Ray Tracing on an AMD Graphics Card beats nvidia

[chrcoluk]

Just a quick thought, why are CPU's only double threaded, why not 3 or 4 threads?

Dual threaded per core on average provides a performance boost of about 10% when the cpu is fully loaded, its not very efficient. The affects of SMT/HTT in cinebench is far from reality from most software. Encoding is almost the only workload that gets anything decent from it.

I expect potential gains from adding 3rd or more threads to a cpu core would be near zero.

 

[Rroff]

Dual threaded per core on average provides a performance boost of about 10% when the cpu is fully loaded, its not very efficient. The affects of SMT/HTT in cinebench is far from reality from most software. Encoding is almost the only workload that gets anything decent from it.

I expect potential gains from adding 3rd or more threads to a cpu core would be near zero.

Generally there is some overhead from having it enabled - depending on CPU and config enabling HT decreases pure single thread performance around 5% but you usually offset that by increased threaded performance. I don't know but I'd assume that having the architecture try to balance the utilisation of additional threads would increase the overhead with a bigger drop in single thread performance with diminishing ability to satisfy the needs of the additional threads from the resources available.

Personally I've found HT a lot more useful than most make out with an overall increase in system smoothness with additional programs running and quite a few instances of decent performance gains.

 

[chrcoluk]

Single threaded workloads would slow down if another process tires to use the other logical core sharing the physical core but a good cpu scheduler would avoid that situation.

Under light to moderate load a extra thread on a core is almost zero impact, so is this improved system smoothness whilst you have encoding in the background or something? Also is it measured or just "feels smoother" (possible placebo).

You not limited to running programs equal to number of cores, programs dont tie down cores or anything like that.

 

[chrcoluk]

With this peformance data now released, I would say RT acceleration is underwhelming.

So e.g. in BF5, a game optimised specifically for the RT hardware, the performance on a RTX card with comparable compute performance is only around double over a non RTX card. (2080 vs 1080ti).

If you think thats decent then look at how much difference specialised hardware usually makes e.g. compare crypto performance on a AESNI cpu vs a non AESNI cpu, and you will see this is underwhelming. Seriously I was expecting the 1080 ti to be running at about 5-10fps with BF5 RT.

 

[humbug]

With this peformance data now released, I would say RT acceleration is underwhelming.

So e.g. in BF5, a game optimised specifically for the RT hardware, the performance on a RTX card with comparable compute performance is only around double over a non RTX card. (2080 vs 1080ti).

If you think thats decent then look at how much difference specialised hardware usually makes e.g. compare crypto performance on a AESNI cpu vs a non AESNI cpu, and you will see this is underwhelming. Seriously I was expecting the 1080 ti to be running at about 5-10fps with BF5 RT.

That performance, even on the 2080TI It doesn't have to be like this, why do you do this nVidia?

 

[Grim5]

Battlefield is using one of the lighter implementations by just doing reflections

Metro exodus is a good example of heavy Ray Tracing thanks to its global illumination

And here the gap between the 1080ti and 2080ti is 300%

In any case, I think it's a good thing that the 1080ti is still capable of playing games with ray tracing on, you may need to turn down the ray tracing to Low, but you can get playable framerates, so the amount of people who can experience it now in games is majorly improved

 

[D.P.]

With this peformance data now released, I would say RT acceleration is underwhelming.

So e.g. in BF5, a game optimised specifically for the RT hardware, the performance on a RTX card with comparable compute performance is only around double over a non RTX card. (2080 vs 1080ti).


If you think thats decent then look at how much difference specialised hardware usually makes e.g. compare crypto performance on a AESNI cpu vs a non AESNI cpu, and you will see this is underwhelming. Seriously I was expecting the 1080 ti to be running at about 5-10fps with BF5 RT.

reflections are the simplest aspect with less rays needed. When you start using global illumination, AO, shadowing and son on then Turing is over 6x faster than Pascal, which is exactly what you would expect.

 

[Grim5]

reflections are the simplest aspect with less rays needed. When you start using global illumination, AO, shadowing and son on then Turing is over 6x faster than Pascal, which is exactly what you would expect.

Which is why the benchmarks show a much wider margin, because they're using a strong implementation and combining multiple types, such as GI, Shadows and Reflections all in one demo.
Reflections is the easiest form of Ray Tracing to run and we know DICE has already cut down the quality of its Ray Tracing passes a couple times now to boost performance, making Ray Tracing in Battlefield V relatively easy to run now

I can run Battlefield v at 1440p Ultra settings, except Ray Tracing set to High and get 90fps average

 

[chrcoluk]

Battlefield is using one of the lighter implementations by just doing reflections

Metro exodus is a good example of heavy Ray Tracing thanks to its global illumination

And here the gap between the 1080ti and 2080ti is 300%

In any case, I think it's a good thing that the 1080ti is still capable of playing games with ray tracing on, you may need to turn down the ray tracing to Low, but you can get playable framerates, so the amount of people who can experience it now in games is majorly improved

Now that looks a bit better, I was expecting nearer a 10x improvement tho, at absolute minimum 5x advantage. I think its reasonable to assume a optimised for pascal method would double 1080ti performanse so on BF5 would be close to a 2080, and in exodus about 33% slower.

 

[Poneros]

Reflections are actually very heavy on performance, it's only because they're so severely gimped and sparsely used in BF V that they aren't as taxing as other RT implementations in other games.

See: https://youtu.be/EekCn4wed1E?t=968

 

[Grim5]

Now that looks a bit better, I was expecting nearer a 10x improvement tho, at absolute minimum 5x advantage. I think its reasonable to assume a optimised for pascal method would double 1080ti performanse so on BF5 would be close to a 2080, and in exodus about 33% slower.

Well it’s 5 times boost is you turn dlss on
But then your image is slightly softer

 

[Rroff]

Reflections are actually very heavy on performance, it's only because they're so severely gimped and sparsely used in BF V that they aren't as taxing as other RT implementations in other games.

Depending on technique (there is a way of entirely eliminating the need for one of the passes) used reflections generally only need 1 or 2 ray passes though compared to other lighting features that can need several bounce passes to produce high quality results.

 

[doody]

we need crytek ray tracing https://www.youtube.com/watch?v=vvfsM7GKIAE

 

[Gregster]

we need crytek ray tracing https://www.youtube.com/watch?v=vvfsM7GKIAE

Man, that kid gets right on my nerves. I can't stand to watch him.

 

[Double07]

Man, that kid gets right on my nerves. I can't stand to watch him.

His tests were very scientific though, honest.