maddie · Nov 13, 2024

ToTTenTranz said:
I'm not sure how this correlates to discrete GPUs at all.

Amazing the blinkers view some have.

Maybe the UDNA decision resulted in staff duplication. Maybe support staff hires have exceeded the chosen optimum. Maybe AI is indeed replacing some low level report collating staff staff. Maybe some of each. So many possibilities.

SolidQ · Nov 13, 2024

For me like, they merge CDNA and RDNA teams, an some excess people got layoffs

Hitman928 · Nov 13, 2024

jpiniero said:
AMD Confirms Laying Off 4% Of Its Employees To Align Resources With "Largest Growth Opportunities"

Chip designer Advanced Micro Devices, Inc. (AMD) is purportedly laying off 4% of its workforce or 1,000 workers, according to online reports.

wccftech.com

AMD is doing some layoffs. Might this be the beginning of the end of the dGPUs at AMD if they can't get prices up?

AMD has done multiple acquisitions in the past couple of years, a certain amount of these layoffs will be due to redundancies after purchasing these companies.

gdansk · Nov 13, 2024

Hitman928 said:
AMD has done multiple acquisitions in the past couple of years, a certain amount of these layoffs will be due to redundancies after purchasing these companies.

Yeah, AMD didn't say where this was targeted. It seems like a bad idea to layoff anyone with experience designing GPUs, since a GPGPU is responsible for most their revenue growth this year. But there are definitely overlaps with some of Xilinx's AI products.

Now the Radeon sales and marketing teams? Probably unnecessary.

DisEnchantment · Nov 13, 2024

Hitman928 said:
AMD has done multiple acquisitions in the past couple of years, a certain amount of these layoffs will be due to redundancies after purchasing these companies.

HR, Sales, Marketing, IT, Accounting, etc., are typical targets for consolidation. Happens 100% of the time during acquisitions and this probably starts on day one of integration process before anybody even start looking at the Engineering side.
~~This year alone AMD acquired like 4 companies.~~
The past two years AMD acquired the following

tajoh111 · Nov 13, 2024

AMD phrased it in a manner however which does imply engineering cuts.

Aligning resources to highest growing opportunities means diverting resources from underperforming markets to ones that are actually growing and making money.

Discrete GPU's will take a cut absolutely. RDNA 4 having 2 chips only along with AMD consolidating CDN AND RDNA into one unit is strong signs of this along with the most important factor in all of this which is AMD's discrete unit makes very little to no money. 15% marketshare along with a low average selling price mean's AMD might be making anywhere from 150-400 million in quarterly revenue which turns into no profit when you mix in R and D amortization. Compare this to AI which is rapidly growing and along with AMD's attempt at accelerating their CDNA products roadmap to keep up with Nvidia's....it should be clear AMD is cutting into gaming when you have read AMD's response about these cuts.

We can't blame AMD as a result of this and you can't really blame consumers that much either. Graphics cards cost more than ever to make(cost of production and R and D) but AMD's discrete sales have shown little to no growth for years. This means profits are getting smaller and smaller. Nvidia has largely been keeping their pricing within 10% of AMD's value proposition which largely means most consumers have little incentive to buy AMD graphics cards when you add on Nvidia's software stack.

Discrete graphics is not growing at a fast enough rate in terms of TAM to justify too many players when the barriers to entry are so expensive. Intel is proof of this. Intel spent 10 billion dollars on ARC with someone leading them with tonnes of GPU experience but it yielded products which were ultimately failures. AMD only makes 100-300 million on discrete graphics but must spend hundreds of millions to billions on R and D on discrete. Do you think it makes sense to continue to spend money on R and D where there are so many other industries AMD is involved in where the profits are far greater?

adroc_thurston · Nov 13, 2024

tajoh111 said:
AMD consolidating CDN AND RDNA into one uni

They are not.

Hail The Brain Slug · Nov 13, 2024

adroc_thurston said:
They are not.

So you're saying Jack Huynh's words are wrong?

Jack Huynh [JH], AMD: So, part of a big change at AMD is today we have a CDNA architecture for our Instinct data center GPUs and RDNA for the consumer stuff. It’s forked. Going forward, we will call it UDNA. There'll be one unified architecture, both Instinct and client [consumer].

moinmoin · Nov 13, 2024

Hail The Brain Slug said:
So you're saying Jack Huynh's words are wrong?

No, the architecture will be re-unified, unlike currently where CDNA stuck with GFX9 while RDNA kept progressing. But that doesn't mean the designs or the teams will be consolidated into one single unit. Both markets will continue to be targeted separately, with dedicated designs.

GodisanAtheist · Nov 13, 2024

leoneazzurro said:
Just like current 4090M is more efficient than 4080M, you trade chip die size for performance. Efficiency of an architecture is depending on many factors. We could speak about efficiency if comparing dies with similar size and resources at the same power budget. That said, Nvidia has the lead currently and my guess is that it will likely keep an advantage with the next gen, then the unknown will be how good or bad RDNA4 will be compared to the green side's next gen.

- If AMD isn't going to be competing on die size, then they're going to compete on frequency and pushing the voltage curve to push their midrange as much as possible.

Doesn't surprise me in the least that AMD when not competing with a whole stack would throw any kind of efficiency metrics out the window for that last 5-10% of performance.

Hail The Brain Slug · Nov 13, 2024

moinmoin said:
No, the architecture will be re-unified, unlike currently where CDNA stuck with GFX9 while RDNA kept progressing. But that doesn't mean the designs or the teams will be consolidated into one single unit. Both markets will continue to be targeted separately, with dedicated designs.

Another quote from the article explicitly mentions this.

So, instead of having two teams do it, you have one team.

inquiss · Nov 13, 2024

Hail The Brain Slug said:
Another quote from the article explicitly mentions this.

Can't that last quote mean about IP generation? You would still have two completely different designs and implementations

adroc_thurston · Nov 13, 2024

Hail The Brain Slug said:
Another quote from the article explicitly mentions this.

yeah, same ISA, rough feature parity, vastly different WGP skews and packaging fanciness allotments.

dr1337 · Nov 13, 2024

moinmoin said:
Both markets will continue to be targeted separately, with dedicated designs.

Two designs by one team...

moinmoin said:
But that doesn't mean the designs or the teams will be consolidated into one single unit.

but that is literally what the person interviewed said would happen...

honestly yall really need to stop debating semantics because the rate of people aggressively agreeing with each other on the AAT forums is too high.

There are probably teams for everything, as there should be. This is multi-million dollar product design, ofc its not one single group producing all products. I highly doubt that even in the GCN days that the same group made workstation products as gaming cards.

An architecture team is not the same as a product design/marketing team.

moinmoin · Nov 14, 2024

dr1337 said:
There are probably teams for everything, as there should be.

There are usually teams for each IP block, as well as teams assembling them into the final package/product. Doesn't mean there can't be overlaps obviously.

Architecture being unified means that work on that IP is no longer isolated between DC and client as it's currently the case. So the team working on the architecture for Instinct won't have to work on GFX9/Vega any more but can use and contribute to the moving target that is the current architecture for the client markets.

I think the primary motivation of this move is to feed back some of the DC optimizations to client to move both targets a little closer, to catch up with Nvidia and their ability to support all their software ecosystem on all their hardware. AMD made that job harder for itself by letting the architecture diverge, and that should improve with this move.

leoneazzurro · Nov 14, 2024

GodisanAtheist said:
- If AMD isn't going to be competing on die size, then they're going to compete on frequency and pushing the voltage curve to push their midrange as much as possible.

Doesn't surprise me in the least that AMD when not competing with a whole stack would throw any kind of efficiency metrics out the window for that last 5-10% of performance.

What I meant is that efficiency of an architecture is a thing, efficiency of a SKU is another, even if there is a dependence of the latter on the former. AMD seems indeed more focused on saving die size and sacrificing some efficiency in the process.

tajoh111 · Nov 15, 2024

Not a reliable source or anything but much of the cuts involved engineers for gaming and semi custom. A 10% cut in in these departments.

The move apparently pissed off people still at the company and hurt moral as the cuts were not merit based.

soresu · Nov 16, 2024

tajoh111 said:
and hurt moral as the cuts were not merit based

Merit by what standard?

Job perforance merit?

Or future company financial strategy merit?

eek2121 · Nov 16, 2024

GodisanAtheist said:
- If AMD isn't going to be competing on die size, then they're going to compete on frequency and pushing the voltage curve to push their midrange as much as possible.

Doesn't surprise me in the least that AMD when not competing with a whole stack would throw any kind of efficiency metrics out the window for that last 5-10% of performance.

Trust me when I say it won’t matter much.

I’d rather have a small-ish 60CU 4ghz chip that costs $599 and uses 20-30% higher power vs a 96CU chip that costs $1,199 that only runs at 2.5ghz. AMD isn’t Intel, I am sure they will be focusing on efficiency.

(honestly I’d rather have something in the middle ground,72-80 CUs at 3.2-3.5ghz maybe?)

The key to keeping costs down is to keep die sizes down. They have been following NVIDIA for years, maybe they should focus on building high frequency cards instead. I am curious at how much things would scale. Current 16 core Zen 5 chips can boost to 5.75ghz. How big of a GPU could AMD make with a 5ghz+ speed I wonder?

gdansk · Nov 16, 2024

If they can even clock it high enough to compete with the 7900 XT I'll be impressed. That'd be nearly 3.4GHz? It seems too high.

Kepler_L2 · Nov 16, 2024

gdansk said:
If they can even clock it high enough to compete with the 7900 XT I'll be impressed. That'd be nearly 3.4GHz? It seems too high.

That was the target (3.46 GHz).

gdansk · Nov 16, 2024

Kepler_L2 said:
That was the target (3.46 GHz).

well it all adds up. I'm curious to see how it worked out this time.

Discussion RDNA4 + CDNA3 Architectures Thread

Golden Member

Lifer

Senior member

Member

Diamond Member

Senior member

Diamond Member

Diamond Member

Golden Member

Senior member

Diamond Member

Diamond Member

Diamond Member

Diamond Member

Diamond Member

Senior member

Diamond Member

Senior member

Diamond Member

Golden Member

Senior member

Diamond Member

Diamond Member

Diamond Member

Senior member

Diamond Member