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

 

[Kepler_L2]

Full N31 and a cut-down N31? Or no?

Cutdown N31, full N31, full N31 + 3D V-Cache. And none of them are called 7800

 

[leoneazzurro]

Cutdown N31, full N31, full N31 + 3D V-Cache. And none of them are called 7800

7900 7900XT 7950XT 7999,99XT

 

[Timorous]

Cutdown N31, full N31, full N31 + 3D V-Cache. And none of them are called 7800

So N32 is 7800 and probably 7700 for the 192 bit 12GB cut version.

Cool.

As for N31 I guess there will be. 7900XT3D, 7900XT and 7900

 

[GodisanAtheist]

7900 7900XT 7950XT 7999,99XT

"I showed you the 7900 and the 7900XT today... performance leaders in their class to be sure, but there is one more thing... a card so powerful we decided to just launch it as the next next gen today: the 8900XT..."

 

[TESKATLIPOKA]

Cutdown N31, full N31, full N31 + 3D V-Cache. And none of them are called 7800

Ok, so a single N31 is called 7700.
7800 will be 2x N32.
7900 will be 2x N31.
I like this, Ada is finished.

 

[Saylick]

I'm barely holding onto the brake lever here, guys... Please help. This hype train cannot be contained.

https://twitter.com/x/status/1572649543963136000

 

[Revolution 11]

This DLSS3 also makes me think back to the HWUB blacklist issue a while ago. Imagine if media outlets agreed to NVs terms and started having DLSS3 with interpolated frames included in the standard performance graphs. It would so easily skew the real picture and I think this will be telling ad to which outlets are worth reading going forward.

I am far more pessimistic than you are. Most if not all media outlets will eventually get lazy/uninformed/lax about their reporting and start using DLSS3 interpolated frames as raw FPS within a few years. It will be the rare gem in the pile of horse feces that actually continues to report non-DLSS performance in the long-term.

I hope I am very wrong about this.

 

[Karnak]

I'm barely holding onto the brake lever here, guys... Please help. This hype train cannot be contained.

https://twitter.com/x/status/1572649543963136000

That's a lot considering it's N31. Clockwise N32 and N33 should go higher than N31.

 

[Saylick]

That's a lot considering it's N31. Clockwise N32 and N33 should go higher than N31.

I suspect N31 tops out at ~3.5 GHz max boost. N32 probably 100 MHz above that. I suspect N33 won't clock higher, given that it is on N6 even though it is a smaller die, so I'll say ~3.5 GHz max as well. It would be interesting to see what kind of clocks people get with N32, though, since it should be the die with the highest clocking potential.

 

[eek2121]

Pessimists: "There is no way AMD can take on both NVIDIA and Intel. Not possible."

AMD: "HOLD MY BEER!"

🤣🤣🤣

Erm, let's just say that a certain CEO that likes really expensive clothes is going to have virtual mud on his face before the end of the year. 🤣

 

[eek2121]

I'm barely holding onto the brake lever here, guys... Please help. This hype train cannot be contained.

https://twitter.com/x/status/1572649543963136000

The FMax is actually higher. 😉

EDIT: I should add the following disclaimer, the Fmax I have seen is not in final silicon, so it may or may not be higher/lower.

 

[Saylick]

Pessimists: "There is no way AMD can take on both NVIDIA and Intel. Not possible."

AMD: "HOLD MY BEER!"

🤣🤣🤣

Erm, let's just say that a certain CEO that likes really expensive clothes is going to have virtual mud on his face before the end of the year. 🤣

Time to whip out this meme again

 

[Saylick]

Chips and Cheese has some musings on both Ada Lovelace and RDNA 3. Most of it is a recap of what we know, but some is educated guessing. Pretty good insight in my opinion:

Nvidia’s RTX 4090 Launch: A Strong Ray-Tracing Focus

Recently, Nvidia announced their RTX 4000 series at GTC 2022. At a high level, Nvidia took advantage of a process node shrink to evolve their architecture while also scaling it up and clocking it m…

chipsandcheese.com

 

[Joe NYC]

That's indeed the case. Think of it this way: a big part of power usage of chips is transferring and holding data. Transferring the data next to each other within the same die is always the best case. The longer the traces and the more jumps between different materials (e.g. die and substrate) and techniques (e.g. parallel and serial) the more energy will be spent on translation and transportation of the data. That's usually measured in picojoules per bit transferred. And holding data in memory and caches costs energy as well.

I have seen some rumors that RDNA3 may be using CoWoS / RDL.

It would kind of make sense, since the entire complex of GCD + MCDs is not that large, and can fit on a carrier wafer that is not so huge.

I wonder what implications it might have on the power consumption while communicating between MCD and GCD - and if there might be any tricks AMD has up its sleave in the packaging area...

 

[Saylick]

I have seen some rumors that RDNA3 may be using CoWoS / RDL.

It would kind of make sense, since the entire complex of GCD + MCDs is not that large, and can fit on a carrier wafer that is not so huge.

I wonder what implications it might have on the power consumption while communicating between MCD and GCD - and if there might be any tricks AMD has up its sleave in the packaging area...

CoWoS-R doesn't have an silicon interposer, right? It uses a high pitch RDL layer right in the organic substrate itself, so it eliminates much of the cost of traditional 2.5D methods due to the lack of extra silicon, i.e. cheaper than a full interposer and cheaper than using embedded bridges. The bandwidth isn't going to be as high as using silicon due to the lower interconnect density but it's likely enough for the Infinity Cache to the GCD.

 

[beginner99]

Time to whip out this meme again

When your company is slim and all the dead weight that adds nothing but make things more comple for the sake of bureaucracy and justifying their jobs, you can get a lot of stuff done efficiently.

maybe the house cleaning under Rory Read was a good thing and in hindsight he was better than I thought.

talking from experience. When I started at the place I could just do stuff myself. Now I need to go via IT support which then forwards it t specific team which then forwrds it to outsourced service company. all in all it can take a day or to to get a server rebooted. Now imagine this ineffciency in much, much more complex things. many things will simply not be done anymore.

 

[amenx]

For the first time in more than a decade, I'm more interested in an AMD card than Nvidia. Pretty sure they realize this is a golden opportunity for them to steal big market share from Nvidia. All they need is to perform up to par or slightly less than Ada and undercut on price. There is growing revulsion with Nvidia after the 4000 series reveal.

 

[Timorous]

So I was thinking about what we can glean from AMD official information re top N31.

All we have is the greater than 50% perf per watt increase but we don't know the baseline and we don't know the wattage so here are some potential numbers.

Baseline 6900XT (reference)
At 375W we have 1.5x perf / watt x 1.25x more watts which is 1.875x more performance than baseline. A direct match for the 4090 in raster performance and probably a low end estimate given AMD sandbagging on claims recently.

If AMD push to 450W and performance scaling is still decent then we see 1.5x X 1.5x for a 2.25x performance gain over the 6900XT which is in the middle of the rumoured gains.

If perf/watt is closer to 1.6x then at 375W we get 2x performance and at 450W we get 2.4x performance.

6950XT baseline (reference 335W) according to TPU this is 1.07x 6900XT and the 3090Ti is 1.1x ahead of the 6950XT. That puts the 4090 1.76x faster than the 6950XT.

A 375W N31 here with that 1.5x perf/watt would be 1.68x faster than the 6950XT and about 0.9x 4090 performance. 450W would be 2x faster than yhe 6950XT.

1.6x perf/watt would be 1.79x for 375W and 2.15x for 450W.

Also worth noting that in measured power draw the reference 6950XT is actually very frugal and is more efficient than the 6900XT.

Ultimately I think we have reasonable lower to upper bounds here and my estimate is that at worst top N31 will match the 4090 and at best it will be a performance tier ahead depending on final TDP and the exact perf/watt scaling.

Bottom line is I think AMD have levers to pull here and plenty of choices on how to approach the SKU list.

 

[Karnak]

According to Angstronomics and Igor there's a dual 8pin for the reference card, so max. 375W. But given that's the absolute (official) maximum ~350W for the reference is more likely.

 

[Timorous]

According to Angstronomics and Igor there's a dual 8pin for the reference card, so max. 375W. But given that's the absolute (official) maximum ~350W for the reference is more likely.

If this is the case then I think 4090 +/-10% is more likely.

At 350W we have a range of 1.67-1.86x 6900XT performance which seems a little low for the paper spec.

Obviously I am using TBPs rather than measured power numbers and as stated the 6950XT is actually more efficient than the 6900XT when measured despite the 7% perf gain for a 10% TBP increase making it look worse.

These numbers are also a long way short of the rumours so unless AMD heavily sandbagged the perf/watt gain (possible given they said Zen 4 was > 5Ghz and had 8% ipc gain) then it seems a little low to be credible.

 

[moinmoin]

maybe the house cleaning under Rory Read was a good thing and in hindsight he was better than I thought.

I think Rory Read is underappreciated. In hindsight it looks he was hired for a very specific job and he left when that job was done. That job appears to have been cleaning the house and hire a core team of IBM veterans, easing in Su into her current position. Read (23 years at IBM) joined AMD in August 2011, Mark Papermaster (26 years at IBM) joined in October 2011, Dr Lisa Su (12 years at IBM) joined in December 2011.

 

[Tigerick]

AMD Cuts MSRPs of Radeon RX 6000 Series Graphics Cards

AMD has come up with new MSRPs for its Radeon RX 6000 series RDNA2 graphics cards, in a bid to better compete against NVIDIA's RTX 30-series high-end cards, as the product lifecycles of both companies draw to a close. The new price list was finalized on September 15, and is beginning to take...

www.techpowerup.com

Here comes the price cut of RDNA2 to make room for RDNA3 launching, can't wait

 

[Stuka87]

According to Angstronomics and Igor there's a dual 8pin for the reference card, so max. 375W. But given that's the absolute (official) maximum ~350W for the reference is more likely.

That's a long ways from the QUAD 8-pin adapters the AIB 4090's have.

 

[fleshconsumed]

AMD Cuts MSRPs of Radeon RX 6000 Series Graphics Cards

AMD has come up with new MSRPs for its Radeon RX 6000 series RDNA2 graphics cards, in a bid to better compete against NVIDIA's RTX 30-series high-end cards, as the product lifecycles of both companies draw to a close. The new price list was finalized on September 15, and is beginning to take...

www.techpowerup.com

Here comes the price cut of RDNA2 to make room for RDNA3 launching, can't wait

Nice, but it's just a $20-50 price cut for the 6800/6800XT. Doesn't bode well for upcoming 7000 series pricing.

 

[Tigerick]

Nice, but it's just a $20-50 price cut for the 6800/6800XT. Doesn't bode well for upcoming 7000 series pricing.

My estimated selling price of 7800 is between 599 to 799 so still above 6800 series, so should be safe