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

 

[Joe NYC]

Well it seems AMD will announce the new cards in Gamescom later this month ??

https://twitter.com/x/status/1691864001129443634

That's what MLID said, where AMD would launch Navi 32 cards.

 

[igor_kavinski]

Sorry, 3AM random thoughts.

I like the idea of having two dedicated GPUs, one for each eye, to do better VR.

 

[igor_kavinski]

Gamescom is less than a week away!

I hope AMD surprises us!

(not the bad, shocking kind)

 

[DisEnchantment]

[2/3] drm/amdgpu/soc21: add initial GC 11.5.0 soc21 support - Patchwork

patchwork.freedesktop.org

This patch reads like GFX1150 is actually coming in dGPU/SoC21 form too

 

[Gideon]

(not the bad, shocking kind)

there is also the possibility of "shockingly bad"

 

[Timorous]

Gamescom is less than a week away!

I hope AMD surprises us!

(not the bad, shocking kind)

After their lies in the 7900XTX/XT reveal I will be mentally taking 10% off at least and waiting for reviews. Prior to that reveal they had a solid track record of their presentation numbers being pretty on the money, no idea why they flushed that down the loo.

 

[Joe NYC]

After their lies in the 7900XTX/XT reveal I will be mentally taking 10% off at least and waiting for reviews. Prior to that reveal they had a solid track record of their presentation numbers being pretty on the money, no idea why they flushed that down the loo.

If MLID is to be believed, AMD had to sacrifice some of the RDNA3 performance for stability and bug workarounds in the drivers.

 

[maddie]

If MLID is to be believed

Are you deliberately annoying some here? Kudos to you, I'm enjoying the rabid reply posts.

 

[Panino Manino]

One week for the next disappointment, I'm so excited...
Surprise me AMD.

 

[GodisanAtheist]

AMD's product launches are like "Gurl, prepare yourself for the most disappointing 5 minutes of your life"

 

[gdansk]

AMD's product launches are like "Gurl, prepare yourself for the most disappointing 5 minutes of your life"

It's hit or miss. Zen team keeps announcing killer product after killer product. Meanwhile the RDNA team is disappointing even itself.

 

[Hail The Brain Slug]

It's hit or miss. Zen team keeps announcing killer product after killer product. Meanwhile the RDNA team is disappointing even itself.

I can agree with that, sadly. N21 products (6900XT/6950XT/6800XT/6800) were pretty well a smash hit. Everything else below that, and everything else newer since has been a bit.. disappoint.

 

[Joe NYC]

One week for the next disappointment, I'm so excited...
Surprise me AMD.

Well, we have the baseline, by calculating what percentage Navi32 is vs. Navi31.

So the likelihood is that Navi32 will not fall below this baseline. So then, there is baseline with some room on the upside. That's what we should be watching for. Some upside over the base line or none.

 

[Shmee]

I can agree with that, sadly. N21 products (6900XT/6950XT/6800XT/6800) were pretty well a smash hit. Everything else below that, and everything else newer since has been a bit.. disappoint.

IDK, the 7900XTX had some decent cards at the high end, but it was best to get the chonker cards from Sapphire/Powercolor etc. And the other RDNA2 cards were pretty good in general, particularly the 6700XT and 6750XT. The 6600XT and 6600 were also decent. The main issue was the RX 6500, and the sometimes overpriced 7900 series.

 

[TESKATLIPOKA]

IDK, the 7900XTX had some decent cards at the high end, but it was best to get the chonker cards from Sapphire/Powercolor etc. And the other RDNA2 cards were pretty good in general, particularly the 6700XT and 6750XT. The 6600XT and 6600 were also decent. The main issue was the RX 6500, and the sometimes overpriced 7900 series.

N22 had a bad balance of specs in my opinion. Too much BW and IC for only 40CU.
Should have been 48CU, 192-bit and 64MB IC for the same die size or a bit less.

N24 was bad as you said. Their design choices crippled It a bit too much.
There were 3 big misses.
1. missing media encoding/decoding support
2. 4GB Vram combination with PCIe4 4x
3. very high clocks for 6500XT causing very high TDP

Paired with 8GB Vram It performed decently.
Should have limited TBP to 75W and lowered boost to 2250-2300MHz, higher clocks were already pointless when It just didn't have the required BW.
It was evident If you compared It to RX 6400. Only 28% difference at 1080p despite having 61% more TFLOPs(+33% CU and 21% higher boost).

 

[Aapje]

N24 was bad as you said. Their design choices crippled It a bit too much.
There were 3 big misses.
1. missing media encoding/decoding support
2. 4GB Vram combination with PCIe4 4x
3. very high clocks for 6500XT causing very high TDP

I think that you need to remember that this chip was probably never intended for desktop and was only released there due to the shortages.

In a laptop, two of your three misses aren't as much of an issue:
1. All mobile Zen 3 chips have a basic GPU and can thus be used for media encoding. In general, laptops with a discrete GPU pretty much always have an APU where the discrete GPU is only used for heavy tasks.
3. In a laptop, the clock speeds are kept much lower, so the chip is very efficient

The mobile Zen 3 APU's are limited to PCIe 3, so that is still an issue, although with the lower clocks it might not matter as much.

 

[Tuna-Fish]

N22 had a bad balance of specs in my opinion. Too much BW and IC for only 40CU.
Should have been 48CU, 192-bit and 64MB IC for the same die size or a bit less.

That was not a viable configuration with the elements they had. IC is tied to memory controllers, with 192-bit it's either 48 or 96MB. They decided 48 was too little.

 

[TESKATLIPOKA]

That was not a viable configuration with the elements they had. IC is tied to memory controllers, with 192-bit it's either 48 or 96MB. They decided 48 was too little.

Not 100% sure If It's really tied to It, but let's say It is.
N22 wit 192-bit bus has 24*4MB=96MB or 4*4MB per 32-bit GDDR6.

So It can be either 24MB, 48MB, 72MB or 96MB. 72MB could have been a good compromise although a bit unusual number.

 

[soresu]

[2/3] drm/amdgpu/soc21: add initial GC 11.5.0 soc21 support - Patchwork

patchwork.freedesktop.org

This patch reads like GFX1150 is actually coming in dGPU/SoC21 form too

I must be missing something that is dGPU specific in there?

Is it this line with 'external' in it?

adev->external_rev_id = adev->rev_id + 0x80;

 

[Tuna-Fish]

Not 100% sure If It's really tied to It, but let's say It is.

It's a memory-side cache, where the IC is directly bound to and part of the same logical structure as a single memory controller channel.

N22 wit 192-bit bus has 24*4MB=96MB or 4*4MB per 32-bit GDDR6.
View attachment 84579
So It can be either 24MB, 48MB, 72MB or 96MB. 72MB could have been a good compromise although a bit unusual number.

While GDDR6 chips are 32-bit, it actually has two 16-bit channels per chip⁰, and each of those has it's own IC. With 4MB slices, you can have one or two per channel, meaning 12*4= 48MB or 12*4*2 = 96MB.

0: See page 2 of the standard. (free, but requires registration&login)

 

[TESKATLIPOKA]

While GDDR6 chips are 32-bit, it actually has two 16-bit channels per chip⁰, and each of those has it's own IC. With 4MB slices, you can have one or two per channel, meaning 12*4= 48MB or 12*4*2 = 96MB.

0: See page 2 of the standard. (free, but requires registration&login)

That 4MB slice doesn't need to be that size, right?
Then It can be 1MB, 2MB, 3MB, 4MB or even more.
In total 24MB, 48MB, 72MB or 96MB with 2 slices per channel.

 

[PJVol]

I must be missing something that is dGPU specific in there?

You don't, because it's an APU.

@@ -2358,6 +2361,7 @@  int amdgpu_discovery_set_ip_blocks(struct amdgpu_device *adev)
     case IP_VERSION(10, 3, 7):
     case IP_VERSION(11, 0, 1):
     case IP_VERSION(11, 0, 4):
+    case IP_VERSION(11, 5, 0):
         adev->flags |= AMD_IS_APU;
         break;
     default:

[1/3] drm/amdgpu: add new AMDGPU_FAMILY definition - Patchwork

patchwork.freedesktop.org

 

[soresu]

You don't, because it's an APU.

@@ -2358,6 +2361,7 @@  int amdgpu_discovery_set_ip_blocks(struct amdgpu_device *adev)
     case IP_VERSION(10, 3, 7):
     case IP_VERSION(11, 0, 1):
     case IP_VERSION(11, 0, 4):
+    case IP_VERSION(11, 5, 0):
         adev->flags |= AMD_IS_APU;
         break;
     default:

[1/3] drm/amdgpu: add new AMDGPU_FAMILY definition - Patchwork

patchwork.freedesktop.org

Oh lol didn't even bother to check pt 1 and 3 😂

 

[Tuna-Fish]

That 4MB slice doesn't need to be that size, right?
Then It can be 1MB, 2MB, 3MB, 4MB or even more.
In total 24MB, 48MB, 72MB or 96MB with 2 slices per channel.

In principle, absolutely yes, but that's more development work. AMD designed a single 4MB slice, and used it in all the RDNA2 products, with N23 and N24 using one slice per channel, and N21 and N22 using two per channel.

 

[RnR_au]

Finally! Maybe...

Fixed Issues​

* Improvements to high idle power when using select multi-display setups with mixed high-resolution and high refresh rate displays on Radeon™ RX 7000 series GPUs with variable refresh rate enabled. Further optimizations are being investigated to improve idle power on additional display configurations.

https://videocardz.com/driver/amd-radeon-software-adrenalin-23-8-1