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

 

[eek2121]

I will say maybe 7950 XTX. Faster memory and better binned for a bit more moola than the 7900 XTX. I don't think anything GDDR6 faster than 20 is shipping though.

Not enough power connectors.

 

[jpiniero]

Not enough power connectors.

2x8-pin plus the slot is 375 W. I doubt AMD wants to go any higher than the 355 W that the stock 7900 XTX is (without adding another power connector) but perhaps better binning would help counter the presumed higher power draw of any faster memory.

 

[Ajay]

Not enough power connectors.

Uh, add one. It works for Board Partners for the RTX 4900.

 

[jpiniero]

No more AMD Advantage: MSI Alpha 17 launched with Ryzen 9 7945HX, RTX 4070, and world's first Qualcomm FastConnect 7800 Wi-Fi 7

MSI has announced an all-new Alpha 17 lineup with AMD Ryzen 9 7945HX Dragon Range APUs and a choice of Nvidia GeForce RTX 4070 and RTX 4060 Laptop GPUs. This year's Alpha 17 offerings come with improved QHD 240 Hz displays that cover 100% DCI-P3 and are also the world's first laptops to feature...

www.notebookcheck.net

Doesn't seem like any RDNA3 laptops were announced at Computex. Even this MSI Alpha, which the current one uses N23, is being refreshed with 4060/4070. Still comes with an AMD processor (7945HX)

 

[GodisanAtheist]

This is where AMD getting absolutely massacred in steam charts comes from. OEMs are basically all aboard the Nvidia train and NV is absolutely happy to feed them.

AMD is only now just barely making inroads with their CPUs after multiple successful product generations (and a much better TDP to performance profile on their processors) than Intel.

Embedded processors seem to be the real niche AMD has carved out for itself but those will never be counted on Steam unless they're a Steam Deck, and who knows if even those get counted.

 

[igor_kavinski]

I bet 90% of Chinese e-sports gamers have nVoodoo cards. That alone is enough to skew the Steam results in favor of Jensen's black hole, err...bank account.

 

[Heartbreaker]

Embedded processors seem to be the real niche AMD has carved out for itself but those will never be counted on Steam unless they're a Steam Deck, and who knows if even those get counted.

This entry is steam deck:

AMD Custom GPU 0405 0.30%+0.12%

 

[TESKATLIPOKA]

Also with the current design parameters of Navi 3x series, I think what your more likely to see is a jump in specs akin to Navi 32 to Navi 31 for 400mm2 of add die. Basically a 33% increase in specs vs current Navi 31. What your suggesting is something asymmetrical to the current design of the Navi 3 series making the engineering even more difficult. Also packing it too tightly would limit clocks and just increase the problems AMD already has. You would also need to a couple more MCD making this a monstrous 700mm die and if AMD current problems with Navi 31 regarding clocks and power still existed(likely magnified with a larger die), you would likely just get something equal to the current RTX 4090 maybe even lose, while Nvidia releases a fully enabled AD102.

I calculated It based on the table Locuza made, and I got 50% more, but didn't change interconnects, CP and MM&PCIe. Hopefully, he is right with this table.

If Ada102 with comparable specs is comfortable with 384bit bus and 96MB L2, I don't think this beefed up N31 would need a couple more MCDs, but who knows.
Naturally, I wouldn't keep the same clocks as N31, but ~10-15% lower. This will help to keep TBP to 450W.
RTX 4090 is only 23% faster than RX 7900 XTX, so this should be at worst as fast as RTX 4090 even with 15% lower clocks in raster.

If I wanted to be sure to win by a few % against full Ad102, then I could go big with N30: 428mm2 GCD + 8 MCDs + 32GB Vram with 10% lower clocks and 500W TBP.
I could pack inside a 428mm2 GCD 160CU(+66.7%) : 10240SP(+66.7%) : 640TMUs(+66.7%) : 256ROPs(+33.3%) + 8 MCDs would provide me with 128MB IC(+33%) and 1280 GB/s BW(+33%).

There would be no monstrous Die, that's Ad102, only a monstrous package.
428mm2 5nm GCD + 300mm2(8*37.5mm2) 6nm MCD = 728mm2 in total, but 41% of that is on a much cheaper process than Ada102, so making It shouldn't cost more than Ad102.

With these specs, It would win against full Ad102 in raster and I think It would be at worst pretty close to RTX 4090 in RT.
Price It at $1399(+$400), this would be more than enough for the higher production cost and extra Vram, they would have even a decent extra profit on It compared to RTX 7900 XTX.
RTX 4090 with $1599 would need to go down with the price, because It wouldn't be worth the extra +$200 cost even with DLSS3, considering AMD is probably doing something similar to Frame generation.

Of course, this N30 GCD is only a wishful thinking, even that beefed up N31. AMD will at most release N32, but there was this possibility, which they didn't use.
Hopefully RDNA4 will be much better.

 

[PJVol]

I calculated It based on the table Locuza made, and I got 50% more, but didn't change interconnects, CP and MM&PCIe. Hopefully, he is right with this table.

Idk how you came up with 144 CUs for your "n30". IIRC the full RDNA3 gfx config is 8SE x 16CU, i.e. 128CUs total (area increased by a third). Also this GPU would fit perfectly in the 450W TDP with the same clocks as the 96CU N31.
I believe the main reason it doesn't exist is because they did miss the original perf/W target.

 

[TESKATLIPOKA]

Idk how you came up with 144 CUs for your "n30". IIRC the full RDNA3 gfx config is 8SE x 16CU, i.e. 128CUs total (area increased by a third). Also this GPU would fit perfectly in the 450W TDP with the same clocks as the 96CU N31.
I believe the main reason it doesn't exist is because they did miss the original perf/W target.

My N30 has 160CU. 144CU(+50%) was for a beefed up N31, didn't give It a proper name.
Are you sure RDNA3 is limited to 8SE x 16CU?
N32 supposedly has 3SE x 20CU. If 8SE is the limit, then with 20CU per SE I still end up with 160CU.

 

[MrTeal]

Prices on N22 are just in freefall, at least in the US; $310 will buy you a 6700XT at Newegg. I'm not sure what N32 will bring, but given that the 7600 is already selling under MSRP they're going to need to be much more aggressive on pricing than they were with N33.

 

[Kepler_L2]

My N30 has 160CU. 144CU(+50%) was for a beefed up N31, didn't give It a proper name.
Are you sure RDNA3 is limited to 8SE x 16CU?
N32 supposedly has 3SE x 20CU. If 8SE is the limit, then with 20CU per SE I still end up with 160CU.

gfx11 supports up to 8 SEs, gfx12 up to 32. There is no technical limit on CUs per SE (i.e Arden has 28 CUs per SE).

 

[TESKATLIPOKA]

Prices on N22 are just in freefall, at least in the US; $310 will buy you a 6700XT at Newegg. I'm not sure what N32 will bring, but given that the 7600 is already selling under MSRP they're going to need to be much more agressive on pricing than they were with N33.

With $499 price tag, N32 will have a similar perf/$ as N33.
For the same price Nvidia will give you only RTX 4060Ti 16GB.
Full N32 could perform 20-25% better in raster and similar RT in my opinion.

 

[PJVol]

Are you sure RDNA3 is limited to 8SE x 16CU?

About CU # per SE - not really, although 20CU config in a gfx11 family is yet to be confirmed either. Honestly, the rumored 3 se/60 cu layout for the n32 seems weird, from both architectural and market positioning points.

 

[jpiniero]

TPU has N32 being 64 CUs. The 60 talk could just be that the 7800 XT was intended to be slightly cut down.

With $499 price tag, N32 will have a similar perf/$ as N33.

See that's the thing. I doubt they would be very excited to sell N32 that low.

 

[MrTeal]

With $499 price tag, N32 will have a similar perf/$ as N33.
For the same price Nvidia will give you only RTX 4060Ti 16GB.
Full N32 could perform 20-25% better in raster and similar RT in my opinion.

Sure, and that's probably where they're going to need to position it at most. 4060 Ti at 400 by all accounts isn't moving, and the 16GB at 500 is probably going to be even worse. If 32CU scaling at 128bit bus width is anything to go by for the 6650XT->7600, CU to CU performance is pretty much flat. Full N32 will probably split the difference between the 6800 and 6800XT, and 6800XT's at least in the US where there's still supplies are sitting in stock at $500.

Realistically though the N32 GCD is the same size as N33 on the same process at 200mm², and the GCDs are 146mm² on an older process. With the MCD packing say it costs twice as much as N33 to make. There's twice as many VRAM chips, but PCB, assembly, packaging, distribution and marketing certainly won't cost double for N32 vs N33. There's already 7600's selling in the US for $257. AMD might not want to sell at $500 or below, but if they can sell a 7600 at $257, they can sell N32 at $500 or less.

 

[TESKATLIPOKA]

About CU # per SE - not really, although 20CU config in a gfx11 family is yet to be confirmed either. Honestly, the rumored 3 se/60 cu layout for the n32 seems weird, from both architectural and market positioning points.

It's from Angstronomics. They were right about N31 and N33.

What is so weird about 3SE x 20CU from architectural point?
N22 and N21 also have 20CU per shader engine.
N32 has 1/2 shader engines compared to N31, the same as N22 vs N21, but they added 25% more CUs per engine, so the difference in performance wouldn't be as big as It was with N22 vs N21.

I don't understand your point about weird market positioning at all.
Both 3SE x 20CU or 4SE x 16CU would have similar market positioning, because they would perform pretty similarly.

TPU has N32 being 64 CUs. The 60 talk could just be that the 7800 XT was intended to be slightly cut down.



See that's the thing. I doubt they would be very excited to sell N32 that low.

TPU has no inside info, they just write what they think is correct.
If you think there would be 7800XTX with 64CU, then 60CU is too close to It and there is no point in having It in reserve when both Ada104 and Ada106 is already released.

I doubt they will be more excited with $549 price tag and unsold cards on the shelves.

 

[TESKATLIPOKA]

Sure, and that's probably where they're going to need to position it at most. 4060 Ti at 400 by all accounts isn't moving, and the 16GB at 500 is probably going to be even worse. If 32CU scaling at 128bit bus width is anything to go by for the 6650XT->7600, CU to CU performance is pretty much flat. Full N32 will probably split the difference between the 6800 and 6800XT, and 6800XT's at least in the US where there's still supplies are sitting in stock at $500.

Specs for N32 are the same as RX6800 except IC size, memory speed and GPU clock.
RX 6800XT is only 15.3-15.5% in every resolution based on (TPU).
RX 6800's boost is officially 2105MHz, but TPU measured higher median clocks.
Median:
RX 6800 - 2218MHz (TPU)
RX 7900XT - 2692MHz (TPU)
RX 7900XTX - 2631MHz (TPU)
With 20% higher clocks, you are at 2662MHz, which is in the middle of what N31 managed.
This should be enough for N32 to be on par with RX 6800XT in raster and be faster in RT.

Realistically though the N32 GCD is the same size as N33 on the same process at 200mm², and the GCDs are 146mm² on an older process. With the MCD packing say it costs twice as much as N33 to make. There's twice as many VRAM chips, but PCB, assembly, packaging, distribution and marketing certainly won't cost double for N32 vs N33. There's already 7600's selling in the US for $257. AMD might not want to sell at $500 or below, but if they can sell a 7600 at $257, they can sell N32 at $500 or less.

I also did some calculations and I got 2.64x times higher production cost for just dies(GCD + 4 MCDs) compared to N33 die, excluding the more expensive packaging.

The rest are only assumptions, don't take It like It's a fact.
N33 -> $25 + $5 for packaging
N32 -> $66 + $14 for packaging
N31 -> $102 + $18 for packaging

Let's say AMD asks 37% of the MSRP as the price for N33 from the AIB.
Let's say AMD asks 44% of the MSRP as the price for N32 from the AIB.
Let's say AMD asks 40% of the MSRP as the price for N31 from the AIB.

N33: $269 -> $100 per N33 chip, production cost is $30 and gross profit $70.
N32: $499 -> $220 per N32 chip, production cost is $80 and gross profit $140.
N31: $999 -> $400 per N31 chip, production cost is $120 and gross profit $280.

As you said, N32 could cost less than $499, they will have profit, just not as much.
They can ask for the chip $40 less, AIB will reduce Its gross margin by -$10 and we are already at $449.
But let's be honest, they will ask at least $499 for N32.

 

[PJVol]

TPU took the bait?
https://www.techpowerup.com/308136/amds-radeon-rx-7800-xt-spotted-in-geekbench-vulkan-test
result log shows it's 6950XT )

"id": 5019,
"value": "0x1002",
"ivalue": 4098

"id": 5020,
"value": "0x73a5",
"ivalue": 29605

 

[PJVol]

What is so weird about 3SE x 20CU from architectural point?
N22 and N21 also have 20CU per shader engine.

Actually this, N31 has 16 CUs per SE.
Perhaps there was a reason CUs were down to 16 in rdna 3, although this could be purely layout considerations.

 

[Timorous]

About CU # per SE - not really, although 20CU config in a gfx11 family is yet to be confirmed either. Honestly, the rumored 3 se/60 cu layout for the n32 seems weird, from both architectural and market positioning points.

Makes perfect sense from a layout respective though. Allows for a squarer die and a lot of routing work is already done since N31 is 2 banks of 3 already.

 

[TESKATLIPOKA]

Actually this, N31 has 16 CUs per SE.
Perhaps there was a reason CUs were down to 16 in rdna 3, although this could be purely layout considerations.

I think layout is the culprit.
N23 -> 2SE x 16CU
N22 -> 2SE x 20CU
N21 -> 4SE x 20CU

N33 -> 2SE x 16CU
N32 -> 3SE x 20CU
N31 -> 6SE x 16CU

To me N22 never made a sense, too much IC(+200%) and BW(+50%) for too little extra execution units(+25%) compared to N23.

 

[jpiniero]

TPU has no inside info, they just write what they think is correct.
If you think there would be 7800XTX with 64CU, then 60CU is too close to It and there is no point in having It in reserve when both Ada104 and Ada106 is already released.

I guess since N31 is 96 CU and N33 is 32 CU, it only makes sense that N32 would be 64?

 

[Mopetar]

Realistically though the N32 GCD is the same size as N33 on the same process at 200mm², and the GCDs are 146mm² on an older process.

Navi33 is on 6nm, not 5nm.

 

[TESKATLIPOKA]

I guess since N31 is 96 CU and N33 is 32 CU, it only makes sense that N32 would be 64?

That doesn't really matter.
Look at RDNA2 specs. 32CU N23, 40CU N22 and then 80CU N21.