DX12, DX12U, DXR, API bias, and API evolution ..

In the future i think we will need an api for cpu + gpu (apu)'s, bc they enjoy a lot of benefits like no need to do deep copies and low cpu to gpu communication, the things a discreate gpu could never take advantage of.
Don't the console APIs and Apple's Metal already make use of an APU/SoC's advantages? And on the PC side, I doubt developers would be interesting in using a new API just to squeeze more performance out of APUs.
 
In the future i think we will need an api for cpu + gpu (apu)'s, bc they enjoy a lot of benefits like no need to do deep copies and low cpu to gpu communication, the things a discreate gpu could never take advantage of.

Why would this need a new API? There are already functions in DX12 that allow discrete GPU based systems to behave a little like APU's (the ability to read and write to both memory pools for example). Having a separate API just for APUs would be totally unusable.
 
Why would this need a new API? There are already functions in DX12 that allow discrete GPU based systems to behave a little like APU's (the ability to read and write to both memory pools for example). Having a separate API just for APUs would be totally unusable.
sorry I am but a humble compiler dev, i don't know the ins and outs of gpu api's except for some opengl stuff that I learned back in collage as a elective 10 years ago.
 
Any details yet on what new tech they added to this benchmark?

The 7900 XTX went from 44% faster than the 6950XT in Time Spy to 61% faster in Nomad. That's quite a jump.
 
In the new UL benchmark Normad Vulkan is up to 11% faster on nVidia than DX12: https://www.computerbase.de/2024-05..._directx_12_vs_vulkan_mit_geforce_radeon__arc
The gap between 4090 and 7900XTX increases from +37% to +50%...
PCGH presented the results with the best scores. Here are the 4K max results:

4090: 103 fps
4080S: 72 fps
7900XTX: 66 fps
3090Ti: 60 fps
6950XT: 43 fps
2070S: 24 fps
5700XT: 21 FPS


 
Wonder what that's stressing, first thought was compute but it could be triangle performance as it does seem to roughly match the GPC/SE distributions. Drivers maybe, wouldn't be the first time we've seen signifncant underperformance fixed after a few driver updates
 
I wonder if the new synthetic benchmark does anything interesting in terms of deferred texturing/visibility buffer or GPU driven rendering/bindless as it doesn't feature virtual geometry ...
 
I wonder if the new synthetic benchmark does anything interesting in terms of deferred texturing/visibility buffer or GPU driven rendering/bindless as it doesn't feature virtual geometry ...

That would be nice but I would be very surprised. 3dmark has been playing catch up for a long time now and it’s tech isn’t up to par with the best engines. I’m just hoping this one at least looks good.
 
It's out now and I gave it a run. It's about 5% faster in Vulkan vs DX12 on my 3090. Explorer mode is a nice addition.

It looks like a solid UE4 implementation but doesn't really push the envelope or do anything we haven't seen before. Lots of screen space stuff that defined last generation. Environment detail is pretty good. I don't understand why UL is unable to generate more impressive graphics in these very constrained demos when full fledged games can do better. It's like they're 4-5 years behind the cutting edge which kind of makes their benchmarks pointless. 3dmark had its heyday when it was 4-5 years ahead of the best games could offer.

A few things immediately stood out to me.

- No contact hardening or soft shadows
- Stiff cloth physics
- Unconvincing materials, especially on the armor and clothes of the soldiers
 
Yeah I meant UE4's feature set. 3dmark is running on its own custom engine as usual.
I wonder how much of the feature deficiency is because they wanted a "singular" benchmark that can run on desktop and mobile (making the results a bit more comparable).
 
I wonder how much of the feature deficiency is because they wanted a "singular" benchmark that can run on desktop and mobile (making the results a bit more comparable).

Don’t really need special hardware for better material shading or soft shadows though. I think they could’ve made the benchmark a bit heavier especially since it’s so constrained in scope. A 4060 Ti already gets over 30fps so in a generation or two mainstream cards will breeze through it.
 
Don’t really need special hardware for better material shading or soft shadows though. I think they could’ve made the benchmark a bit heavier especially since it’s so constrained in scope. A 4060 Ti already gets over 30fps so in a generation or two mainstream cards will breeze through it.

This is their more general benchmark, Speedway is still their premier one
 
Back
Top