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

 

[adroc_thurston]

that was original 33%? Because currently seeing 1tb vs 1.8tb

It's 33% more bus width plus the speed bump going from 21 or 20.5 G6X to 28 G7.
GB202 is a really chungus membw bump by modern standards. The perf is kinda MIA though.

 

[SolidQ]

It's 33% more bus width plus the speed bump going from 21 or 20.5 G6X to 28 G7.
GB202 is a really chungus membw bump by modern standards. The perf is kinda MIA though.

Maybe after seeing 5090, AMD will rethink about HALO cards with UDNA2 or UDNA3

 

[Win2012R2]

No, they just modelled a huge uplift based off 50% more SM, new Uarch, GDDR7 and 33% moar memory width.

Same as MLID sauces...

 

[SolidQ]

Same as MLID sauces...

Maybe they have own MLID?

 

[adroc_thurston]

Maybe after seeing 5090, AMD will rethink about HALO cards with UDNA2 or UDNA3

who tf knows but at least people can scream at Jack now rightfully.

Same as MLID sauces...

Or they just modelled like 70% perf bump for NV halo.
Which is good, you always gotta assume the comp is extra mean. That's how they've won server CPUs.
But it's also bad when it kills your halo parts.

 

[Gideon]

It looks like noone is even considering the possibility of "RDNA 4 Performance Leaks are Wrong" meaning it performs worse than the leaks specularte (and worse than RTX 5070.) E.g. a bit below "7900 GRE", losing 5-10% in raster and 20+% in RT

I'm setting the bar to this low in order to be pleasantly surprised by AMD beating it, instead of dissapointed *again*:

Price: 529$ cause AMD
Perf: 5% below 7900 GRE losing to the RTX 5070 (after all it has less resources than a 7900 GRE)

What I think they need to do to at least keep their market share:

Price. 450$ max (same as the 7700 XT MSRP)
Perf: Raster >5% over RTX 5070, RT at least within 5-10%

 

[gdansk]

a bit below "7900 GRE"

At that point why even make it? Just cancel it and keep N44 for budget max.

 

[Gideon]

At that point why even make it? Just cancel it.

Cause the RT perf is better and it has a DLSS competitor?

N32 with its chiplet packaging surely isn't cheap to make. I find it very unlikely N48 is actually more expensive to produce.

But knowing AMD's love story with margins, this of course wont be passed on to the client.

I surely hope I'm wrong, but judging by their past performance I wouldn't be even a bit surprised if they'll charge 699$ for RX 9070 XT if it's even 5% faster in Raster than RTX 5070 non-Ti (cause AMD just loves horrible reviews at launch).

Then they'll be sure to make a RX 9070 which is 5-10% slower (in raster) than the RTX 5070 and sell it for 499$ (gimping the memory to 12GB of course).

 

[adroc_thurston]

N32 with its chiplet packaging surely isn't cheap to make

Not expensive either, InFO slop is cheap by design.

At that point why even make it? Just cancel it and keep N44 for budget max.

Bingo

 

[gdansk]

I find it very unlikely N48 is actually more expensive to produce

Being cheap to produce is useless if you won't sell any and it would be very much on that end.

 

[SolidQ]

"RDNA 4 Performance Leaks are Wrong

There was real two leaks, GRE and 4080, but none of 7900XT, just pure speculation

who tf knows but at least people can scream at Jack now rightfully.

Why Radeon Division almost always unlucky with directors, at least Scott was trying compete with NV

 

[adroc_thurston]

There was real two leaks, GRE and 4080, but none of 7900XT, just pure speculation

The truth is always in-between.

Why Radeon Division almost always unlucky with directors, at least Scott was trying compete with NV

I don't think it's an SVP issue. More product/program manager level.

 

[gdansk]

Confirmed Radeon problems:
1. Marketing (we all agree)
2. Product planning (where is Big Navi 4x?)
3. Execution (Navi 31/32 in particular)
4. Software
5. Leadership *new*

 

[insertcarehere]

At that point why even make it? Just cancel it and keep N44 for budget max.

Wouldn't surprise me if AMD encountered unexpected bugs post silicon tape-out ala what supposedly happened with RDNA3, preventing them from hitting target clocks at a given power envelope. Which would also somewhat explain why a supposedly a mid-range chip is getting board designs at launch that look like a 7900XTX, (3x 8-pin, 3 slot coolers etc).

 

[Win2012R2]

Being cheap to produce is useless if you won't sell any and it would be very much on that end.

You make it sound like the only way to win is not to play...

 

[adroc_thurston]

Wouldn't surprise me if AMD encountered unexpected bugs post silicon tape-out ala what supposedly happened with RDNA3, preventing them from hitting target clocks at a given power node. Which would also somewhat explain why a supposedly a mid-range chip is getting board designs at launch that look like a 7900XTX, (3x 8-pin, 3 slot coolers etc).

They do not repeat stuff.
And 3 slot coolers are just a cope thing because this time AICs do not have any big boy part to tape high margin overkill designs onto.

 

[gdansk]

You make it sound like the only way to win is not to play...

Or execute. They could just do that.

 

[eek2121]

It looks like noone is even considering the possibility of "RDNA 4 Performance Leaks are Wrong" meaning it performs worse than the leaks specularte (and worse than RTX 5070.) E.g. a bit below "7900 GRE", losing 5-10% in raster and 20+% in RT

I'm setting the bar to this low in order to be pleasantly surprised by AMD beating it, instead of dissapointed *again*:

Price: 529$ cause AMD
Perf: 5% below 7900 GRE losing to the RTX 5070 (after all it has less resources than a 7900 GRE)

What I think they need to do to at least keep their market share:

Price. 450$ max (same as the 7700 XT MSRP)
Perf: Raster >5% over RTX 5070, RT at least within 5-10%

It will be under $500.

 

[del42sa]

It will be under $500.

it will be free of charge

 

[beginner99]

Wouldn't surprise me if AMD encountered unexpected bugs post silicon tape-out ala what supposedly happened with RDNA3, preventing them from hitting target clocks at a given power envelope. Which would also somewhat explain why a supposedly a mid-range chip is getting board designs at launch that look like a 7900XTX, (3x 8-pin, 3 slot coolers etc).

I had exactly the same thought. Coolers look way too beefy and why a 3x8 pin option? I think once again didn't hit clock targets and hence power usage will be very high. I guess their target was 5070 Ti for the XT for the naming to make sense. Therefore I also think they absolutely needed to wait for NV official pricing to not look ridiculous with too high a price or too low a margin.

 

[GTracing]

The idea that AMD cancelled the launch because of Blackwell pricing doesn't make sense to me. They're going to have to announce RDNA4 sooner or later, and delaying the announcement doesn't makes it look better.

I think they simply want to use their limited CES time to focus on their strengths. Their CPUs are far more competitive. Non-enthusiasts may only read one article about AMD's presentation. It's better for AMD as a whole if their laptop chips get more press. AMD wants to look good for investors.

Plus, they can sell more RDNA3 stock without a fire-sale.

 

[jpiniero]

The idea that AMD cancelled the launch because of Blackwell pricing doesn't make sense to me. They're going to have to announce RDNA4 sooner or later, and delaying the announcement doesn't makes it look better.

I think they were unsure about Blackwell's actual gaming performance (ie: not using Multi Frame Gen). Don't think they were anticipating Blackwell being so bad. Figure they were going to do the -$100 5070 but the 9070 XT should be easily faster than it in raster.

 

[marees]

Wouldn't surprise me if AMD encountered unexpected bugs post silicon tape-out ala what supposedly happened with RDNA3, preventing them from hitting target clocks at a given power envelope. Which would also somewhat explain why a supposedly a mid-range chip is getting board designs at launch that look like a 7900XTX, (3x 8-pin, 3 slot coolers etc).

Silicon would have taped out 9 months ago

 

[Win2012R2]

Silicon would have taped out 9 months ago

Same as RDNA3 at a similar junction?

 

[marees]

AMD explains the missing RDNA 4 GPUs in its CES 2025 livestream​

News
By Paul Alcorn Contributions from Jarred Walton published 11 hours ago

Solving the case of missing RDNA 4 GPUs.

AMD's CES 2025 livestream was a jam-packed 45-minute show with a slew of product announcements on the company's latest tech. However, one key omission is drawing criticism: AMD didn't mention its next-gen RDNA 4 GPUs at all, even though the company teased the new architecture to the press before the event. After the event, we met with AMD executives David McAfee and Frank Azor in a small group meeting, and the pair immediately addressed the elephant in the room


"There's no way, with everything else we had to do at this show, that you would have left happy if we had included RDNA 4 in the keynote. And not just because of time [limits], but the pricing perspective, performance review, FSR4 capabilities, architecture, and the incumbent [Nvidia] is going to launch something later tonight." Azor said. "Why not wait and see what they're going to do? Whatever we do to respond would just benefit the market. So it's good for gamers."

"This was a no-win scenario," Azor continued. "We debated all these different options and whether or not our customers, the gamers, the market, were going to walk away from today and say, 'Wow, that was amazing, I was blown away by it' if we included it in there for five or eight minutes. Why? Why do that? Let's give it its proper time. Let's put together a recipe that's going to win — positioning, performance, pricing, and the time that it deserves, and show people that we actually do care about gaming."

AMD explains the missing RDNA 4 GPUs in its CES 2025 livestream

Solving the case of missing RDNA 4 GPUs.

www.tomshardware.com