Question Speculation: RDNA3 + CDNA2 Architectures Thread

[uzzi38]

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 is much bigger than from RDNA1 to RDNA2. We should expect many big improvements in GFX11. 🤔" / Twitter

 

[TESKATLIPOKA]

Pretty sure they're talking about the mobile SKUs they leaked earlier. So the clocks (up to 2700MHz boost) and perf estimates are for N33 @120W max (RX 7700M).
Desktop (RX 7600XT, not releasing anytime soon) will probably spring for 20Gbps GDDR6, higher clocks and obviously, higher TBP (150W+).

You are probably right. He was also comparing PHX IGP to RX 570.

 

[amenx]

Wonder what sort of inventory will be available at launch. Have feeling it will be hard to get an XTX at or near MSRP. And that it will sell out quickly and ebay will be full of them at $1500 ea. Would be a shame, but this may very well end up happening.

 

[Aapje]

Whatever there is will be sold out quick. Just prepare for the usual, including impatient people whining all over the place.

 

[gdansk]

Whatever there is will be sold out quick. Just prepare for the usual, including impatient people whining all over the place.

Related to that, when does AMD webstore add stock for new GPU launches? 8AM PST? Does anyone know

 

[scineram]

Launch is pretty late so there'll be only expensive air shipments.

 

[Kaluan]

You are probably right. He was also comparing PHX IGP to RX 570.

Yeah the public is much more aware of how desktop GPUs perform rather than IGPs and mobile GPUs (with wildly varying TDP and cooling configurations). Less confusion.

So I appreciate it when I they use stuff like dGPUs as performance baseline for IGP and mGPU.

The old rumor about PHX's IGP = 3060M @ 60W TGP was a mess.

 

[Glo.]

In order to tie with 3060M it would need to score around 8000 pts in 3dMark TS.

I think its much more realistic that this kind of performance we will see from Strix Point, at the earliest.

 

[Kaluan]

In order to tie with 3060M it would need to score around 8000 pts in 3dMark TS.

I think its much more realistic that this kind of performance we will see from Strix Point, at the earliest.

Well my point was that mobile GPU performance is all over the place. So unreliable baselines to compare against.

The 8000 TS graphics score you mention is @ ~95W TGP (or an average of 30 laptops with TGPs from 65W to 140W. At 60W (lowest spec) it gets under 6200.
nVidia's Dynamic Boost (+15W) further complicates things. Was it on or off during that run? etc

I don't think you can even find 3060M designs that run at 60W. Lowest I found was 65W and it also had +15W Dynamic Boost feature.

680M @ 45W does 2440 TS graphics
And RX 570 does ~4000
But I think Time Spy doesn't like GCN so much, so not ideal for these 3-way comparisons, I usually use Fire Strike for lower end or older gen comparisons. Bit more neutral.

 

[TESKATLIPOKA]

8000 in TimeSpy Graphic?
That is pretty much impossible with Strix Point.
28CU RX 6700S manages 8300, so you can see how big GPU you would need.
Then you would need enough power to feed It, 45W is clearly not enough, at least BW could be fixed by HBM.

I see Strix Point with 16CU IGP max, so ~5500 points.

 

[TESKATLIPOKA]

Early AIB RX 7900 series.

AIB RX 7900 XTX up to: Boost AIB 2677 MHz.
Game AIB 2542 MHz.

AIB RX 7900 XT up to: Boost AIB 2611 MHz.
Game AIB 2262 MHz.

AIB clock inconsistent with reference. Some OC AIB card TBP >= 400w.

Link
I think this was expected.

 

[Glo.]

8000 in TimeSpy Graphic?
That is pretty much impossible with Strix Point.
28CU RX 6700S manages 8300, so you can see how big GPU you would need.
Then you would need enough power to feed It, 45W is clearly not enough, at least BW could be fixed by HBM.

I see Strix Point with 16CU IGP max, so ~5500 points.

Well, if 35W 6WGP/12 CU with 2.6 GHz iGPU of Phoenix will score around 4000 pts in graphics time spy, and Strix point will clock much, much higher, will have 33% more WGPs/CUs, and most importantly - Infinity Cache, I think we should start considering increasing the perf targets, a little .

Personally, I expect desktop 5600XT/RTX 2060 performance, from highest TDP variant, highest end variant of Strix Point.

And yeah, this will be most likely my next upgrade in the form of a NUC PC .

 

[Kaluan]

Link
I think this was expected.

They're also confirming the N31 hardware bug. And that new tap-outs that fix it will be H2 2023, along with N32. 😒

 

[Kaluan]

Well, if 35W 6WGP/12 CU with 2.6 GHz iGPU of Phoenix will score around 4000 pts in graphics time spy, and Strix point will clock much, much higher, will have 33% more WGPs/CUs, and most importantly - Infinity Cache, I think we should start considering increasing the perf targets, a little .

Personally, I expect desktop 5600XT/RTX 2060 performance, from highest TDP variant, highest end variant of Strix Point.

And yeah, this will be most likely my next upgrade in the form of a NUC PC .

That begs the question for me if Phoenix implementation of RDNA3 (Graphics Core v11.0.1) also suffers from such v/f crippling "bugs". And if N33 (GC v11.0.3) and "PHX2" (v11.0.4, allegedly, initially thought to be Dragon Range, but we know that uses RDNA2), coming out H2 2023 also suffer from it.

Strix uses "RDNA3+", Graphics Core v11.5.0, so a different type of beast. Presumably on a TSMC N3-class node.

 

[TESKATLIPOKA]

They're also confirming the N31 hardware bug. And that new tap-outs that fix it will be H2 2023, along with N32. 😒

Where did you get info about N32 being H2 2023?

 

[Saylick]

Where did you get info about N32 being H2 2023?

https://twitter.com/x/status/1599053260643479555

 

[amenx]

That would be one huge lengthy gap for AMDs upper midrange under $899 products. Hope its not true.

 

[TESKATLIPOKA]

https://twitter.com/x/status/1599053260643479555

The "power bug" should have been fixed for N32 and N33, so why would It be delayed to H2? Or that's just desktop parts.
I take this with suspicion.

 

[scineram]

I think it all depends on how long Navi21 stocks last. And what Nvidia does.

 

[Kaluan]

I think it all depends on how long Navi21 stocks last. And what Nvidia does.

Yes. That and if the so called bug is present on N33 silicon as well.
Tho mobile N33 seems to have the green light for Q1 2023. Desktop... is a different subject.

(mobile N33 is corroborated by the leaked 2023 ASUS G16 laptop specs)

 

[maddie]

Yes. That and if the so called bug is present on N33 silicon as well.
Tho mobile N33 seems to have the green light for Q1 2023. Desktop... is a different subject.

(mobile N33 is corroborated by the leaked 2023 ASUS G16 laptop specs)

Why assume the bug, if it exists, is in the logic blocks. My thinking is the metal signal layers, maybe the fan-out design. Signal integrity @ design V&F and such. Much easier (relative) to fix.

 

[biostud]

https://twitter.com/x/status/1599053260643479555

If these rumors are true, then a 3Ghz N32 should be close to a 7900XT in speed, but price wise it should be quite a bit lower, so it would make sense to launch N32 with a faster N31, when first gen N31 is being fased out. Then raise the price for N31 gen2.

 

[scineram]

I think we could see Navi32 earlier than H2. I am expecting Q2.

 

[Kaluan]

Link
I think this was expected.

The AIB cards they're talking about are these:
(7-12% higher clocks)

XFX Speedster MERC319 RX 7900 XTX ULTRA Specs

AMD Navi 31, 2677 MHz, 6144 Cores, 384 TMUs, 192 ROPs, 24576 MB GDDR6, 2500 MHz, 384 bit

www.techpowerup.com

(9-12% higher clocks)

XFX Speedster MERC310 RX 7900 XT Black Specs

AMD Navi 31, 2559 MHz, 5376 Cores, 336 TMUs, 192 ROPs, 20480 MB GDDR6, 2500 MHz, 320 bit

www.techpowerup.com

XFX 7900XTX Merc319 Speedster ULTRA
and
XFX 7900XT Merc319 Speedster ULTRA

Now, historically (looking back at XFX's RX 6900XT lineup) the ULTRA SKUs are far from "up to" in terms of max factory OC on their custon designs. Yet these 7900XT/XTX still seems to have a healthy boost over stock.
Here's their 6900XT ULTRA:
(4-5% higher clocks)

XFX Speedster MERC319 RX 6900 XT ULTRA Specs

AMD Navi 21, 2340 MHz, 5120 Cores, 320 TMUs, 128 ROPs, 16384 MB GDDR6, 2000 MHz, 256 bit

www.techpowerup.com

And here's their fastest 6900XT SKU:
(12-21% higher clock)

XFX Speedster ZERO RX 6900 XT WB Specs

AMD Navi 21, 2525 MHz, 5120 Cores, 320 TMUs, 128 ROPs, 16384 MB GDDR6, 2000 MHz, 256 bit

www.techpowerup.com

Check the stock/factory OC clock deltas between all of these. Just thought it was worth pointing out.

 

[Mopetar]

If these rumors are true, then a 3Ghz N32 should be close to a 7900XT in speed, but price wise it should be quite a bit lower, so it would make sense to launch N32 with a faster N31, when first gen N31 is being fased out. Then raise the price for N31 gen2.

That's what I had wondered about. If there was some design defect that could be fixed for N32, it would have meant the performance would be close to N31 in cases where the memory bandwidth wasn't a bottleneck.

I think we could see Navi32 earlier than H2. I am expecting Q2.

Not much sense in releasing something that would basically supplant an existing product if there's no replacement for it and would just result in no one buying that higher tier product.

Right now a 7900 is at least a better value than a 4080 if you're just looking for raster performance. Better to sell those and get rid of the stock while you can if they're going to be replaced in half a year or so.

The interesting bit will be how AMD prices everything. I suspect they'd like to launch N32 at prices just below where N31 is at now and have the N31 refresh sell for even more. Hopefully the market (driven by the glut of used mining GPUs and excess Ampere stock) makes that strategy less viable, but we can hope.

 

[linkgoron]

That's what I had wondered about. If there was some design defect that could be fixed for N32, it would have meant the performance would be close to N31 in cases where the memory bandwidth wasn't a bottleneck.

Not much sense in releasing something that would basically supplant an existing product if there's no replacement for it and would just result in no one buying that higher tier product.

If they can make a profit with the lower performance, they'll sell the card with some hardware block to cap performance or some kind of driver solution. There's no reason to store GPUs in some storage and wait for 2 or 3 quarters if they can make a profit from them, IMO. Supposing that there really is a hardware bug, it makes more sense, IMO, to sell whatever they have and then just refresh next year.