Top AMD RDNA4 Part Could Offer RX 7900 XTX Performance at Half its Price and Lower Power

[btarunr]

We've known since way back in August 2023, that AMD is rumored to be retreating from the enthusiast graphics segment with its next-generation RDNA 4 graphics architecture, which means that we likely won't see successors to the RX 7900 series squaring off against the upper end of NVIDIA's fastest GeForce RTX "Blackwell" series. What we'll get instead is a product stack closely resembling that of the RX 5000 series RDNA, with its top part providing a highly competitive price-performance mix around the $400-mark. A more recent report by Moore's Law is Dead sheds more light on this part.

Apparently, the top Radeon RX SKU based on the next-gen RDNA4 graphics architecture will offer performance comparable to that of the current RX 7900 XTX, but at less than half its price (around the $400 mark). It is also expected to achieve this performance target using a smaller, simpler silicon, with significantly lower board cost, leading up to its price. What's more, there could be energy efficiency gains made from the switch to a newer 4 nm-class foundry node and the RDNA4 architecture itself; which could achieve its performance target using fewer numbers of compute units than the RX 7900 XTX with its 96.



When it came out, the RX 5700 XT offered an interesting performance proposition, beating the RTX 2070, and forcing NVIDIA to refresh its product stack with the RTX 20-series SUPER, and the resulting RTX 2070 SUPER. Things could go down slightly differently with RDNA4. Back in 2019, ray tracing was a novelty, and AMD could surprise NVIDIA in the performance segment even without it. There is no such advantage now, ray tracing is relevant; and so AMD could count on timing its launch before the Q4-2024 debut of the RTX 50-series "Blackwell."

View at TechPowerUp Main Site | Source

 

[theouto]

That would be an incredible upgrade proposition.

 

[napata]

Déjà Vu.

 

[tfdsaf]

I don't see a reason why they would abandon the high end when the biggest margins are there! The issue with AMD this generation is the price, they've got cheaper silicon compared to Nvidia, yet their pricing has been extremely bad, especially at launch.

Imagine the RX 7900XT launching at $800, it would have been an instant hit, it would have been the go to card with enthusiasts, but at $900 it was overexpensive and lacked value, even the 7900XTX was better value.
Then they launched the 7600 at $270 which was a reduction from the $300 price they initially went for, imagine this card launching at $240, it would have been an entry level hit, it would have been the go-to card for people looking for value and as an entry card.

They also screwed up with the 7700XT price, it should have cost $420 at start and it would have been just as popular, if not more than the 7800XT. At $450 it was worse value than the more expensive 7800xt.

 

[Vayra86]

It could also scratch my back and wash my car

but alas

 

[Kohl Baas]

Not happy about this. The RX7900XTX is a bit lacking in 4K and apparently I won't have any options anytime soon to remedy that problem...

 

[Dirt Chip]

Very good, if it will actually materialise.
Most don’t want a new high end GPU (except to read about them in reviews) with a new record high price.
Most want the same performance level at reasonable cost.
If AMD will pull it over, they will see much applause from the mass, imo.

 

[low]

What i read:

Top AMD RDNA4 Part Could halve the value of your current gpu.

​

I don't see a reason why they would abandon the high end when the biggest margins are there! The issue with AMD this generation is the price, they've got cheaper silicon compared to Nvidia, yet their pricing has been extremely bad, especially at launch.

Imagine the RX 7900XT launching at $800, it would have been an instant hit, it would have been the go to card with enthusiasts, but at $900 it was overexpensive and lacked value, even the 7900XTX was better value.
Then they launched the 7600 at $270 which was a reduction from the $300 price they initially went for, imagine this card launching at $240, it would have been an entry level hit, it would have been the go-to card for people looking for value and as an entry card.

They also screwed up with the 7700XT price, it should have cost $420 at start and it would have been just as popular, if not more than the 7800XT. At $450 it was worse value than the more expensive 7800xt.

Yep, thats unreal. AMD and Intel lied a lot in the past.

 

[MachineLearning]

I desperately hope they don't abandon the high-end. At least one other company needs to compete with nVidia there, don't care if it's Intel or AMD. Just, some kind of pressure...

 

[remekra]

It's not impossible since 7800XT MSRP is 499 so half of 6900XT while according to TPU relative performance is 97%, it's really not that impressive if they will do it.
I surely hope they will not abandon high end, as I will have to switch back to nvidia and get a 5080.

 

[theouto]

I desperately hope they don't abandon the high-end. At least one other company needs to compete with nVidia there, don't care if it's Intel or AMD. Just, some kind of pressure...

I imagine AMD will come back after this generation, and the current gen is just them getting chiplets right on GPUs. And if they manage the goal, then that would open the room for a lot of scalability and perhaps a high end mid-gen gpu? Though that last one is a hail mary of a long shot.

 

[T1beriu]

Another of MLID's junk.

 

[Patriot]

It is finally going to have proper tensor cores, not the half assed matrix cores in the 7xxx series.

Examining AMD’s RDNA 4 Changes in LLVM

As 2024 continues on, because time never stops, AMD has been working on their upcoming RDNA 4 architecture. Part of this involves supporting open source projects like LLVM. If done right, merging t…

chipsandcheese.com

 

[theouto]

It is finally going to have proper tensor cores, not the half assed matrix cores in the 7xxx series.

Examining AMD’s RDNA 4 Changes in LLVM

As 2024 continues on, because time never stops, AMD has been working on their upcoming RDNA 4 architecture. Part of this involves supporting open source projects like LLVM. If done right, merging t…

chipsandcheese.com

This would mean that it is theoretically capable of running DLSS and other algorithms that make use of tensor cores, wonder what the excuse will be for nvidia then. I also wonder if it means that the community could retrofit dlss to run on RDNA4 cards, interesting if true.

 

[remekra]

It is finally going to have proper tensor cores, not the half assed matrix cores in the 7xxx series.

Examining AMD’s RDNA 4 Changes in LLVM

As 2024 continues on, because time never stops, AMD has been working on their upcoming RDNA 4 architecture. Part of this involves supporting open source projects like LLVM. If done right, merging t…

chipsandcheese.com

Don't see anything about "tensor cores" there, just adding new instructions and sparsity. It's not an issue anyway as RDNA3 can do AI when implemented properly (see SHARK):

Stable Diffusion Performance - NVIDIA GeForce VS AMD Radeon

Stable Diffusion is seeing more use for professional content creation work. How do NVIDIA GeForce and AMD Radeon cards compare in this workflow?

www.pugetsystems.com

It's more about nvidia actually using it in gaming (DLSS, RR) while AMD not, which I hope they will finally start to do.

 

[Broken Processor]

I don't see a reason why they would abandon the high end when the biggest margins are there! The issue with AMD this generation is the price, they've got cheaper silicon compared to Nvidia, yet their pricing has been extremely bad, especially at launch.

Imagine the RX 7900XT launching at $800, it would have been an instant hit, it would have been the go to card with enthusiasts, but at $900 it was overexpensive and lacked value, even the 7900XTX was better value.
Then they launched the 7600 at $270 which was a reduction from the $300 price they initially went for, imagine this card launching at $240, it would have been an entry level hit, it would have been the go-to card for people looking for value and as an entry card.

They also screwed up with the 7700XT price, it should have cost $420 at start and it would have been just as popular, if not more than the 7800XT. At $450 it was worse value than the more expensive 7800xt.

They haven't abandoned it they simply can't get multi GCD's working correctly if the rumours are to be believed. So this gives them room to get the following generation working correctly rather than wasting resources fire fighting. It boils down to Radeon group not having the resources it needed and being forced to make hard choices but thankfully with AI it appears that's likely to change but if that helps or hinders GPU development who knows.

 

[Patriot]

This would mean that it is theoretically capable of running DLSS and other algorithms that make use of tensor cores, wonder what the excuse will be for nvidia then. I also wonder if it means that the community could retrofit dlss to run on RDNA4 cards, interesting if true.

It already is, the 7xxx isnt wimpy on performance, it is just feature stripped, It has some features from CDNA1 and some from CDNA2 but not full support of everything from either.
RDNA4 looks like CDNA3 /mi300 feature set.

RDNA2 does bfloat16 nicely which is what the majority of generative AI is using.

RTX4080 bfloat16 is: 97.5/195 regular sparsity
RTX4090 bfloat16 is: 165.2/330.4

Most things do not support sparsity which is why the 7900xtx does things like stable diffusion so well relative to 4080/4090

 

[aktpu]

"Could" but won't

 

[remekra]

This would mean that it is theoretically capable of running DLSS and other algorithms that make use of tensor cores, wonder what the excuse will be for nvidia then. I also wonder if it means that the community could retrofit dlss to run on RDNA4 cards, interesting if true.

Any GPU can could theoretically run DLSS as shown by 1.9 version of it in Control that was running on shaders, not on Tensor cores. Retrofitting will never happen as DLSS is closed source and uses CUDA in it as shown by the leak of it source code some time ago.

 

[Patriot]

Don't see anything about "tensor cores" there, just adding new instructions and sparsity. It's not an issue anyway as RDNA3 can do AI when implemented properly (see SHARK):

Stable Diffusion Performance - NVIDIA GeForce VS AMD Radeon

Stable Diffusion is seeing more use for professional content creation work. How do NVIDIA GeForce and AMD Radeon cards compare in this workflow?

www.pugetsystems.com

It's more about nvidia actually using it in gaming (DLSS, RR) while AMD not, which I hope they will finally start to do.

That is simply because you do not know what tensor means. Tensor is Matrix math.
Google has TPUs tensor cores that specialize in INT8, AMD simply calls the Matrix cores for instinct or AI cores now in RDNA3

How to accelerate AI applications on RDNA 3 using WMMA

This blog is a quick how-to guide for using the WMMA feature with our RDNA 3 GPU architecture using a Hello World example.

gpuopen.com

The function to call the AI cores in RDNA is WMMA ^ that is where the data for my spreadsheet came from.

I don't know if it is better or worse that AMD chose to not just call them tensor cores... It is not a trademarked word, it can't be it is a mathematical term.

 

[Broken Processor]

I desperately hope they don't abandon the high-end. At least one other company needs to compete with nVidia there, don't care if it's Intel or AMD. Just, some kind of pressure...

Sadly for Blackwell generation Nvidia wont have any competition in the high end it seems.

 

[3valatzy]

I desperately hope they don't abandon the high-end. At least one other company needs to compete with nVidia there, don't care if it's Intel or AMD. Just, some kind of pressure...

It's not impossible since 7800XT MSRP is 499 so half of 6900XT while according to TPU relative performance is 97%, it's really not that impressive if they will do it.
I surely hope they will not abandon high end, as I will have to switch back to nvidia and get a 5080.

They haven't abandoned it they simply can't get multi GCD's working correctly if the rumours are to be believed. So this gives them room to get the following generation working correctly rather than wasting resources fire fighting. It boils down to Radeon group not having the resources it needed and being forced to make hard choices but thankfully with AI it appears that's likely to change but if that helps or hinders GPU development who knows.

It's not just the rumours - it is pretty straightforward visible that the Navi 31 in RX 7900 series lacks performance. Nvidia's RTX 4090 with disabled AD102 is 25% faster in ordinary raster and whooping 60-65% faster in ray-traced gaming. AMD failed miserably with the chiplets approach, what do they think?

 

[The Shield]

"Could" but won't

I agree. Someone still considers AMD a "good price/quality brand", but this was true many many years ago. Nowadays, AMD offer is fully in line with Nvidia marketing practice of low volumes and over-high margins.

 

[wolf]

The price had better be compelling, 50% faster than a 6800XT / 3080, 4 Years later? Price will perhaps be the only way it can impress.

 

[stimpy88]

So AMD's best card will be about the performance of a low-end RTX 5070 or 5060. AMD really are incompetent. Good thing they were gifted an amazing CPU architecture that is easy to iterate on, otherwise they would be bankrupt by now.

If this is true, nGreedia are going to completely screw us all over.