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

 

[marees]

Has anyone backed up MLID's claim that 7900xtx production has beeen stopped a couple of months ago 🤔

 

[marees]

How big is laptop market in comparison to desktop dgpu market & can n48/n44 compete against Blackwell there 🤔

https://twitter.com/x/status/1832988987335512246

 

[adroc_thurston]

How big is laptop market in comparison to desktop dgpu market

Smaller but not that much smaller.

can n48/n44 compete against Blackwell there

No, AMD torched their track record with RDNA3 and has to start from scratch. Again.

 

[Mahboi]

As long as RDNA4 works well on release AMD will be fine.

Fingers crossed. I don't want another gen of
- VR broken (7 months)
- new FSR will be here soon (11 months)
- compute/AI support eventually (ongoing...)

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

I'm not sure how much I trust Daniel Owen's opinions on the market, but it's true that you are getting upscaling enabled by default now. It's getting so baked in I even see idiotic stuff like framegen in competitive FPS.
It's getting to be an obvious choice for most things (antialiasing, upscaling, all kinds of visual upgrades come from software now).
Hopefully whatever AI they bake into FSR 4 does get a better image than what they had.

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.

When's MI350X for again?

 

[Mahboi]

How big is laptop market in comparison to desktop dgpu market & can n48/n44 compete against Blackwell there 🤔

https://twitter.com/x/status/1832988987335512246

The laptop market is huge.
The gaming laptop market is pretty big.
The gaming laptop market that wants APUs with a lot of CUs and little membw is errrr...we'll see when Halo comes out, see how it fares. I'm not expecting a smashing success unless that famed pre-Zen 6 I/O actually gets a ton of throughput.

I don't think they can beat BW for much, but they can seriously bite through their lower level cards. This was already a well expected change.

 

[Shmee]

So is it still looking like RDNA4 is going to be pretty meh? Honestly it seems like RDNA2 and RDNA3 were pretty good, and may continue to be the ones to look for.

 

[marees]

So is it still looking like RDNA4 is going to be pretty meh? Honestly it seems like RDNA2 and RDNA3 were pretty good, and may continue to be the ones to look for.

If the 6800 is still on sale in your region buy that.

Check for winter sales for 6750xt

Check for winter sales of 7900xt

Other than that RDNA 4 should beat RDNA 3 in all respects. So wait for it, if your budget is in $350 to $650 range

 

[ToTTenTranz]

So is it still looking like RDNA4 is going to be pretty meh? Honestly it seems like RDNA2 and RDNA3 were pretty good, and may continue to be the ones to look for.

RDNA4 should be exciting enough if it's bringing back a comeback of a substantial increase to performance/price ratio to the mid/high-end, which is something we haven't really had within the past 7 years.

 

[marees]

RDNA4 should be exciting enough if it's bringing back a comeback of a substantial increase to performance/price ratio to the mid/high-end, which is something we haven't really had within the past 7 years.

Another reason to get excited is currently AMD sales are at an all time low but last year the nov/dec sales were a peak.

So they have to do something to replicate that magic this year also

 

[branch_suggestion]

When's MI350X for again?

2025, Q4 presumably.
Yearly cadence of minor/major changes, you can thank AI bros for that.

So is it still looking like RDNA4 is going to be pretty meh? Honestly it seems like RDNA2 and RDNA3 were pretty good, and may continue to be the ones to look for.

Depends, for the <$500 market if things go to plan, it could be the biggest perf/$ bump since 16nm.
Otherwise there is nothing, yes.

 

[Shmee]

Yeah, I am speaking mainly as someone who already has a 7900XTX. And it is a nice chonky Nitro+ from Sapphire.

 

[ToTTenTranz]

Yeah, I am speaking mainly as someone who already has a 7900XTX. And it is a nice chonky Nitro+ from Sapphire.

Odds are the top Navi 48 might be substantially faster than the 7900XTX in games with heavy raytracing.

 

[marees]

Yeah, I am speaking mainly as someone who already has a 7900XTX. And it is a nice chonky Nitro+ from Sapphire.

Best to skip 2 gen & wait for RDNA 6

 

[linkgoron]

Odds are the top Navi 48 might be substantially faster than the 7900XTX in games with heavy raytracing.

I wouldn't bet on it, just given how small Navi 48 is.

 

[SolidQ]

just given how small Navi 48 is.

4070ti have 294mm, and faster in RT, than XTX

 

[Mahboi]

Odds are the top Navi 48 might be substantially faster than the 7900XTX in games with heavy raytracing.

A 4070 non-super can beat an XTX in raytracing. It's really not news.

2025, Q4 presumably.
Yearly cadence of minor/major changes, you can thank AI bros for that.

Impressive if on time. Wasn't MI300X late last year? If it's a really big leap again, it'll be something.

Depends, for the <$500 market if things go to plan, it could be the biggest perf/$ bump since 16nm.

How so exactly? Top die is expected to be roughly above a 7900 xt that's already for about 700 dollars. Unless you're selling for $450, which seems to be a hard stretch...

 

[linkgoron]

4070ti have 294mm, and faster in RT, than XTX

You're suggesting that Navi 48 at 240mm^2 will compete with the 4070ti at 294mm in raytracing? Do you think that the 240mm^2 will be "substantially faster" in RTX than the 7900xtx (Just the GCD is 304mm^2)?

 

[marees]

Top die is expected to be roughly above a 7900 xt that's already for about 700 dollars.

Let us benchmark prices once 6800 & 6750 xt are sold out

N31 is already sold for $530 as the 7900 GRE

 

[Mahboi]

You're suggesting that Navi 48 at 240mm^2 will compete with the 4070ti at 294mm in raytracing? Do you think that the 240mm^2 will be "substantially faster" in RTX than the 7900xtx (Just the GCD is 304mm^2)?

One of the big improvements is a BVH walker, that should remove one of the worst bottlenecks for RT.
As for general raster performance, here's my copium calculations:

Page 80 - Discussion - RDNA4 + CDNA3 Architectures Thread

Page 80 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

 

[SolidQ]

One funny moment, soon PS5pro reveal, and we gonna know about RDNA4 RT, earlier than from AMD

 

[marees]

One funny moment, soon PS5pro reveal, and we gonna know about RDNA4 RT, earlier than from AMD

As per leaks today is the D-day

 

[Kepler_L2]

CDNA requires all hands on deck as far as chiplets are concerned.

Oh yeah I have seen some employee product assignment data at AMD. MI350/MI400 are the biggest projects within the company, with more than twice the engineers of Zen6.

 

[reaperrr3]

You're suggesting that Navi 48 at 240mm^2 will compete with the 4070ti at 294mm in raytracing? Do you think that the 240mm^2 will be "substantially faster" in RTX than the 7900xtx (Just the GCD is 304mm^2)?

IIRC, the latest rumors are pointing more in the 270mm² direction.

But generally speaking, size isn't necessarily a good indicator. Not all architectural improvements cost much area, sometimes newly designed IP blocks even occupy less area.
Case in point: Despite more L0 and L1 cache, improved RT and doubled maximum FP32 throughput, Navi33 is 25mm² smaller than N23, and normally the N6 process doesn't offer enough density improvement over N7 to accomplish that, so they must've reduced the size of some units while still improving their performance.

For RDNA4, the known info and rumors suggest RT-Performance per CU per clock may more than double. That doesn't mean double the FPS, but the perf hit taken from enabling RT may get cut in half.
Basically, in games where a 7900XTX may drop from 100 to 60 fps if you enable RT, an N48 card may only drop from 100 to 80 (or 90 to 72).

Besides, if the architecture is better optimized for clocks/efficiency than its predecessor, a ~20% higher clockspeed can do a lot for performance per area, and RDNA4 is suggested to clock significantly higher than RDNA3.

 

[Hans Gruber]

The only real benefit AMD has this round is the silicon. Nvidia will be using essentially the same silicon as the 40 series while AMD and Intel will be on what looks to be N4P. Intel will make huge efficiency gains going from N6 down to N4P. I assume AMD will use N4P across their product line. For AMD's sake, I hope they use 20gbit DDR6. Nvidia is going to smoke everybody with DDR7 on their high end products down to the 5070. The claimed efficiency gains from the Hynix DDR7 is crazy good if true. Nvidia will get little to no efficiency gains from their silicon this round.

I would hold off crediting efficiency gains for RDNA4. I think it's all about the silicon and the N4P stuff looks to be really good compared to N5 and N6 used on RDNA3 products.

It's not apples to apples but I would look for 120W for 4060 performance from RDNA4 and if they are at 4070 performance 170-180w would be the power usage for RDNA4 to equal what Nvidia got out of 4N last generation.

 

[Mahboi]

Oh yeah I have seen some employee product assignment data at AMD. MI350/MI400 are the biggest projects within the company, with more than twice the engineers of Zen6.

AIMD king of Compute
I may crap on AI all day but it is a legitimately amazing all-around capable HPC chip too so...yay of big stuff.

IIRC, the latest rumors are pointing more in the 270mm² direction.

But generally speaking, size isn't necessarily a good indicator. Not all architectural improvements cost much area, sometimes newly designed IP blocks even occupy less area.
Case in point: Despite more L0 and L1 cache, improved RT and doubled maximum FP32 throughput, Navi33 is 25mm² smaller than N23, and normally the N6 process doesn't offer enough density improvement over N7 to accomplish that, so they must've reduced the size of some units while still improving their performance.

Yeah RDNA 3 is basically the cheapest arch since forever. Rumours have already come out that its Vega 2.0 and will be (well 3.5 will be) in iGPUs and portables/handhelds until 2027. It's in for the long haul cause it's so cheap area-wise. And since its real weak point is RT, and RT in the first place won't really be usable in those kinds of small devices, it's a perfect fit. It was really born to be a cheap but highly scalable and efficient gen. If only the silly ROP problem, or whatever part it was at the end of raster pipeline that fails, hadn't been there...

For RDNA4, the known info and rumors suggest RT-Performance per CU per clock may more than double. That doesn't mean double the FPS, but the perf hit taken from enabling RT may get cut in half.

Hype train ran over that, what I saw was a broad 25% improvement. Considering that the RT penalty is typically nearly 45-50% with RDNA3, that's more like 15% faster.