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]

Why would it be any less than it's predecessor? Especially given current GPU pricing (if we are talking about a 7800XT). Sadly for AMD - NV has really made the whole AI and Ray-tracing branding stick as prominent checkboxes for GPU purchase considerations. Maybe that, and lagging sales, will keep the price down a bit.

N32 XTX is cheaper to produce than N21 XT.

 

[insertcarehere]

N32 XTX is cheaper to produce than N21 XT.

Doesn't mean anything when N21 is winding down production (my market stopped stocking RDNA2 for months) and N32 has its own set of R&D expenses which need to get amortized down.

Not to mention that full N21 XT launched at $999 MSRP..

 

[PJVol]

Doesn't mean anything when N21 is winding down production

Exactly, since N32 and N21 are to occupy different market niches performance-wise.

Why would it be any less than it's predecessor?

It wouldn't considering its predecessor Navi 22 XT was launched at $479 (and Navi 22 KXT now is $429 at amd.com), a little fact that many here seem to have forgotten.

 

[maddie]

Limbo time.

AMD Radeon RX 6300 entry-level RDNA2 desktop GPU with 32-bit memory has been spotted for less than $60

AMD Radeon RX 6300 entry-level RDNA2 desktop GPU with 32-bit memory has been spotted for less than $60 - VideoCardz.com

AMD Radeon RX 6300 with 32-bit memory bus appears on Chinese market A low-profile Radeon RX 6000 card with a single-fan design represents the lowest tier in AMD product family. AMD Radeon RX 6300 desktop GPU, Source: Goofish Technically speaking, AMD has not confirmed a Radeon RX 6300 desktop...

videocardz.com

 

[TESKATLIPOKA]

Why would it be any less than it's predecessor? Especially given current GPU pricing (if we are talking about a 7800XT). Sadly for AMD - NV has really made the whole AI and Ray-tracing branding stick as prominent checkboxes for GPU purchase considerations. Maybe that, and lagging sales, will keep the price down a bit.

$599 is more than what N22 cost as @PJVol already mentioned.
With performance at 6800xt-6900xt level, AMD can't ask much more.
In my opinion, $649 is the limit they can ask. It would have comparable perf/$ to 7900XT, but I don't think It would sell well.

P.S. Maybe most will go to laptop, there It could do better.

 

[TESKATLIPOKA]

Limbo time.

AMD Radeon RX 6300 entry-level RDNA2 desktop GPU with 32-bit memory has been spotted for less than $60

AMD Radeon RX 6300 entry-level RDNA2 desktop GPU with 32-bit memory has been spotted for less than $60 - VideoCardz.com

AMD Radeon RX 6300 with 32-bit memory bus appears on Chinese market A low-profile Radeon RX 6000 card with a single-fan design represents the lowest tier in AMD product family. AMD Radeon RX 6300 desktop GPU, Source: Goofish Technically speaking, AMD has not confirmed a Radeon RX 6300 desktop...

videocardz.com

32-bit memory means only 2GB Vram. If It had at least good video encoding(decoding) capabilities, but It does not. Useless card in my opinion.

 

[beginner99]

Honestly, I don't believe It will be this cheap, $599 looks much more likely.

Even $599 seems highly optimistic given the last couple years and current release cycle. that is $200 less than for a 7900 xt. it will only happen if the gap is bigger than expected and N32 also has issues and can't clock as high as expected (eg close to 3 ghz).

 

[Kaluan]

Even $599 seems highly optimistic given the last couple years and current release cycle. that is $200 less than for a 7900 xt. it will only happen if the gap is bigger than expected and N32 also has issues and can't clock as high as expected (eg close to 3 ghz).

7900XT keeps coming down in price and desktop N32 is likely many many months away still. Let's give them some benefit of the doubt guys/gals... 😅

I think the soon(er) to be released 7600/7600 XT pricing will give us the best insight on N32 pricing. More than any other idle speculation.

$299 7600 and $349 7600 XT? Then sure, I could see them go for $599 N32 or less.

$399 for a N33? No chance they price 7800XT less than $649.

 

[Tuna-Fish]

Which N32 model are you talking about?

I think it's about 99% certain that there will be at least two, one with 4 MCDs and 16GB ram and another with 3 MCDs and 12GB ram. I think the weaker model will definitely retail for less than $600, even at launch. The stronger model might get down there too, depending on the time until launch and how the economy looks at that point.

 

[jpiniero]

Which N32 model are you talking about?

AMD did only release one N22 desktop model initially.

 

[GodisanAtheist]

AMD did only release one N22 desktop model initially.

- AMD is trying to cover this entire gen with 3 chips and their largest single die is 300mm2, we'll undoubtedly see 2 iterations of the N22. MCD's are an almost too perfect place to do some product segmentation.

 

[beginner99]

Which N32 model are you talking about?

The fully enabled / highest SKU. A lower SKU with less vram and less CUs yeah that could be at $599 but it will obviously also perform worse.

 

[Kaluan]

Which N32 model are you talking about?

I think it's about 99% certain that there will be at least two, one with 4 MCDs and 16GB ram and another with 3 MCDs and 12GB ram. I think the weaker model will definitely retail for less than $600, even at launch. The stronger model might get down there too, depending on the time until launch and how the economy looks at that point.

The full-fat + 4 MCD one.
I can't see them asking $599-649 for the (cut N32) 7700 XT.

 

[biostud]

The full-fat + 4 MCD one.
I can't see them asking $599-649 for the (cut N32) 7700 XT.

If the performance lands ~6950XT for the full N32 it will be a hard sell if it is more than $599. (Which is still very little improvement in price/performance ratio over N21 cards)

 

[Kaluan]

If the performance lands ~6950XT for the full N32 it will be a hard sell if it is more than $599. (Which is still very little improvement in price/performance ratio over N21 cards)

Hmm, SKU branding-wise I agree it wouldn't look great for just a 15-20% bump from 6800XT to 7800XT. But chip tier-wise (N22 to N32) it would be pretty massive.
Also, that's mostly raster talk, RT will likely spring quite a bit above anything RDNA2.

We just don't know what's the hold-up. Reasonable speculation says they held the release back so they can have a better showing. Which might mean working 3GHz in-game clocks or more and less silicon bugs. But they can still save it even if it's quasi-broken like N31, they just need not be morons about the pricing.

Jeez, even RX 7900 launch could have looked objectively better with just2 minor changes:

RX 7900 XTX - RX 7900 XT
RX 7900 XT $899 -> RX 7900 $799

That being said, if they bork up the pricing, personally I may just get second hand or the best value N23/33 SKU in my market and just weather the storm until next year.

 

[Timorous]

N32 with 4 MCDs is less silicon than N22 is. Probably costs a bit more because 200mm is N5 and the packaging costs but I don't think it will be massively more.

That means I think a $550-$600 full fat N32 would actually have higher margin than the $480 MSRP 6700XT and it would be far far higher margin than the $650 MSRP 6800XT. From that perspective AMD charging that kind of price is entirely on the cards, especially if it ends up being close to the 6950XT in performance.

I can see AMD going for that win win. Offer a product that performs well at an attractive price point while increasing their margin at a given price point can't be too shabby, even more so if the 4070 is actually $749.

 

[jpiniero]

- AMD is trying to cover this entire gen with 3 chips and their largest single die is 300mm2, we'll undoubtedly see 2 iterations of the N22. MCD's are an almost too perfect place to do some product segmentation.

The N32 GCD is allegedly 200 mm2. Between the high yields (and presumably low volume), AMD should in theory be able to fufill the partially busted GCD with mobile parts just fine. That being said, I don't know if missing the boat is going to cost them the mobile OEM deals. But if it does, it would be more of an argument to cancel N32 entirely rather than make 2 desktop models. With the 7900 XT seemingly not selling at all, I don't think AMD is very excited.

 

[GodisanAtheist]

The N32 GCD is allegedly 200 mm2. Between the high yields (and presumably low volume), AMD should in theory be able to fufill the partially busted GCD with mobile parts just fine. That being said, I don't know if missing the boat is going to cost them the mobile OEM deals. But if it does, it would be more of an argument to cancel N32 entirely rather than make 2 desktop models. With the 7900 XT seemingly not selling at all, I don't think AMD is very excited.

-- It'd be interesting to see if the failure rate on these parts is too low, AMD to actually put a hard clock limit on the lower tier SKUs along with dropping an MCD. AMD has put in clock limits before with stuff like the 6800 (non-XT) because they knew it would get too close to the 6800XT at equal clocks.

So you get a 60CU GCD no matter what, but the XT edition runs at 3Ghz w/ 4 MCDs and the Non-XT edition runs at 2.5 Ghz w/ 3MCDs.

It would bring that CPU design philosophy down to GPUs, where the primary differences between some skus is just energy and temp savings.

 

[Tuna-Fish]

How good the clipped N32 is depends entirely on how they clip it. If they disable a single entire SE, it will be much slower than if they evenly disable 9 WGPs, even if the flops numbers will be nearly identical.

 

[Kronos1996]

In what reality can 64CU gpu be close to the 84CU one?

The 6800 has 50% more CU’s then the 6750 XT while only being 10-20% faster. The 7900 XT has 40% more CU’s then what the 7800 XT likely will. The math ain’t that hard.

Scaling will probably be similar if not worse, depending on if Navi 31 has hardware flaws that Navi 32 fixes. Even in a best case scenario for the 7900 XT though, it’s unlikely to be more then 10-15% faster then a card that cost substantially less to make. Possibly even worse if the Navi 32 silicon turns out much better then Navi 31 did.

 

[GodisanAtheist]

The 6800 has 50% more CU’s then the 6750 XT while only being 10-20% faster. The 7900 XT has 40% more CU’s then what the 7800 XT likely will. The math ain’t that hard.

Scaling will probably be similar if not worse, depending on if Navi 31 has hardware flaws that Navi 32 fixes. Even in a best case scenario for the 7900 XT though, it’s unlikely to be more then 10-15% faster then a card that cost substantially less to make. Possibly even worse if the Navi 32 silicon turns out much better then Navi 31 did.

- TBF the 6750XT had a gameclock of 2495Ghz while the 6800 had a gameclock of 1815Mhz. That's a whopping 37% clockspeed advantage on the part of the 6750XT which really helps it make up lost ground.

The 7900Xt has really conservative clocks @ 2025Mhz and would only have a 40% CU advantage. N32 @ 2.7Ghz would basically land right on top of it.

 

[Kronos1996]

- TBF the 6750XT had a gameclock of 2495Ghz while the 6800 had a gameclock of 1815Mhz. That's a whopping 37% clockspeed advantage on the part of the 6750XT which really helps it make up lost ground.

The 7900Xt has really conservative clocks @ 2025Mhz and would only have a 40% CU advantage. N32 @ 2.7Ghz would basically land right on top of it.

Pretty much what I was thinking yeah. The cut down memory bandwidth might hurt it more at 4K but I wouldn't be shocked if they’re barely 10% apart.

 

[insertcarehere]

The 6800 has 50% more CU’s then the 6750 XT while only being 10-20% faster. The 7900 XT has 40% more CU’s then what the 7800 XT likely will. The math ain’t that hard.

GPUs are not only comprised of CUs

6800 vs 6750xt:
-50% more CUs (60 vs 40)
-18.5% more vram bandwidth (512GB/s vs 432GB/s)
-33% more ICache (128mb vs 96mb)
-50% more ROPs (96 vs 64)

Not to mention the significantly higher clock speeds (and power draw) of the latter. 6800 is basically N21 with 1/4 SEs turned off, so it is 3 Shader Engines vs full N22 with 2 Shader Engines.

7900xt vs 7800xt (full N32):
- 40% more CUs (84 vs 60)
- 25(?)% more VRAM bandwidth (320bit vs 256bit)
- 25% more ICache (80mb vs 64mb(?))
- 100% more ROPs (192 vs 96(?))

As per Angstro's RDNA3 leaks, N32 is supposedly a 3SE part, while 7900XT has all 6 of N31's SEs activated (192 ROPs is full count for N31). So 7900xt is (cut) 6SE design vs the 3SE design for N32.

In many respects the gap between 7900xt and full N32 should be larger than between 6800 and full N22.

 

[PJVol]

The 6800 has 50% more CU’s then the 6750 XT while only being 10-20% faster. The 7900 XT has 40% more CU’s then what the 7800 XT likely will. The math ain’t that hard.

Indeed, 6800 has 1.5x CUs of 6700xt and if we take game/boost clocks as the base, its 0.75 - 0.8x deficit makes it only 12 - 20% faster, though based on actual average gaming clock the difference is more like 25%.
So, yeah... I'd say the math turned out to be even easier ))

With regards to 7900XT, thefpsreview.com measured its average clock in games ~2600mhz, therefore to compensate CU deficit of 1.4x the 7800xt should run above 3600 mHz on average in games. And I suspect running at these frequencies (which in itself seems higly unlikely) may well require more power than 7900XT consumes.

I could call 3ghz avg. a safe bet, hence the ~6800XT-6900XT territory, though given the current state of rdna 3 execution, even that seems quite optimistic. Thus my question is still valid I think.

 

[leoneazzurro]

I think no one realistically thinks about full N32 being as fast as the 7900XT, the possible target is the 6900XT.