AMD Radeon RDNA 4/RX 9000/FSR 4

Sadly, FSR 4 at this point in time will be exclusive to RDNA 4. This is what has been shared with partners. Though this may have changed. Tomorrow more will be known, but I don't think the event will be about RDNA 3.

Yes, RDNA 3 does have AI accelerators; however, the architecture lacks dedicated support. IF, FSR 4 ever comes to RDNA 3, then it has been trained as such, or it will be another model.
 
With how stupid Nvidea 50XX card pricing is, im very interested to see where the AMD 90XX cards land

If they can hit some bangers under the 20K/15K/12K mark I might switch to team red

I mean the 7900 XTX is hovering around R20 000 atm and that is an excellent card, its competition is the 4080 Super and that card is still over 30 000.
 
FSR 4 might also spell an end to AMD's open compatibility. It becomes an ecosystem, then; nothing is known whether it will be compatible with Nvidia's or Intel's product stacks.

This might be a new branch.

The FidelityFX-SDK is updated with FSR 3.1.3, which is now also available in Unreal Engine.

 
With how stupid Nvidea 50XX card pricing is, im very interested to see where the AMD 90XX cards land

If they can hit some bangers under the 20K/15K/12K mark I might switch to team red

I mean the 7900 XTX is hovering around R20 000 atm and that is an excellent card

AMD hasn't announced either the 7900 XT or XTX to be discontinued. It is rumored that the 7800 XT is discontinued, but I won't take it as the truth until it is announced by AMD. I do expect the 9070/XT to slide into the GRE slot, scaling up and between the 7900 XT and 7900 XTX depending on the model.

It is not going to be as cheap as people want it to be. The truth will be known tomorrow.

Having seen the leaks, and some reviewers giving them credence, AMD is not trying to sell it to current 7900 XT and XTX owners. I mean, they compared it to the 7900 GRE.
 
I also think that reviewers have already started with benchmarks. Reviewers won't make a peep while under NDA. There have been subtle hints.

Firmware and drivers have been available since December, as indicated by leaks, obviously optimized over time. How reliable the leaks are, well, that will be known on the 5th (or whichever date is announced tomorrow).

The chances that the driver will be on its own branch also seem high, given the tech stack. What a way to cross the bridge to UDNA.
 
I also saw something that FSR 4 might work like DLSS 4 does currently where you'll be able to swap it yourself in the software.

That's absolutely something they need, they don't want to wait and hope devs add or update games for FSR 4, they need it readily available immediately in all the popular games.
 
For those who have seen the Linux benchmark leaks... To give some insight:


Open-Source RADV Radeon Driver Support For RDNA4: "Should Be Good Enough"​


Lead RADV developer Samuel Pitoiset of Valve's Linux graphics driver team has provided some insight into the support expectations for this open-source Radeon Vulkan driver with the upcoming Radeon RX 9070 "RDNA4" graphics cards.

This month there's been many RADV improvements around RDNA4 and its GFX12 graphics IP to land in Mesa 25.0. Ahead of the imminent Mesa 25.0 feature freeze / code branching, Samuel has provided a statement on the RDNA4 support expectations. Presumably the Valve Linux graphics team has already been seeded early hardware from AMD as part of their RADV driver work.

...

Mesa 25.0 is working its way toward a stable release hopefully by the end of February, thus should be out ahead of the Radeon RX 9070 series graphics cards expected to ship in March. The RadeonSI Gallium3D (OpenGL) driver support should be in good shape too given that it's worked on by AMD engineers. It's nice knowing the RADV support for RDNA4 should also be in "good enough" shape for launch day. Over in kernel space, it's likely Linux 6.13~6.14 as the baseline for the necessary AMDGPU kernel graphics driver support. We'll see when the Radeon RX 9070 series graphics cards ship with our Linux testing.

The team would have a sample to work with. MESA 25.0.0 is released:


It does support GFX12, however, I have no idea how stable it is, and I think kernel 6.13 is the baseline. Not sure.

The leaked benchmarks were done on Ubuntu 24.10 kernel 6.11. Something doesn't add up there.

As best as always, it is best to wait until the reviewers publish their scores.
 
I also saw something that FSR 4 might work like DLSS 4 does currently where you'll be able to swap it yourself in the software.

That's absolutely something they need, they don't want to wait and hope devs add or update games for FSR 4, they need it readily available immediately in all the popular games.

Yes, but only with FSR 3 (I think 3.1.0 and up), which is a signed AMD FidelityFX API DLL integration. Obviously newer versions will be supported as such.

EDIT:

Yup,


  • FSR 3.1 introduces the FSR API, which makes integrations of FSR future proof, so applications using this interface can be updated to upcoming versions of FSR more easily.
 
RX 7900 (REFERENCE) INFORMATION TABLES, FOR COMPARISONS:

CLOCK SPEEDS:RX 7900 GRE (TDP 260 W)RX 7900 XT (TDP 300 W)RX 7900 XTX (TDP 355 W)
Base Clock1287 MHz1387 MHz1929 MHz
Shader Clock1880 MHz2025 MHz2365 MHz
Game Clock1880 MHz2025 MHz2365 MHz
Boost Clock2245 MHz2394 MHz2498 MHz
Memory Clock2250 MHz (18 Gbps effective)2500 MHz (20 Gbps effective)2500 MHz (20 Gbps effective)

MEMORY:RX 7900 GRERX 7900 XTRX 7900 XTX
Memory Size16 GB20 GB24 GB
Memory TypeGDDR6GDDR6GDDR6
Memory Bus256 bit320 bit384 bit
Bandwidth576.0 GB/s800.0 GB/s960.0 GB/s

RENDER CONFIG:RX 7900 GRERX 7900 XTRX 7900 XTX
Shading Units512053766144
TMUs320336384
ROPs160192192
Compute Units808496
RT Cores808496
L0 Cache64 KB per WGP64 KB per WGP64 KB per WGP
L1 Cache256 KB per Array256 KB per Array256 KB per Array
L2 Cache6 MB6 MB6 MB
L3 Cache64 MB80 MB96 MB

THEORETICAL PERF.RX 7900 GRERX 7900 XTRX 7900 XTX
Pixel Rate359.2 GPixel/s459.6 GPixel/s479.6 GPixel/s
Texture Rate718.4 GTexel/s804.4 GTexel/s959.2 GTexel/s
FP16 (half)91.96 TFLOPS (2:1)103.0 TFLOPS (2:1)122.8 TFLOPS (2:1)
FP32 (float)45.98 TFLOPS51.48 TFLOPS61.39 TFLOPS
FP64 (double)1,437 GFLOPS (1:32)1.609 TFLOPS (1:32)1.918 TFLOPS (1:32)

I am not going to post 9070/XT specs until they are made known by AMD.

Note that RDNA 4 has newer-generation everything (almost), and may have dedicated DGF decoding hardware (similar to Nvidia's Blackwell GPUs). The latter is not established; it is only written in a white paper and may only be introduced with UDNA. It could be big; the compression tech will boost these GPUs. GFX12 also has FP8, among other new stack supports.
 
For those wanting to know more about DGF:


They say it will come to newer GPUs, but their target sample was an RX 7900 XT, however, there is a cost, and that cost can be negated with specialized hardware. Will this tech come to RDNA 3... Only AMD can provide the answers.

Nvidia has employed a similar technology, but not quite the same approach. Their white paper on it predates AMD by quite some time. I am not going to discuss Nvidia things in this thread.
 
Apparently about $600. Would this be priced around R15-16K? I was thinking of getting the 5070 ti but I can't justify 20k on a GPU. If this is around 16K I'm going AMD.
 
Apparently about $600. Would this be priced around R15-16K? I was thinking of getting the 5070 ti but I can't justify 20k on a GPU. If this is around 16K I'm going AMD.

The launch event was alledgedly already held in China, as reported by VideoCardz (using Chiphell and some Weibo user as their source)


AMD confirms Radeon RX 9070 XT is 42% faster than RX 7900 GRE at 4K​



AMD RDNA4 officially presented in China: Radeon RX 9070 XT priced at 4999 RMB (~$599), RX 9070 at 4499 RMB (~$549)​


I would wait until the real launch, only a little over 3 more hours to go.
 
With how stupid Nvidea 50XX card pricing is, im very interested to see where the AMD 90XX cards land

If they can hit some bangers under the 20K/15K/12K mark I might switch to team red

I mean the 7900 XTX is hovering around R20 000 atm and that is an excellent card, its competition is the 4080 Super and that card is still over 30 000.

True. I use a 7900XTX and a 4080S. The 1% lows in FPS are way better in my 4080S as compared to the 7900XTX. It also just feels 'smoother' and has less stutter/lags on my 4080S for some reason. Could be Driver related.

The 4080S is ~R10k more but it is worth the price imo.
 
Looking at all the presentation slides. AMD sure used the "Artistic Render" GPU everywhere. That GPU did exist, in what quantities I don't know. I am calling it; that GPU was pulled so that they don't have to attach a sticker price (MSRP) to a base model that they have made under their own label.

We have already seen this with the RTX 5070 Ti.

I hope that I am wrong.
 
Top
Sign up to the MyBroadband newsletter