AMD Navi 23 ‘NVIDIA Killer’ GPU Rumored to Support Hardware Ray Tracing, Coming Next Year

queamin · 5 Sep 2020 at 17:36

Lokken86 said:
If it was actually 160CUs that would be a bit of a 4870 repeat. Surely impossible on 7nm.

They been on 7nm a while who knows.

melmac · 5 Sep 2020 at 17:37

Gerard said:
AMD cpu's never had driver issues haunting them for years, that's the big problem that stops more people buying radeon cards. Even if their drivers were rock solid it's still something that's been talked about enough to stop people buying amd cards.

For a few years they were great with drivers, right from the November of the year they released the 7970 up until the 5700 release. Sure they had minor slip ups, but no drivers from either company are perfect and there will always be small issues. They were just getting over the stigma of bad drivers when they messed up with the black screens on the Navi 1 cards.

The other problem is that during the times they have had good drivers, they have really messed up launches. Tahiti, Hawaii, Fury, Polaris, Vega.

Hopefully for RDNA 2, the launch will be good and Drivers will be good.

gpuerrilla · 5 Sep 2020 at 17:41

Poneros said:
Not much to say, it's in-line with other rumours & guesswork. I'm still sad over HBM. For me what's most important is what they do software-wise, AMD generally doesn't fail with the chips themselves. But obviously we won't know anything about that until they actually reveal them.

Yeah your right, looks like no HBM this round.

Chuk_Chuk · 5 Sep 2020 at 17:43

danlightbulb said:
Interesting.

Does this suggest AMD have doubled up on the CUs as well, like NVidia has done? There are 8 blocks of 20 CUs (2 blocks per shader engine) shown there which is 160 CUs total.

And there are 12 ram boxes could indicate 12 Gb.

RDNA1 had double CUs, so it is just a continuation of that design.

Gerard · 5 Sep 2020 at 17:45

melmac said:
For a few years they were great with drivers, right from the November of the year they released the 7970 up until the 5700 release. Sure they had minor slip ups, but no drivers from either company are perfect and there will always be small issues. They were just getting over the stigma of bad drivers when they messed up with the black screens on the Navi 1 cards.

The other problem is that during the times they have had good drivers, they have really messed up launches. Tahiti, Hawaii, Fury, Polaris, Vega.

Hopefully for RDNA 2, the launch will be good and Drivers will be good.

They have had good drivers for many years then the black screen bug for navi rears it's head, which seems to be almost a legacy bug when there used to be a grey screen bug in older cards.

The big issue with their launches is the drivers always seem rushed, you would think for the launch of a new card they would have a really polished set of drivers ready to go, but more often than not it seems like they've been slapped together at the last minute. The other major issue is performance improves over time, all well and good but surely you would want most of that performance to be available for reviews to show the card in a better light? First impressions are everything, people want to know the performance when reviews hit, not 3-6 months down the line where it gets a re-review to see how much performance has increased. By that point most people have already bought their upgrade based on reviews a few months earlier.

Just baffling that it takes AMD months to get to grips with their drivers on virtually every release, think they need to start putting more resources into their software dept.

humbug · 5 Sep 2020 at 18:17

Chuk_Chuk said:
Just to add to this, from the MS hot chips presentation the XSX has 26 active dual CUs or 52 CUs. If this image is true and if it can properly utilise its CUs this will be significantly faster than the XSX.

the XSX has 5Mb L2 Cache, if i am reading this diagram right, this only has 6MB which seems low considering the size difference between the two.

Wait what? Buy that measure this is either 26 dual FP CU's vs 80 dual FP CU's or 52 singular FP CU's vs singular FP 160 CU's.

I can't believe that, because if it is true Navi 21 will crush a 3090, no, i have to believe its Dual FP 52 CU's.

EastCoastHandle · 5 Sep 2020 at 18:20

It's been said way back they had a 160CU card. But from my understanding this is for professional use not gaming.

TNA · 5 Sep 2020 at 18:22

humbug said:
Yes, you must know me well enough by now...

Unattainable clock speeds with my early silicon, very unattainable...

You waited this long with that RX 580?

Haha. I was slightly concerned that I may have over teased you with floppy poppy

Ah yes, I forgot about my CPU OC

The RX 580 is a stop gap card. It has enough grunt to play at least 50% of my Steam library so does the job for now.

muon · 5 Sep 2020 at 18:23

It's not a 4x increase in CUs. Someone already explained it above.

If it was the FLOPs number would also be 4x. But it is only 2x.

Chuk_Chuk · 5 Sep 2020 at 18:28

humbug said:
Wait what? Buy that measure this is either 26 dual FP CU's vs 80 dual FP CU's or 52 singular FP CU's vs singular FP 160 CU's.

I can't believe that, because if it is true Navi 21 will crush a 3090, no, i have to believe its Dual FP 52 CU's.

humbug · 5 Sep 2020 at 18:29

muon said:
It's not a 4x increase in CUs. Someone already explained it above.

If it was the FLOPs number would also be 4x. But it is only 2x.

It may not be the gaming card, tho i thought that had 128 CU's? in any case i don't know where the blue tab on the right side came from but it looks to be making the same mistake people did with Ampere, the reason Ampere has 2X the TF of what was leaked and rumoured is because no one realized Nvidia are using dual thread cores, each CU has twice the FP units that Turing had. 1 Turing vs 2 Ampere.

This, in RDNA2 is exactly the same thing.

humbug · 5 Sep 2020 at 18:32

humbug said:
It may not be the gaming card, tho i thought that had 128 CU's? in any case i don't know where the blue tab on the right side came from but it looks to be making the same mistake people did with Ampere, the reason Ampere has 2X the TF of what was leaked and rumoured is because no one realized Nvidia are using dual thread cores, each CU has twice the FP units that Turing had. 1 Turing vs 2 Ampere.

This, in RDNA2 is exactly the same thing.

Look to put it bluntly, just like Turing vs Ampere, it does not have 5120 Shaders, it has 10240.

humbug · 5 Sep 2020 at 18:37

Chuk_Chuk said:

That's 28 Dual, or 56 Singlar Shaders..... *Gulp*

4K8KW10 · 5 Sep 2020 at 19:12

humbug said:
That's 28 Dual, or 56 Singlar Shaders..... *Gulp*

26 dual compute units is 3328 shaders.

Lokken86 · 5 Sep 2020 at 19:15

Is it officially 56CUs now or what?

muon · 5 Sep 2020 at 19:16

Lokken86 said:
Is it officially 56CUs now or what?

No one knows anything.

Apparently it could be 160 CUs, which is a little absurd.

Chuk_Chuk · 5 Sep 2020 at 19:21

Lokken86 said:
Is it officially 56CUs now or what?

If you are talking about the XSX, it has 56 CUs for yield reasons.

gpuerrilla · 5 Sep 2020 at 20:09

Chuk_Chuk said:
If you are talking about the XSX, it has 56 CUs for yield reasons.

Thought it was 54 with 2 disabled?

Chuk_Chuk · 5 Sep 2020 at 20:11

Th0nt said:
Thought it was 54 with 2 disabled?

You are correct, i shouldn't have been lazy with my response.

"56 CUs on the die, 4 of which are disabled to improve yields for a total of 52 active CUs"

4K8KW10 · 5 Sep 2020 at 20:12

Th0nt said:
Thought it was 54 with 2 disabled?

They are 52 out of 56, or 26 dual out of 28 dual. 4 CU cut for redundancy, I guess every chip is cut in a different area... so to increase usability of the dies.