Nvidia Ampere Discussion [2020-05-14]

Is it still limited by FP32 calculations ?

 

I would guess that it's shading performance, and low utilization of shading cores. Even though Ampere doubled L1 cache, spatial locality during ray-tracing is probably so low that you stall cores. The RT cores are probably just fast enough that they don't bottleneck and making them a lot faster wouldn't bring appreciable results. There are all kinds of strategies for ray-sorting, and there will be DXR 1.1 games coming, so software developers may find ways to lessen the extent of those issues.

 

There's really no way for them to be on the same metric, if they were even RTX 2080S would be over 2,6x faster than XSX in RT, AMD/Sony/MS wouldn't have bothered with RT if it was that slow.

 

There are raytracing heavy games, like BFV and TOMBR, and the panelty for raytracing is very real.
From early reviews, it looks not like 3080, has shifted the balance for raytracing to be much less of a panelty.

I did notice nothing new yet in DX12 that would require games to be rewritten to better take advantage of new hardware features.
I would hope the hardware scheduler is able to better use GPU resources.

 

RT workload, not overall. AMD had to put RTA units into the chip because the market is going this way. And console games are getting a even lower quality RT implementation than PC games. Have you seen DMC5:SE? With RT it is 1080p/60FPS on a PS5. You can play Battlefield 5 on a 1660TI with nearly 60FPS with RT on Rotterdam...

 

There is a synthetic benchmark for that, 3Dmark port royal:
https://www.overclock3d.net/reviews/gpu_displays/asus_rtx_3080_oc_tuf_review/17

 

Or minecraft rtx or quake2 rtx or blender to really measure if ray tracing has improved in ampere over turing.

To use ampere optimally scheduling of compute and dxr jobs(rays) in game engine might have to change. Ampere is better at running those things parallel.

 

Just like Vega then.

It's less about what it's limited by and more about the additional calculations which are required. Once you trace a ray (which is done on RT core) you need to incorporate this data into rendering somehow which usually means more and more precise / detailed shading.
Basically, RT leads to more shading needed which in turn leads to slower overall performance with RT no matter how fast you perform the step of actually tracing rays. So even with a 30 TF GPU you will still have performance loss with RT over rendering the same scene without RT - and this is unlikely to change, ever. I have no idea why people were/are expecting RT to become "free" on new h/w.

Well, that's what they've said during Ampere announcement I think.

 

The modes for raytracing on the consoles for DMC5:SE with RT are 4k 30FPS and 1080P 60FPS. Whilst I think to a degree you are correct and the performance of the consoles wrt wont be that great, I do not think it will be as bad as you are making it out to be.

 

Rendering resolutions, yeah. RT seem be quarter that though in GT7 and what seems to be even lower than that in DMC5SE. The BVHs are also heavily simplified, with dynamic geometry being absent from RT reflections in Spiderman, for example. So it's a bit early to compare this directly to PC implementations. We'll have see how PC RDNA2 will fare here on the same code.

 

Leaked review of the RTX 3090:

https://videocardz.com/newz/nvidia-geforce-rtx-3090-gaming-performance-review-leaks-out

It's faster than the RTX 3080 by 4.7 to 11.5%:





Realistically, if we only look at the games that are in the 50-70FPS range (i.e. the ones that probably aren't CPU limite), it seems to be around 10% faster.

I'm not sure that's worth twice the price. Although it doesn't seem like there will be many RTX 3080 available at MSRP anyway.

 

3090 has ~12-20 % more resources, no wonder it cannot distance itself more with power limit <10 % higher.

3080 is hardly overclockable. On the other hand power limited to 250 Watt reduces performance only by several percent.
This would explain why is there hardly any improvement in power efficiency despite major node transition- Nvidia went full RX 590.

 

I suppose we can rule out bandwidth limitations for Ampere.

 

No benchmarks for professional apps yet?

 

https://wccftech.com/nvidia-geforce-rtx-3090-teclab-review-leaked/

 

3080 has some ML features disabled, at least that's what I read on dev forums. For some people it's worth it.

 

If these numbers stick I’m abandoning my quest for a 3090.

 

Nvidias product segmentation this time round seems really peculiar.

Why have a 3080 with 10gb memory rather than 12?

Spec wise the 3090 is 20% more powerful than the 3080, closer than any other generation. Performance wise this seems to translate to 10% in games. No room to introduce a 3080ti many are speculating will come.

Wouldn't a rumoured 3070ti 16gb card undermine the marketing potential of the 3080 10gb?

Don't rumours of cards with more memory(3080 20gb) impact sales potential now? Why are aibs leaking this information.

This is the least coherent launch from NVIDIA i have seen in multiple generations

 

If what we’re seeing is a front-end geometry processing bottleneck even at 4K then it seems Nvidia hasn’t struck the right balance with Ampere.