AMD Radeon RDNA2 Navi (RX 6500, 6600, 6700, 6800, 6900 XT)
AMD's Q4 GPU portfolio was high-end based. No new mainstream products, no new low-end products. Polaris is very old now and AMD doesn't offer any replacement for this price-point yet. Despite lower market share, their margins and maybe event profit could be better because of much higher ASP.
Bondrewd
Veteran
They don't give a shit about lower end client PC/dGPU.
They're very busy hiring people and penetrating channel wherever they can (look up their recent hires and all).
Yeah that's exactly what they're doing by taping shitton of parts out to cover the market and hiring big channel ppl to make sure said products penetrate said markets.
Yeah but the revs were still down YoY (that's despite the mining boom cleaning everything resembling dGPs off the shelves).
Leoneazzurro5
Regular
They are still selling everything they are producing with demand for more, the issue is TSMC being limited in 7nm capacity and with the need to share it with everything AMD does. Not suprising, it was clear with months in advance that APU/laptop and CPU business was first priority for AMD. Moreover they started selling new GPUs in late November, while Nvidia had the full quarter for their RTX 3000 series (with the 3070 only in slight delay).
Rootax
Veteran
I wonder how the mismanagement of production will cost them in the end. Especially on the cpu side, where they have better products than Intel for some sectors, but are not able to gain a lot of market share due to low availability (their are other factors of course, but still).
D
Deleted member 13524
Guest
Some leaked/rumored performance comparisons:
https://wccftech.com/amd-radeon-rx-6700-xt-12-gb-graphics-card-gaming-vulkan-opencl-benchmarks-leak/
If we assume Geekbench's compute benchmark as a measurement of pure compute performance, then the 6700XT has to be clocking like crazy.
The RX6800 averages at 2200MHz which results in around 16.9 TFLOPs. It's getting a 127000 score with 16.9 TFLOPs.
In that leaked result the 6700XT is getting 103000 score, which is the equivalent of a 13.7 TFLOPs in compute performance.
To get to 13.7 TFLOPs, the 40 CUs in the 6700XT would need to be running at ~2680MHz.
We might be looking at the first GPU to ever cross the 3GHz mark on air.
Unless TSMC misses the mark on their performance predictions, it seems like a safe bet to assume almost all desktop RDNA GPUs on 3nm will be running at over 3GHz.
https://wccftech.com/amd-radeon-rx-6700-xt-12-gb-graphics-card-gaming-vulkan-opencl-benchmarks-leak/
If we assume Geekbench's compute benchmark as a measurement of pure compute performance, then the 6700XT has to be clocking like crazy.
The RX6800 averages at 2200MHz which results in around 16.9 TFLOPs. It's getting a 127000 score with 16.9 TFLOPs.
In that leaked result the 6700XT is getting 103000 score, which is the equivalent of a 13.7 TFLOPs in compute performance.
To get to 13.7 TFLOPs, the 40 CUs in the 6700XT would need to be running at ~2680MHz.
We might be looking at the first GPU to ever cross the 3GHz mark on air.
Unless TSMC misses the mark on their performance predictions, it seems like a safe bet to assume almost all desktop RDNA GPUs on 3nm will be running at over 3GHz.
Bondrewd
Veteran
Mismanagement? Oh dear.
Going stonks with CZN/LCN availability being OK and Milan going GA in ~4 days.
The rich man's brand yea.Lolexplains it all.
Heat density would prevent that from happening but still.
Leoneazzurro5
Regular
Contracts for production are signed with a LOT of advance, and "backporting" products on other processes is costly and it takes a lot of time. Not only for the contractor, but for the supplier, too. TSMC itself would have been delighted to know with a 5 years advance that their products would have been so well received by the market and that a global pandemic would have fueled computer devices demand sky high, so they could have easily built the needed capacity and their already groundbreaking results would have been even more stellar.
Last edited:
Throw enough money at it and everything is possible, even ASML building more machines
So far it really looks like AMD mismanaged the situation. AMD planned the whole 2020-2021 mix of RDNA 1/2, Ryzen 3k/5k, APUs 4k/5k, CDNA, EPYC 2/3, two consoles, etc. on a single TSMC N7 or whatever it's called. The need of capacity for the console ramp was known years prior.
Leveraging the EUV variants of TSMC 7nm would surely be costlier but free the available capacity, wouldn't it? I mean putting the high margin RDNA 2 and CDNA 1(2) there would be cool, right?
They're building as fast as they can already.
Yes, but had this demand been known beforehand like the hypothetical scenario above assumed for TSMC, they could have built more manufacturing capacity in those 5 years.
D
Deleted member 13524
Guest
Revenues are up YoY according to their latest FY results.
Do you mean revenues from their discrete GPU business?
Maybe, but the chip isn't that small at 250mm^2.
Bondrewd
Veteran
Oh noes, no.
They don't exist anymore; N6 is later.
Neither of those are even remotely high-margin (or even relevant; ahoy, AMD's winning the mobile game for the first time in their history) versus CPU parts of the portfolio.
dGPUs are a silly accretive business from a CPU house POV.
ASML has a gigantic backlog of stuff that doesn't seem to shrink at this point.
Yeah, we're talking dGPs here and all.
N22 is ~340mm^2 or so.
The issue is obviously hotspots.
Could be the prime motivating factor behind Sony's decision to use liquid metal thermal compound.
Bondrewd
Veteran
LM is also super-nice for CPU temps, and CPU hotspots are a big thing this gen (since, well, first gen in a while with real CPUs).
Frenetic Pony
Veteran
Eh, there's a packaging shortage for chips, among other components. Supposedly AMD has over a billion dollars worth of otherwise finished inventory just waiting around because they can't get non sexy but necessary components, shipping costs are skyrocketing globally, etc. etc.
It's entirely within reason to suspect AMD has a decent amount of actual dies and fab capacity but are simply blocked elsewhere, like literally everyone else on the planet is as well.
It's entirely within reason to suspect AMD has a decent amount of actual dies and fab capacity but are simply blocked elsewhere, like literally everyone else on the planet is as well.
