AMD RDNA3 Specifications Discussion Thread

Yes, that's why I reckon 7900XTX is a $700 card. But AMD thinks it can sell this for $999.

Also, notice, I said 3080 with (Ti) in brackets :)

As far as future games and ray tracing is concerned, I think AMD's only escape route is UE5...

Not happy to hear AMD seemingly talk about proprietary RT code for RDNA 3.


Your CPU would not be fast enough for a 4090, it seems and maybe not for a 4080 16GB, either... I can't tell what resolution you use for games.

On the other hand, I hope that ray tracing in more games will make talk about CPU-limited graphics cards redundant. RTX 4090 should have no trouble demonstrating real value against RX 7900XTX.

I can't help thinking that RDNA 4 needs to arrive autumn 2023 or else AMD can join Intel in the irrelevant graphics card tech business.

I have a 34 inch 4k monitor free sync 30-60 fps and a 34 inch ultra wide 3440x1440p 144hz with freesync (dell s3422dwg). I think i'd be able to turn up the graphical settings on the newer games enough that the cpu wouldn't be a big issue
 

Some simple calculated performance figures.

This is EXCELLENT performance for the price. Wow! AMD is killing it.

While I really love RT I realize for the vast majority of gamers, AMD is the far better choice compared to Nvidia. Most people don't care about RT.

Most games will be based on UE5 in the future, and as we know from the UE5 thread, it does have a very competent Software RT solution which still looks great in most instances. So theoretically, if devs were to use that, the 7900XTX would still perform very close to the 4090 even in a next generation game because no Ray accelerators / RT cores are in use then.
Software RT in UE5 currently requires a separate exe/Project File/Game export from the editor

I am not sure If UE5 games realistically will ship with Software AND Hardware, rather just one or the other. I Imagine that fact means Games will ship with Just one or the other only
 
Software RT in UE5 currently requires a separate exe/Project File/Game export from the editor

I am not sure If UE5 games realistically will ship with Software AND Hardware, rather just one or the other. I Imagine that fact means Games will ship with Just one or the other only
You are right about that. Still, even if you enable HW-RT for your shipped project, the fallback mode for cards that are not capable of HW-RT is still present automatically. And given there is a CVar that disables HW-RT, I imagine you could create a setting ingame that disables HW-RT in a game configured with HW-RT enabled easily.
 
Add 10% more performance for the 4090 because TPU used the regular 5800X CPU, which makes the 4090 CPU limited even at 4K, TPU doesn't also test with max ray tracing settings .. I will do proper ray tracing comparisons soon.
The numbers the user made were in error as two of the games were non raytraced numbers

1667514212796.png

Here is the updated numbers from the comparision. In traditional rasterization its a beast and at least based on tehse numbers it would be above a 3090ti in most ray tracing applications. Obviously we need to know how it really does in non marketing benchmarks. But Depending on how the 4080 16gig and whatever the 408012 gig become it could be a decent alternative

Taking a look at the 40x0 series is interesting

The 4090 is 16384 cuda / 2.52ghz clocks / 384bit memory /1018GB
The 4080 16 9728 cuda / 2.51ghz clocks / 256bit memory / 742GB/s bandwidth
The 4080 12 7680 cuda / 2.61ghz clocks/192bit memory /557GB

The pricing is $1600,$1200,$900. The two amd cards have a good chance of out performing the 4080 in rasterization and loosing a bit in ray tracing but coming in at $200/$300 less and the 4080 12 gig which we all know is going to come back since its already made and wil just be renamed has only half the cuda cores of the 4090 and a little more than half the bandwidth.

I think it will be really interesting to see comparisons.
 
Man, hauling ~5.3TB/s offdie on 2.5D is an utterly nuts idea in retrospective (which is why they're only doing it once, really).
a good chance of out performing the 4080 in rasterization
It'll skullfuck it alright, that's by design.
Now can someone normalize RT perf per clock per WGP/TPC for science's sake.
Please.
 
Something seems off about RDNA3. I have the feeling they wanted to do at least one dual GCD card to compete at the top but couldn't get it work properly

It does seem that way. All of the presenters seemed a bit deflated even Lisa. And the constant prompting for applause didn’t help.

355w for 61TF also seems a bit high. Maybe the chiplet interconnect needs a lot of power.
Man, hauling ~5.3TB/s offdie on 2.5D is an utterly nuts idea in retrospective (which is why they're only doing it once, really).

It'll skullfuck it alright, that's by design.
Now can someone normalize RT perf per clock per WGP/TPC for science's sake.
Please.

In a pure RT workload? Probably not pretty. We haven’t seen those benches yet though.
 
Maybe the chiplet interconnect needs a lot of power.
Of course it does, it's 5.3TB/s aggregate over 2.5D.
That's nuts.
Ada L2 is ~5TB/s and that's on-die.
In a pure RT workload? Probably not pretty. We haven’t seen those benches yet though.
I just want to see something iso CU/SM count.
AD102 has 33% more SMs even chopped in 4090 i.e. literally more ray pew-pew hardware.
 
The numbers the user made were in error as two of the games were non raytraced numbers

View attachment 7431

Here is the updated numbers from the comparision. In traditional rasterization its a beast and at least based on tehse numbers it would be above a 3090ti in most ray tracing applications. Obviously we need to know how it really does in non marketing benchmarks. But Depending on how the 4080 16gig and whatever the 408012 gig become it could be a decent alternative

Taking a look at the 40x0 series is interesting

The 4090 is 16384 cuda / 2.52ghz clocks / 384bit memory /1018GB
The 4080 16 9728 cuda / 2.51ghz clocks / 256bit memory / 742GB/s bandwidth
The 4080 12 7680 cuda / 2.61ghz clocks/192bit memory /557GB

The pricing is $1600,$1200,$900. The two amd cards have a good chance of out performing the 4080 in rasterization and loosing a bit in ray tracing but coming in at $200/$300 less and the 4080 12 gig which we all know is going to come back since its already made and wil just be renamed has only half the cuda cores of the 4090 and a little more than half the bandwidth.

I think it will be really interesting to see comparisons.

Alright, looking at the charts...

Next years 7950xt will be faster than the 4090... in some games.

Looks like the RT re-ordering they added has a really, really variable performance boosts. Great on Watchdogs and Cyberpunk but not so much on Metro and even RE Village (ironic that one).

So... $1299? I wonder if they'll stop making the XTX when the 24gbps ram gets here, no reason to use the full good dies on that anymore.
 
You are right about that. Still, even if you enable HW-RT for your shipped project, the fallback mode for cards that are not capable of HW-RT is still present automatically. And given there is a CVar that disables HW-RT, I imagine you could create a setting ingame that disables HW-RT in a game configured with HW-RT enabled easily.
Lumen could improve improve even further on the quality of their software RT by implementing planar reflections for flat and highly specular surfaces ...

Next gen consoles and new GPU have plenty of room in rasterization performance by now to play around with the idea of this feature more commonly and best of all there's no acceleration structure to worry about either. Planar reflections also have the benefit of being higher quality (no noise and potentially use the highest LoD models) as well ...
 
Software RT in UE5 currently requires a separate exe/Project File/Game export from the editor

I am not sure If UE5 games realistically will ship with Software AND Hardware, rather just one or the other. I Imagine that fact means Games will ship with Just one or the other only
You are right about that. Still, even if you enable HW-RT for your shipped project, the fallback mode for cards that are not capable of HW-RT is still present automatically. And given there is a CVar that disables HW-RT, I imagine you could create a setting ingame that disables HW-RT in a game configured with HW-RT enabled easily.

Actually, I just remembered there was a little UE5 Ants demo available to download that does exactly that.

https://dylserx.itch.io/antausventure

Here you can toggle HW-RT on and off ingame.
 
It does seem that way. All of the presenters seemed a bit deflated even Lisa. And the constant prompting for applause didn’t help.

355w for 61TF also seems a bit high. Maybe the chiplet interconnect needs a lot of power.


In a pure RT workload? Probably not pretty. We haven’t seen those benches yet though.

It's bandwidth limitations from 18-20gbps ram. The performance increase over 6950xt assuming 18-20gbps/384bit is X * 1.5 * (112.5?) == 1.5 - 1.7x, thus the performance boosts we see.

The 24gbps just isn't ready on Samsung's part. The ALU clocks should push it much higher, X * 2.2 x * 0.9 == 2x performance or more (that odd split domain thing). Thus the bandwidth limitations come clearly into play.

Samsung just couldn't get the ram they announced up to production levels in time. Was wondering where the announcement of full production was.
 
Last edited:
best of all there's no acceleration structure to worry about either.
What's about huge perf dips in Cyberpunk 2077 when you simply stand in front of a mirror with a single quarter res planar reflection?

Planar reflections also have the benefit of being higher quality (no noise and potentially use the highest LoD models) as well
Gosh, noise comes from physically correct brdf with stochastic sampling. RT reflections obviously don't produce any noise for absolutely flat surfaces, but they do for rough ones. Planar reflections don't support stochastic sampling due to rasterization limitations and thus can't be physically correct at all.
 
Last edited:
Metro Exodus is fast enough even on RDNA2 .. it was designed for RDNA2 consoles after all.

At lower settings, no RT reflections, and reaching 1080p in some areas. Maybe AMD tested it at those low settings, but it's not a 'light' RT game when it's using all of its RT features on PC at all.
 
Anybody got a clue what the GCD/MCD connection is? At 5.7TB/s peak that's already 45W at 1pJ/bit.
I assume it is TSMC based tech... but I have to idea based on those speeds. M1 Ultra interconnect is 2.5TB/s and I haven't seen any good details or deep-dives into their special sauce.
TSMC's CoWoS LowPower interconnect is .56pj/bit but is comparable speed to AMD's IF that is ~2pj/bit.
 
Last edited:
Good, so here are more comparisons:

Cyberpunk 2077, native 4K:

4090: 40fps
3090Ti: 23fps
6900XT LC: 11fps
7900XTX: 17fps (50% faster than 6900XT LC)

The 4090 is 2.3X faster than 7900XTX, the 3090Ti is 35% faster.


Hitman 3 native 4K:

4090: 43fps
3090Ti: 23fps
6900XT LC: 16fps
7900XTX: 26fps (85% faster than 6900XT LC)

The 4090 is 65% faster than 7900XTX

Dying Light 2 native 4K:

4090: 44fps
3090Ti: 24fps
6900XT LC: 11fps
7900XTX: 20fps (56% faster than 6900XT LC)

The 4090 is 2.2X faster than 7900XTX, the 3090Ti is 20% faster

Does the 6900XT LC and 6950XT have the same performance?
 
Back
Top