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

 

[Timorous]

Look at the bright side, only 3 more days of making performance charts and then you will be free (at least until RDNA4 leaks)

Speculating and guessing is the fun bit. Once you know the performance it is boring.

 

[uzzi38]

I haven't been keeping fully up to date so I'm sure I've missed something or another, so @DisEnchantment , anything you want to add to the list here? (Anything you reply here I'll just add to the thread)

uzzi38 on Twitter: "Figured since we're so close to launch, might as well do a summary as to what we know about RDNA3. Not going to cover the specs, because I think that's a done deal at this point, but that's not the only part that can be discussed." / Twitter

 

[exquisitechar]

I haven't been keeping fully up to date so I'm sure I've missed something or another, so @DisEnchantment , anything you want to add to the list here? (Anything you reply here I'll just add to the thread)

uzzi38 on Twitter: "Figured since we're so close to launch, might as well do a summary as to what we know about RDNA3. Not going to cover the specs, because I think that's a done deal at this point, but that's not the only part that can be discussed." / Twitter

Awesome summary. I'm particularly curious about what exactly is the Ray Arbiter stuff that was mentioned in the replies. Seems like it hints at some potentially massive additions to the RT hardware acceleration.

 

[utahraptor]

Am I correct in that the leaked picture floating around today RDNA 3 in the flesh is not proof the Radeon RX 7900 XTX is limited to a maximum power draw of 375W? This could just be the Radeon RX 7900 XT?

 

[exquisitechar]

Am I correct in that the leaked picture floating around today RDNA 3 in the flesh is not proof the Radeon RX 7900 XTX is limited to a maximum power draw of 375W? This could just be the Radeon RX 7900 XT?

The ref design is 2 8-pin.

 

[GodisanAtheist]

Something I don't really see anyone discussing is the return of the XTX moniker. I think that alone denotes that AMD has a winner on its hands.

After the absolute disaster that was the 2900XT DAAMIT completely scrapped the 2900XTX launch and has never revived the XTX moniker for its consumer line since.

Part of me thinks this might actually be a sort of unspoken shanty at DAAMIT, that the XTX moniker was either cursed or never to be used until they felt like they had a card that could live up to the name. Otherwise I see no reason for XTX to literally disappear from the product stack for 15 years and reappear now unless AMD feels that the 7900XTX is worthy.

 

[Shmee]

Seems to me that some cards, such as Red Devil or Nitro+/Toxic models might have the opportunity to have a 3rd 8 pin? With a custom PCB. This could allow for more OC room and performance. Anyway, glad to see the 8 pin connectors and no 12 pin in sight.

 

[DisEnchantment]

I haven't been keeping fully up to date so I'm sure I've missed something or another, so @DisEnchantment , anything you want to add to the list here? (Anything you reply here I'll just add to the thread)

uzzi38 on Twitter: "Figured since we're so close to launch, might as well do a summary as to what we know about RDNA3. Not going to cover the specs, because I think that's a done deal at this point, but that's not the only part that can be discussed." / Twitter

We summarized here before

The could be bottlenecked by the vector L0. But, I have seen @Kepler_L2 already mentioned doubled L0.
Fortunately for RDNA, the IFC/MALL also works very well and is going to be in its second generation.
To be seen how they address BW/Latency issues and shader occupancy, divergent memory access.

Several interesting things on N31/2 already in open source commits

• NGG only geometry pipeline
• OREO
• Unified GFX ring/MES only schedule
• 1 Cycle Wave64
• WMMA
• True 16bit ops/similar to mobile
• Many new low precision vector ops, derived from CDNA2+/GFX940
• CDNA2+ features like Architected flat scratch/packed work item ids
• End to End DCC

• NGG only geometry pipeline
• OREO
• Unified GFX ring/MES only schedule --> HW Schedule mainly
• 1 Cycle Wave64 (when supported by 6 VGPR banks)
• WMMA
• True 16bit ops/similar to mobile --> I think this one should be very interesting for Android. VGPR 0-127 (of each bank) can handle dual 16 bit ops
• Many new low precision vector ops, derived from CDNA2+/GFX940
• CDNA2+ features like Architected flat scratch/packed work item ids
• End to End DCC
• Dual Issue (opportunistic for N33/always for N31) for VOP3 ops (VOPD)
• All < 3 operand ops VOPC/VOP2/VOP1 can be dual issued.

I think RDNA3 has a HW BVH traversal stage.

Spoiler: Patents

• 20200380761 COMMAND PROCESSOR BASED MULTI DISPATCH SCHEDULER
  • https://www.freepatentsonline.com/y2020/0380761.html
• 10692271 Robust ray-triangle intersection
  • https://www.freepatentsonline.com/10692271.html
• 10706609 Efficient data path for ray triangle intersection
  • https://www.freepatentsonline.com/10706609.html
• 10930050 Mechanism for supporting discard functionality in a ray tracing context
  • https://www.freepatentsonline.com/10930050.html
• 20210209832 BOUNDING VOLUME HIERARCHY TRAVERSAL
  • https://www.freepatentsonline.com/y2021/0209832.html
  • Possible HW BVH Traversal
• 20210287421 RAY-TRACING MULTI-SAMPLE ANTI-ALIASING
  • https://www.freepatentsonline.com/y2021/0287421.html
• 20210287422 PARTIALLY RESIDENT BOUNDING VOLUME HIERARCHY
  • https://www.freepatentsonline.com/y2021/0287422.html
• 11158112 Bounding volume hierarchy generation
  • https://www.freepatentsonline.com/11158112.html
  • HW Assisted BVH structure generation
• 20210304484 BOUNDING VOLUME HIERARCHY COMPRESSION
  • https://www.freepatentsonline.com/y2021/0304484.html
  • BVH Compression
• 20210407175 EARLY CULLING FOR RAY TRACING
  • https://www.freepatentsonline.com/y2021/0407175.html
• 20210407176 EARLY TERMINATION OF BOUNDING VOLUME HIERARCHY TRAVERSAL
  • https://www.freepatentsonline.com/y2021/0407176.html

 

[utahraptor]

The ref design is 2 8-pin.

Of which model?

 

[alexruiz]

In case you werent aware, top tier cards are not produced in same volumes as lower tier cards. There were not that many 4090s produced initially. Therefore easy for a new, just released flagship product to go out of stock.

I am perfectly aware of the product mixes.
The point is about "who is going to buy them?"
There are people buying them. Even if it is only 30 per store, the top parts will sell out quickly.

Considering that demand of top video cards tends to be inelastic regarding of price, I would even speculate that any GPU of the newer generation that outperforms the top current one while staying close in value for the money will sell well.
The ones that are more elastic are the lower tiers, those ones will need to outperform their predecessors at the same price point to sell decently.

 

[Stuka87]

I still think it would be really strange for them to launch with a 7950 model, instead of just the base 7900. Unless they are going back to the 5K-7K naming scheme of 7950 for cut down, and 7970 for full die.

But if they stick with the naming of the other RDNA cards, this will be a 7900, and then down the road, there will be a 7950 refresh.

 

[In2Photos]

We summarized here before

• OREO

This right here is reason enough to buy AMD over Nvidia. Better cookies!

 

[gdansk]

What are the odds AMD obsoletes the 4080 on paper before its launch on the 16th?
I'm guessing 90%

 

[Saylick]

What are the odds AMD obsoletes the 4080 on paper before its launch on the 16th?
I'm guessing 90%

Inb4 Nvidia unlaunches the 4080 16GB version as well?

 

[A///]

7900Xt matches the 4090 and is priced at 1100, the xtx with 4 GB extra vram will be more powerful than the 7900xt and pass the 4090. No real answer for dlss 3. Rt should be a lot better, maybe equal to nvidia. xtx will be priced at 1300-1400. Both use close but less power to the nvidia 4090.

Other amd cards like the 7800xt wipe the floor with nvidia.

Seems to me that some cards, such as Red Devil or Nitro+/Toxic models might have the opportunity to have a 3rd 8 pin? With a custom PCB. This could allow for more OC room and performance. Anyway, glad to see the 8 pin connectors and no 12 pin in sight.

yes. 8 pins are rated for 150 watts but can pull more temporarily. 150 watts is the safe pull rate. aic can use a third 8 pin for more power depending on what there's left to tap out of the dies. is it ugly? yes it is but it's a lot safer than the melted plastic and fires plaguing their competitor.

 

[Kaluan]

So where are all the RTX 4090s?
OOS everywhere.

OOS claims literally mean nothing without shipment/stock volume data.

Jeez, we've just been through one big lesson of how market manipulation works over the last 2 years, it's sad to see some have learned nothing.

And it's not like we don't have confirmation newb scalpers hedged on this launch being another 30/6000 series gold mine as well.

We summarized here before

• NGG only geometry pipeline
• OREO
• Unified GFX ring/MES only schedule --> HW Schedule mainly
• 1 Cycle Wave64 (when supported by 6 VGPR banks)
• WMMA
• True 16bit ops/similar to mobile --> I think this one should be very interesting for Android. VGPR 0-127 (of each bank) can handle dual 16 bit ops
• Many new low precision vector ops, derived from CDNA2+/GFX940
• CDNA2+ features like Architected flat scratch/packed work item ids
• End to End DCC
• Dual Issue (opportunistic for N33/always for N31) for VOP3 ops (VOPD)
• All < 3 operand ops VOPC/VOP2/VOP1 can be dual issued.

I think RDNA3 has a HW BVH traversal stage.

Spoiler: Patents

• 20200380761 COMMAND PROCESSOR BASED MULTI DISPATCH SCHEDULER
  • https://www.freepatentsonline.com/y2020/0380761.html
• 10692271 Robust ray-triangle intersection
  • https://www.freepatentsonline.com/10692271.html
• 10706609 Efficient data path for ray triangle intersection
  • https://www.freepatentsonline.com/10706609.html
• 10930050 Mechanism for supporting discard functionality in a ray tracing context
  • https://www.freepatentsonline.com/10930050.html
• 20210209832 BOUNDING VOLUME HIERARCHY TRAVERSAL
  • https://www.freepatentsonline.com/y2021/0209832.html
  • Possible HW BVH Traversal
• 20210287421 RAY-TRACING MULTI-SAMPLE ANTI-ALIASING
  • https://www.freepatentsonline.com/y2021/0287421.html
• 20210287422 PARTIALLY RESIDENT BOUNDING VOLUME HIERARCHY
  • https://www.freepatentsonline.com/y2021/0287422.html
• 11158112 Bounding volume hierarchy generation
  • https://www.freepatentsonline.com/11158112.html
  • HW Assisted BVH structure generation
• 20210304484 BOUNDING VOLUME HIERARCHY COMPRESSION
  • https://www.freepatentsonline.com/y2021/0304484.html
  • BVH Compression
• 20210407175 EARLY CULLING FOR RAY TRACING
  • https://www.freepatentsonline.com/y2021/0407175.html
• 20210407176 EARLY TERMINATION OF BOUNDING VOLUME HIERARCHY TRAVERSAL
  • https://www.freepatentsonline.com/y2021/0407176.html

It's missing this:

https://twitter.com/x/status/1587140993442250753

Towards which I have a question:
Can this help confirm mid-gen console refresh (or next gen) will feature RDNA3 tech? Or is the cache directly addressable by developers part "viewable" as just an expected/natural progression of last level cache like that? Reminds of the ESRAM on XBox One in a sense.

 

[A///]

So where are all the RTX 4090s?
OOS everywhere.
I guess those people buying it didn't hear about the recession.
The cards at the top will sell well if the performance delivers

top tier cards aren't made in volume. Scalpers bought most of the cards and began selling them on ebay. nvidia has allegedly halted 4090 production so more of their fab time can go to their enterprise products to meet sales before the china export ban takes place. i mentioned this pages ago. Most places have not gotten in any stock since their initial shipment.

 

[Kaluan]

YT "premiere" thingy is already up on AMD's channel, almost 3 days in advance. The previous/Raphael one only "premiered" 1 or 2 hours before the actual reveal.

Can someone please check the metadata and see how long the announcement video is?

HYPE

 

[A///]

Certain tech sites had the raphael premiere link up before amd's official youtube made it present.

 

[Kaluan]

Certain tech sites had the raphael premiere link up before amd's official youtube made it present.

Yes, links towards content that didn't exist yet AFAIK.
Anyway it doesn't matter, I just found it peculiar. Seems the final touches/editing is (confirmed) done. Now we're just waiting for the date to be (RDNA) the 3rd. 😅

 

[A///]

Yes, links towards content that didn't exist yet AFAIK.
Anyway it doesn't matter, I just found it peculiar. Seems the final touches/editing is (confirmed) done. Now we're just waiting for the date to be (RDNA) the 3rd. 😅

You can set up a future live stream whenever you want down to 5 minutes before the event but hide it through unlisted video. you can send the link out to others. You won't find it without the link. this different from a private link that you need to be authorized to view.

 

[KompuKare]

7900Xt matches the 4090 and is priced at 1100, the xtx with 4 GB extra vram will be more powerful than the 7900xt and pass the 4090. No real answer for dlss 3. Rt should be a lot better, maybe equal to nvidia. xtx will be priced at 1300-1400. Both use close but less power to the nvidia 4090.

But why should they have an answer for fake-frames DLSS 3.0 aside from answering Nvidia's hype machine?
What possible reason use does anyone have for higher but faked frames which add latency?

 

[RnR_au]

What possible reason use does anyone have for higher but faked frames which add latency?

Single player games.

 

[Tuna-Fish]

Something I don't really see anyone discussing is the return of the XTX moniker. I think that alone denotes that AMD has a winner on its hands.

Is there actually anything from AMD about the XTX moniker? I thought that people just speculate it's going to be there because there will likely be 3 SKUs of the top die, instead of 2. (Full with vCache, full, clipped.)

If that is so, I'd expect a new suffix for the vCache product, instead of recycling XTX. Something that ties into the X3D used with CPUs. XT3D?

 

[Saylick]

So I dug into that Twitter thread that @uzzi38 mentioned, and it seems like RDNA 3 does include some sort of ray coherency sorting just like Nvidia and Intel. Nvidia calls it Shader Execution Reordering, Intel calls it Thread Sorting, and I guess AMD calls it Ray Arbiter. Either way, if this is all true, then it explains how AMD are able to catch up faster in RT performance over just the increase in number of FP units.



My apologies for linking to a Tweet with profanity.

Spoiler: NSFW profanity Tweet

https://twitter.com/x/status/1587172727575773187

https://twitter.com/x/status/1587178429778206722