No DX12 Software is Suitable for Benchmarking *spawn*

Is that an actual technical limitation though?
Yes. Lots of stuff with RT doesn't fit nicely into the "the CPU will spoon-feed the CPU one command at a time" model of previous APIs. It is fundamentally a lot more retained mode, laying out data structures (often asynchronously) in GPU memory to trace against.

Obviously you could add a significantly more advanced memory and binding model onto DX11 in theory but then... that's sorta what the new APIs do. I'm sure we could bolt some form of raytracing onto Glide but that doesn't mean it's a good fit.

There are increasingly a myriad of reasons why the old APIs are become unsuitable for GPU-driven render pipelines. Efficient memory management of transient resources in a frame is another big one.

Furthermore FXC is just exploding under the weight of modern "shaders"/compute forcing a variety of unsustainable hacks like passing them through DXC first, then outputting simplified shader permutations back into FXC and crossing your fingers (and twiddling you thumbs).

DX11 worked well for a long while but it is definitely showing its age and is not a sufficient vehicle for the future. Luckily DX12 is quite established at this point, hence being able to tie the new graphical features of UE5 to DX12/Vulkan+.
 
Last edited:
I’m surprised there aren’t any articles yet. Maybe 3Dmark is truly dead.
I got it and tried it with my 2080ti. Scored around 3000... But I'll be honest... it's a pretty small, limited demo. Probably not worth the money.

The developers said people were wanting an "interactive" test, instead of a "demo" with sound and more complexity... I don't really believe that. I want demos like TimeSpy and FireStrike. This is just a single test, in a single scene. No CPU or physics really to speak of.

And I think that's why we never saw anything about this test until it was out. There were no videos uploaded of this test beforehand.. it just released. If I'd have known, I wouldn't have bought it. I was expecting a race around a speedway with multiple bikes and nice RT visuals.... and while this looks nice.. it's not nearly as "next gen" as I expected.. and that's likely because UL doesn't have the budget to make something like that.
 
New nvidia drivers are supposed to have dx12 improvements. Curious to see if any sites do comprehensive benchmarks and if it changes the story vs dx11
 
New nvidia drivers are supposed to have dx12 improvements. Curious to see if any sites do comprehensive benchmarks and if it changes the story vs dx11
I thought I saw a post saying those improvements were included in the reviews already.
 
New nvidia drivers are supposed to have dx12 improvements. Curious to see if any sites do comprehensive benchmarks and if it changes the story vs dx11
If the press driver is the same as the driver that came out today then the DX12 improvements are only for Ampere Turing.
This Game Ready Driver also includes significant DirectX 12 performance optimizations which can dramatically improve performance for GeForce 30 Series GPUs.
 
Last edited by a moderator:
3DMark Speed Way - Average scores from day 1:


RTX 2060 - 1373
RTX 2070 - 2035
RTX 2070 S - 2076
RTX 2080 - 2203
RTX 2080 Ti - 3008

RTX 3060 - 2218
RTX 3060 Ti - 3023
RTX 3070 - 3474
RTX 3070 Ti - 3754
RTX 3080 - 4697
RTX 3080 Ti - 5413
RTX 3090 - 5480
RTX 3090 Ti - 5959
RTX 4090 - 9903

RX 6600 - 1359
RX 6600 XT - 1589
RX 6700 XT - 2190
RX 6800 - 2971
RX 6800 XT - 3613
RX 6900 XT - 3810
 

"Speed Way uses DirectX Raytracing Tier 1.1 for reflections and global illumination"

I wonder if that's why Navi 21 is substantially faster than 2080Ti, as opposed to normally being around the same performance...
 
It looks good, but not impressive. I regretted spending 4€ for this.

And it runs terrible despite DX12U efficiency features.
 
Back in the better days no-one expected new benchmark to run well on existing hardware, in fact if it did it was considered automatically outdated and useless.
Hah yes, very good point.
Every 3D mark at launch tends to run horrible at the benchmark settings they use for scores. I swear most of 3Dmark 2005 suite was like 5-10 fps on my X800 XT (except the close shot of the captain on the skyboat).
 
So I thought I'd plot the score in 3DMark Speedway against TFLOPS (from TechpowerUp's GPU database) for Ada, Ampere and Navi:

b3da049.png


Overall it does seem to be strongly correlated with TFLOPS.

For what it's worth I expect RDNA 3 to scale worse than the trendline indicates simply because register bandwidth restrictions will impact computational throughput.
 
I thought I saw a post saying those improvements were included in the reviews already.

Wasn't really bringing this up with respect to the 4090, just the overall perception of dx12 vs dx11. Maybe my memory of this thread is faulty, but I believe a lot of time has been spent discussing whether dx12 has really led to performance improvements over dx11. A large increase in base dx12 performance from new drivers would probably factor into that. Would be interested in seeing some dx11 vs dx12 comparisons in games that support both.
 
Back
Top