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

 

[TESKATLIPOKA]

If TDP for mobile parts is 120W with AIB 6600 XT clock levels, then we could be easily be able to get over 2 GHz, with 96 bit bus clock speeds for 7500 series SKU and no 6 pin connector.

That would be vastly faster than RX 6500 XT, while consuming less power, and delivering more VRAM.

And also be 80-90% larger than RX 6500XT, but ok, being faster It could be sold for more.

And now compare this to a full N33 with 150W TBP? How much performance would be lost?
2.1GHz 28CU 96bit N33 vs 2.8GHz 32CU 128bit N33
15 TFLOPs vs 22.9 TFLOPs (+52%)

What price would you set for this GPU If It has only ~70% of performance? This could be sold for $329 and the full N33 for $449, but wouldn't the full N33 be too expensive? $379 for full N33 would look ok, but then this one should cost only $269 and that's too low in my opinion.

 

[Glo.]

And also be 80-90% larger than RX 6500XT, but ok, being faster It could be sold for more.

And now compare this to a full N33 with 150W TBP? How much performance would be lost?
2.1GHz 28CU 96bit N33 vs 2.8GHz 32CU 128bit N33
15 TFLOPs vs 22.9 TFLOPs (+52%)

What price would you set for this GPU If It has only ~70% of performance? This could be sold for $329 and the full N33 for $449, but wouldn't the full N33 be too expensive? $379 for full N33 would look ok, but then this one should cost only $269 and that's too low in my opinion.

All I can tell you is that I expect 7500 SKU with 6 GB VRAM to be priced around 199-249$ mark.

P.S. RX 6500 XT is exactly 33% faster in gaming than RX 6400. So don't exclude possibility that N33 will come cut down and with 70% performance of full SKU.

 

[Joe NYC]

And also be 80-90% larger than RX 6500XT, but ok, being faster It could be sold for more.

And now compare this to a full N33 with 150W TBP? How much performance would be lost?
2.1GHz 28CU 96bit N33 vs 2.8GHz 32CU 128bit N33
15 TFLOPs vs 22.9 TFLOPs (+52%)

What price would you set for this GPU If It has only ~70% of performance? This could be sold for $329 and the full N33 for $449, but wouldn't the full N33 be too expensive? $379 for full N33 would look ok, but then this one should cost only $269 and that's too low in my opinion.

Navi 33 should be cheaper to produce than Navi 23 - due to smaller die and N6 vs. N7.

Navi 23 (6600 - 6650) sells for $220 -$320

$379 for full N33 full die sounds reasonable.

 

[jpiniero]

Navi 33 should be cheaper to produce than Navi 23 - due to smaller die and N6 vs. N7.

Navi 23 (6600 - 6650) sells for $220 -$320

$379 for full N33 full die sounds reasonable.

It will be $500 at least, or it won't come out if AMD can't justify it.

 

[maddie]

It will be $500 at least, or it won't come out if AMD can't justify it.

Are you saying that a N33 - 32CU/8GB card will sell for $500+ which is (1/2) of a $1000 N31 96CU/24GB card?

 

[TESKATLIPOKA]

Are you saying that a N33 - 32CU/8GB card will sell for $500+ which is (1/2) of a $1000 N31 96CU/24GB card?

If It had 1/2 of the performance and I ignore the small Vram, then I would accept that, but N33 won't have that unless It can clock ~40-50% faster and that's 3.5-3.75GHz. I find that pretty much impossible.

 

[Joe NYC]

It will be $500 at least, or it won't come out if AMD can't justify it.

This is not 2020-2021 and AMD is not NVidia. At $500, these cards would probably not sell.

~$379 would be a reasonable price for the current market, and quite profitable for AMD.

If AMD wanted to be aggressive and gain market share in the lower end segment (where NVidia would likely only be competing with older cards), I could see $329-$349 as a possibility too, for full N33.

 

[Joe NYC]

If It had 1/2 of the performance and I ignore the small Vram, then I would accept that, but N33 won't have that unless It can clock ~40-50% faster and that's 3.5-3.75GHz. I find that pretty much impossible.

Right. So it will not have 50% performance, therefore, 50% price ($500) would be too high.

 

[maddie]

If It had 1/2 of the performance and I ignore the small Vram, then I would accept that, but N33 won't have that unless It can clock ~40-50% faster and that's 3.5-3.75GHz. I find that pretty much impossible.

Exactly and remember it also has 1/3 of the bus to feed those needed clocks. This entire continuing inflated pricing push is grotesque.

 

[TESKATLIPOKA]

Exactly and remember it also has 1/3 of the bus to feed those needed clocks. This entire continuing inflated pricing push is grotesque.

This is a bit questionable. If the clock of N31 was planned to be 3-3.3GHz but ended 2.3-2.5GHz, yet memory subsystem wasn't changed, then BW wouldn't be that much of a bottleneck.

 

[Kaluan]

N31 in this state is a FLOP only the price is at least good.
Power draw during Video and multimedia is embarrassingly bad.

Seems that's one of the first issues being partially adressed with the new 22.12.2 (RX 7900 only) driver, build v22.40.00.61

https://www.amd.com/en/support/kb/release-notes/rn-rad-win-22-12-2-rx7900

Improvements to power usage during hardware accelerated video playback. Further power efficiency improvements are planned for future releases.

Following N31 driver progress for these new cards is gonna be a wild ride. Wilder than during RX 5700 times lol

 

[amenx]

Seems that's one of the first issues being partially adressed with the new 22.12.2 (RX 7900 only) driver, build v22.40.00.61

https://www.amd.com/en/support/kb/release-notes/rn-rad-win-22-12-2-rx7900

Following N31 driver progress for these new cards is gonna be a wild ride. Wilder than during RX 5700 times lol

What a shame they couldnt get this fixed on reviews day. First reviews are where a lot of ppl base their impressions of a product upon.

 

[TESKATLIPOKA]

What a shame they couldnt get this fixed on reviews day. First reviews are where a lot of ppl base their impressions of a product upon.

It's not a shame, It's more like a tradition for AMD Graphic division.

 

[KompuKare]

It's not a shame, It's more like a tradition for AMD Graphic division.

Yes, due to the "we are not a budget brand" the Hawaii firesale probably won't get repeated. If release reviews cause poor sales, they'd rather produce less but maintain higher margins than sell more at lower prices. No idea how low sales are meant to get a return the R&D though.

 

[TESKATLIPOKA]

Yes, due to the "we are not a budget brand" the Hawaii firesale probably won't get repeated. If release reviews cause poor sales, they'd rather produce less but maintain higher margins than sell more at lower prices. No idea how low sales are meant to get a return the R&D though.

They are using RDNA3 IP for their IGPs, so most likely R&D is already paid by the CPU division.

 

[moinmoin]

They are using RDNA3 IP for their IGPs, so most likely R&D is already paid by the CPU division.

That and S3 (Strategic Silicon Solutions, the former Semi-Custom Business unit) / consoles etc. And I'd expect some technology sharing between RDNA and CDNA beyond graphics to still exist.

 

[KompuKare]

They are using RDNA3 IP for their IGPs, so most likely R&D is already paid by the CPU division.

Yes, the main part should be. Or at least they enough volume of all the things which use the main IP.

However, creating Navi 31/32/33 in terms of masks, validation, reference designs and coolers, custom work within the drivers and so on?

Well, sales of actual GPUs has to pay for that.

While their previous high volume, lower margins strategy wasn't

EDIT: ...that successful*, their current strategy of high margins nevermind the volume isn't likely to be that successful either.

* Aside from Hawaii, Radeon Group didn't try the higher end stuff when they had the high volume strategy. A high volume, lower margin strategy backed up with high-end halo offerings to get the brand out there hasn't been tried.

edited as I pressed send to soon on the phone, but it has already been quoted.

 

[eek2121]

I believe nVidia is going up to 180 W for the 4090 Mobile.

This is rather sickening to me, but I can understand why.

You can’t really game on an unplugged laptop for a long time anyway. Shoot, even the APU sucks up a ton of juice. On my laptop you might be able to game at poor framerate for 30 minutes.

 

[Tigerick]

Yes, the main part should be. Or at least they enough volume of all the things which use the main IP.

However, creating Navi 31/32/33 in terms of masks, validation, reference designs and coolers, custom work within the drivers and so on?

Well, sales of actual GPUs has to pay for that.

While their previous high volume, lower margins strategy wasnt that

Yeah, creating masks of new process costs fortunes. My estimate of creating masks for N4 process could cost near to $30 millions.

Thus I believe in order to save $60 millions, AMD will choose to use single die of GCX as leaks by RGT. Of course we are still two years away from release, so just fun facts atm.

Anybody wants to create thread for RDNA4 now?

 

[TESKATLIPOKA]

Yeah, creating masks of new process costs fortunes. My estimate of creating masks for N4 process could cost near to $30 millions.

Thus I believe in order to save $60 millions, AMD will choose to use single die of GCX as leaks by RGT. Of course we are still two years away from release, so just fun facts atm.

Anybody wants to create thread for RDNA4 now?

There is no need, RDNA4 thread was created in March 2022 by DisEnchantment.
Link

 

[jpiniero]

Are you saying that a N33 - 32CU/8GB card will sell for $500+ which is (1/2) of a $1000 N31 96CU/24GB card?

It's pretty clear now that AMD intended to sell N31 for more but felt like they couldn't.

If AMD 'fixes' desktop N32 so that it's not much slower than the 7900 XT, I see it being $799 too.

 

[mrcmtl]

It's pretty clear now that AMD intended to sell N31 for more but felt like they couldn't.

If AMD 'fixes' desktop N32 so that it's not much slower than the 7900 XT, I see it being $799 too.

I can't see that happening. The 7900XT is already not selling at 899$, they would have to drop it to at least 849$ to make it a viable option. The full N32 would have to be at <749$. It just seems too cut down to me that I don't see how it would be anywhere close to the 7900XT unless it clocks at 3.5Ghz+. My bet is at 699$ for ~6900XT. Unless AMD brings forth a 7800XT as an even more cut down N31.

 

[Paul98]

Seems there are power consumption issues in some circumstances, guess we will see if this can be corrected in the drivers. Going to have to see if they can get that sorted and what impact it's having on performance, if any.

 

[Kaluan]

https://youtube.com/@AncientGameplays will have a new video on the new driver up in a bit, he tested his findings saying a bit better raster performance in CP2077 (strange, this is were these cards already did well) and around 30W less during video playback.

 

[SmokSmog]

I will tell you a secret.

7800XT will be on Navi31, it will be rebranded 7900XT but with 4CU less. It will arrive in late Q1

7800XT specs
80CU
320bit
20GB
700$ MSRP

Navi32 60CU ( 7700XT) is a 6800XT competitor because 7900XT with 84CU is barely faster than 80CU 6900XT in raster.

Navi33 is a slow dud.

You don't need to thank me.