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

 

[Glo.]

So GDDR7 based on your post not mentioning a type.

HBM500 CONFIRMED!

 

[TESKATLIPOKA]

Back to what Glo. hinted about memory bus.
I am skeptical about only 128bit, I am more in line with 192bit bus and 12GB Vram. If It was only 128bit then N33 would have either 8GB or 16GB. One is too little and the other too much for this new mainstream.

 

[Ajay]

Rdna3 top Gpu will be around 20-30% faster than Nvidias Lovelace in gaming performance (raster) and both gpus will consume about 450W. You can air-cool 450W so these gpus will still have mainstream appeal.

Air cool 450W, sure, if you live north of +45 degrees latitude and are running >300 CFM through your case; everything will be fine - LOL!
Oh, and 4 slot HSF for the win. Otherwise GPU clocks would drop into the gutter.

 

[TESKATLIPOKA]

N33 is monolithic.

And I personally expect for this GPU RX 6900 XT+25% performance.

At least, based on rumors. Will it be true? We will see. If we thought that RDNA2 rumors were crazy, RDNA3 rumors blow them out of the water.

With a 3GHz clockspeed and 80CU why not.
The question is why would N33 need to be so much faster than N21. That would mean N23->N33 would be ~125% increase in a single generation.
This chip would be far larger than 240mm2 and much more expensive than N23.

 

[GodisanAtheist]

N33 is monolithic.

And I personally expect for this GPU RX 6900 XT+25% performance.

At least, based on rumors. Will it be true? We will see. If we thought that RDNA2 rumors were crazy, RDNA3 rumors blow them out of the water.

- Are monolithic & "chiplet/multi-die" mutually exclusive?

Just because AMD separated Logic and IO for Zen chiplets doesn't have to mean they'll take the same approach for GPU chiplets. Its possible having I/O redundancy among the "chiplets" is required for GPU chiplets to work at all.

Just spitballin' here...

 

[Glo.]

To be honest, the more I myself think about this, the more sceptical I am about N33 being 128 bit.

Its just ridiculous thought, in grand scheme of things.

Its not that it is impossible, its ENTIRELY possible. But... Hard to grasp.

- Are monolithic & "chiplet/multi-die" mutually exclusive?

Just because AMD separated Logic and IO for Zen chiplets doesn't have to mean they'll take the same approach for GPU chiplets. Its possible having I/O redundancy among the "chiplets" is required for GPU chiplets to work at all.

Just spitballin' here...

In this context, yes. Chiplet and monolithic GPUs are made on two different nodes.

 

[Graphenewhen]

What if N33 is the largest single GPU chiplet, and all "N3x's" above it are multi-die "chips"?

So you're looking at something really crazy like:

7900XTX = 3x N33 dies, 240CU
7800XT = 2x N33 dies, 160 CU
7700XT = 1x N33 die, 80 CU

Everything below that point can be serviced by an additional N34 die and/or recycled N23/24 dies for the extreme low end.

If we're gonna start the hypetrain, lets go big.

I'd love to see this personally, but I can't see AMD going for that many CUs per chiplet. That's a 6900 XT already, unless 5nm is double the density of 7nm? I could see chiplets be 60 CUs for this sort of configuration:

7900XT = 3x N33 dies, 240CU - 3Ghz
7800XT = 2x N33 dies, 180 CU - 3Ghz
7700XT = 1x N33 die, 120 CU - 3.3Ghz
7600XT = 1x N33 die, 60 CU - 3.3Ghz

That would allow more of the GPU stack to use a single chiplet plus I doubt AMD's going to get all the extra performance from a brute force CU increase, rather than tweaks to the IC and boost speeds.

 

[H T C]

What if N33 is the largest single GPU chiplet, and all "N3x's" above it are multi-die "chips"?

So you're looking at something really crazy like:

7900XTX = 3x N33 dies, 240CU
7800XT = 2x N33 dies, 160 CU
7700XT = 1x N33 die, 80 CU

Everything below that point can be serviced by an additional N34 die and/or recycled N23/24 dies for the extreme low end.

If we're gonna start the hypetrain, lets go big.

Wouldn't it be more realistic something like:

7900XTX = 6 N33 dies, 40CU each for 240 CU
7800XT = 4 N33 dies, 40 CU each for 160 CU
7700XT = 2 N33 die, 40 CU each for 80 CU

Always the same N33 die but different number of such dies: wouldn't that be much better from an economic point of view?

Just a thought.

 

[Ajay]

To be honest, the more I myself think about this, the more skeptical I am about N33 being 128 bit.

I don't know what the current rumors are, but 192 bits/die could make sense with the introduction of a victim cache; vertical stack if heat can be dumped fast enough. Still, FPS' will be consuming even more polys and large textures going forward, it seems like a bit ask.

 

[uzzi38]

People had problems with 300W(AMD) or 350W(NVIDIA) TBP and now 450W should be ok? Coolers would need to be very expensive and gigantic. The likelihood of this is pretty close to zero.

Might I remind you we actually already have 450W GPUs

They may be OC model 3090s with like 4 slot coolers, and offer about 5% higher perf over ref making them totally useless, but hey, they technically exist. We have a laugh about them, but consumers don't actually have issues with buying GPUs like this.

 

[Kepler_L2]

To be honest, the more I myself think about this, the more sceptical I am about N33 being 128 bit.

Its just ridiculous thought, in grand scheme of things.

Its not that it is impossible, its ENTIRELY possible. But... Hard to grasp.

In this context, yes. Chiplet and monolithic GPUs are made on two different nodes.

128bit + 256MB IC is actually higher effective bandwidth than 256bit + 128MB IC. If it's using 3D V-Cache, this 256MB IC can use the same or even less amount of die space than 128MB IC in N21 due to SRAM optimized libraries.

 

[Graphenewhen]

How much IC would it take to do away with GDDR memory entirely? Would it be like for like, or would you need less IC?

 

[Kepler_L2]

How much IC would it take to do away with GDDR memory entirely? Would it be like for like, or would you need less IC?

You would need 95%+ cache hit rate for system DRAM to not be a bottleneck, so it's pretty much infeasible.

 

[Trumpstyle]

Assuming that the memory doesn't create a bottleneck that makes your Navi 33 about 56% faster than Navi 21.

That's kind of nutty.

Yep, slightly below 50% as I don't expect perfect scaling from clock increase, but could be higher if AMD pushes clocks.

If by failed to live up to expectations you mean greatly exceeded what most people expected then okay.

People were really skeptical of a 2x performance increase over the 5700XT even though the perf/watt + TDP budget made that seem entirely doable. As it so happens 6900XT is just over 2x the 5700XT at 4k.

I had Navi21 at 275W and 45% faster than Geforce 2080 ti, it ended up at 300W and 35% faster than Geforce 2080 ti, so for me it was below expectations. That's why I'm going with 2.5x for Navi31 instead of 2.7x

20% higher IPC and 30% higher clocks would mean 56% higher performance like Mopetar said. With a 160CU chiplet GPU you would far exceed 2.5x or 150% performance gain you set.


People had problems with 300W(AMD) or 350W(NVIDIA) TBP and now 450W should be ok? Coolers would need to be very expensive and gigantic. The likelihood of this is pretty close to zero.

I had this discussion before in this thread, we shouldn't expect perfect scaling from higher clocks and more CU's.

450W for both GPUs I have been thinking for weeks now ofc someone just tweeted it out yesterday. But I can't find the tweet, it said expect 450-550W for next-gen gpus. Ofc for amd it should be ~450W if they achieve 1.5x Perf/W and we compare it to the 6800 sku.

 

[Gideon]

450W, while less than ideal room-temperature wise in hot climate, is not that big of a problem when the cooled area is relatively big (which it will surely be with chiplet-gpus).

Still at some.point soon we'll hit a wall. It's crazy to think that the "hot and power hungry" GTX 280 had a tdp of only 236W (vs Radeon 4870's measly 150W) and the dustbuster FX 5800 Ultra only had 44W!

 

[DisEnchantment]

450W, while less than ideal room-temperature wise in hot climate, is not that big of a problem when the cooled area is relatively big (which it will surely be with chiplet-gpus).

450W will be toasty in Northern Hemisphere with the global warming, We are getting 33 degrees and barely 3 months ago we had snow.
Our houses are double walled, the windows have two glass panes, one on the outside and one on the inside, to handle the cold. Now it is trapping heat. And I don't have proper AC, it is nuts. It is complicated to install the AC on the thick walls, and I have the mobile one which is garbage.
My 5700XT makes the room very uncomfortably warm, forget about 450W. I am already underclocking/undervolting my 5700XT and applying power limit. One thing is that I never shutdown my PC though. It runs all the time.
Last month my 980 Pro had WHEA consistently after running for some time due to internal thermal protection I think, I was able to solve the problem turning up the fan curve for all system fans, The system is now louder and I hate it.
With 450W, the system can be kept cool but outside the system it is going to be toasty, you might have to use your PC without your clothes on.

 

[Timorous]

I had Navi21 at 275W and 45% faster than Geforce 2080 ti, it ended up at 300W and 35% faster than Geforce 2080 ti, so for me it was below expectations. That's why I'm going with 2.5x for Navi31 instead of 2.7x

So you expected 2.1x more performance at 4k with a 22% power increase. On the same node?

Seems like the problem was your expectations not what AMD delivered. Not that they were far off since N21 is +40% over a 2080Ti on average at 4k and is frequently closer to +50% with a few outliers that drag the average down (using TPU).

 

[Gideon]

450W will be toasty in Northern Hemisphere with the global warming, We are getting 33 degrees and barely 3 months ago we had snow.
Our houses are double walled, the windows have two glass panes, one on the outside and one on the inside, to handle the cold. Now it is trapping heat. And I don't have proper AC, it is nuts. It is complicated to install the AC on the thick walls, and I have the mobile one which is garbage.
My 5700XT makes the room very uncomfortably warm, forget about 450W. I am already underclocking/undervolting my 5700XT and applying power limit. One thing is that I never shutdown my PC though. It runs all the time.
Last month my 980 Pro had WHEA consistently after running for some time due to internal thermal protection I think, I was able to solve the problem turning up the fan curve for all system fans, The system is now louder and I hate it.
With 450W, the system can be kept cool but outside the system it is going to be toasty, you might have to use your PC without your clothes on.

I feel you , I'm also pretty far north (26° 00' E latitude, though gulf-stream helps a lot). We've had 30+ °C for at least two weeks and the forecast is 32-33 °C for the whole following week, and I'm rocking a RX 6800 (non-XT)! The thick walls and glasses actually help a lot when it's only hot for a couple of days. If it's longer than 2-3 days the walls don't only trap the heat, they heat up themselves and at least for my house take at least a day to cool down, emitting heat inside even when it's milder outside.

Luckily I do keep the rig off quite often and usually game late-evening or at night (when the children are asleep) and while it goes dark only after midnight the temperature at least usually drops to 25-ish degrees after 11 PM which is quite OK with windows open. On some days the temperature doesn't drop at night, and that's hell.

During the day I only really want to use my M1 based work-machine, everything else gets stupidly hot.

 

[TESKATLIPOKA]

Might I remind you we actually already have 450W GPUs

They may be OC model 3090s with like 4 slot coolers, and offer about 5% higher perf over ref making them totally useless, but hey, they technically exist. We have a laugh about them, but consumers don't actually have issues with buying GPUs like this.

I meant only standard versions, and there we won't see 450W most likely. I honestly don't care what monstrosity OEMs make or If someone is willing to buy It, but 450W is crazy.

 

[TESKATLIPOKA]

I had this discussion before in this thread, we shouldn't expect perfect scaling from higher clocks and more CU's.

I don't think anyone expects linear scaling, at least not me.
If we used linear scaling, then IPC, clocks and double CU would mean +204% vs +150%. That's 1/4 higher gain and nonlinear scaling from frequency, and CU shouldn't drop It to only 2.5x.

450W for both GPUs I have been thinking for weeks now ofc someone just tweeted it out yesterday. But I can't find the tweet, it said expect 450-550W for next-gen gpus. Ofc for amd it should be ~450W if they achieve 1.5x Perf/W and we compare it to the 6800 sku.

If you think we will see 450W TBP for standard editions, then so be It. I am just really curious how many people would be willing to buy such a card.

edit: Here is a tweet about 450-550W from Greymon55, not sure If that's what you saw.

 

[Gideon]

If you believe we will see 450W TBP for standard editions, It's your opinion. I am just really curious how many people would be willing to buy such a card.

The multi-die chips with that TBP will probably cost an arm and leg anyway N32 and N33 will probably be a lot more reasonable

 

[TESKATLIPOKA]

The multi-die chips with that TBP will probably cost an arm and leg anyway N32 and N33 will probably be a lot more reasonable

I don't doubt It will cost an arm and leg. N32 is supposedly also multi-die so the only "reasonably" priced would be N33, but that's also questionable, If performance similar or higher than RX 6900XT is really true.

 

[jpiniero]

I don't doubt It will cost an arm and leg. N32 is supposedly also multi-die so the only "reasonably" priced would be N33, but that's also questionable, If performance similar or higher than RX 6900XT is really true.

That might be an argument for AMD/nVidia not going as far as they could with the chiplets. Or perhaps only for Proviz models.

 

[Glo.]

128bit + 256MB IC is actually higher effective bandwidth than 256bit + 128MB IC. If it's using 3D V-Cache, this 256MB IC can use the same or even less amount of die space than 128MB IC in N21 due to SRAM optimized libraries.

Im not arguing against it.

It just like the difference between hearing a story about UFO, believing the person, and believing the actual story, that it actually happened.

I WANT TO BELIEVE!

But its.. out of this world idea.

 

[Mopetar]

I wonder if extra cache could be stacked on top of the GDDR PHY. Even on Navi that eats up a little over 10% of the die area and it's not something like the cores that's going to run particularly hot.

So far we've largely been thinking that they're going to use their technology to either double up on IC or alternatively to keep the same amount but cut the die area used in half. But are there other places it could be put as well.