DirectX 12Futuremark 3DMark Time Spy Benchmarks Thread

[SPBHM]

Ok, got Time Spy working on my brother's PC, had to revert to driver 16.3.2 to do it, possibly something to do with AMD using SSE4, which the Q6600 doesn't support, in DirectX 12 under drivers newer than 16.4.

yes I saw other people reporting this problem,
basically on anything newer than 16.4.2 CPUs without POPCNT (Intel before Nehalem, and AMD before Phenom 1) it doesn't work with DX12, it's a shame specially for the faster C2Qs, like a 4GHz q9550 or Xeon 771 with the mod

 

[Termie]

GTX 1070 SLI @ Reference Clocks with one standard SLI Bridge:

Overall: 8456
Graphics: 9658 (74% higher than a single card)

GTX 1070 SLI @ Reference Clocks with two standard SLI Bridges:

Overall: 8918
Graphics: 10265 (85% higher than a single card)

In other words, looks like running two SLI bridges is making a big difference. This is on my 6700K system, will test my 6900K system soon.

 

[AdamK47]

Jeez...I did something stupid. I loaded up my benchmark settings and ran it at 4.8GHz a couple of times and wondered why my scores had dropped to around ~10000, finally realized I had G-Sync on and it was set to 60Hz. I'll get back around to a 4.8GHz bench later, for now here is my daily 4.5Ghz.

http://www.3dmark.com/spy/50984

That's pretty good. Let us know how the suicide runs go.

 

[Brunnis]

http://www.3dmark.com/spy/69861

Overall: 4100
Graphics: 3953
CPU: 5202

6700K and R9 390 (all stock). See rest of system in sig. Seems to be about where I'd expect it to be.

 

[Azix]

can some of you guys with pascal cards report CPU utilization during the GPU tests? To be compared with AMD cards. The suggestion is that nvidia is doing their async solution partly through software, so one would expect higher CPU utilization during those GPU tests.

should be easy for those with monitoring enabled in 3dmark.

 

[Silverforce11]

can some of you guys with pascal cards report CPU utilization during the GPU tests? To be compared with AMD cards. The suggestion is that nvidia is doing their async solution partly through software, so one would expect higher CPU utilization during those GPU tests.

should be easy for those with monitoring enabled in 3dmark.

https://www.reddit.com/r/nvidia/comments/4t576c/is_nvidia_using_cpu_for_async_compute/

They get lower CPU scores when Async is On.

I don't know where this benchmark is from though, but if it's true:

Software based scheduler, gets busy with Async Compute queues, sorting it via the "Dynamic Load Balancer" on their drivers.

 

[Kenmitch]

can some of you guys with pascal cards report CPU utilization during the GPU tests? To be compared with AMD cards. The suggestion is that nvidia is doing their async solution partly through software, so one would expect higher CPU utilization during those GPU tests.

should be easy for those with monitoring enabled in 3dmark.

I'll try it later tonight and post back. Currently still at work.

 

[Termie]

https://www.reddit.com/r/nvidia/comments/4t576c/is_nvidia_using_cpu_for_async_compute/

They get lower CPU scores when Async is On.

I don't know where this benchmark is from though, but if it's true:

Software based scheduler, gets busy with Async Compute queues, sorting it via the "Dynamic Load Balancer" on their drivers.

These results can't be interpreted like this. The cpu score is derived from a test that uses no GPU power. Basically this is just a result of test to test variation.

 

[Termie]

Just tested async vs async disabled on a single GTX 1070.

Async enabled:
Graphics Test #2: 31.57fps
Maximum CPU Usage: 37%

Async disabled:
Graphics Test #2: 29.69fps
Maximum CPU Usage: 40%

First peak is with Async, second peak is without. GPU is clearly utilized less with Async off.

 

[Kenmitch]

can some of you guys with pascal cards report CPU utilization during the GPU tests? To be compared with AMD cards. The suggestion is that nvidia is doing their async solution partly through software, so one would expect higher CPU utilization during those GPU tests.

should be easy for those with monitoring enabled in 3dmark.

@Termie. Looks like your results match mine as far as cpu usage goes.

I just used the custom setting and didn't test cpu as I was just looking for the data on the cpu usage with and without Async support.

Here is with it enabled.

And this is with it disabled.

HWINFO was reset just before the run was started. Image is from directly after the benchmark finished. Guess you can compare the max and average cpu usages to get an idea.

Zotac GTX 1070 Amp max boost 2025 memory 9558

 

[Silverforce11]

Just tested async vs async disabled on a single GTX 1070.

Async enabled:
Graphics Test #2: 31.57fps
Maximum CPU Usage: 37%

Async disabled:
Graphics Test #2: 29.69fps
Maximum CPU Usage: 40%

First peak is with Async, second peak is without. GPU is clearly utilized less with Async off.

Looks to be within margin of error. ie, no difference.

 

[Termie]

I just used the custom setting and didn't test cpu as I was just looking for the data on the cpu usage with and without Async support.

Here is with it enabled.

And this is with it disabled.

HWINFO was reset just before the run was started. Image is from directly after the benchmark finished. Guess you can compare the max and average cpu usages to get an idea.

Zotac GTX 1070 Amp max boost 2025 memory 9558

This is exactly in line with what I found. CPU Usage is higher with Async disabled.

I'd say the theory that Nvidia is offloading Async to the CPU has been debunked.

 

[dzoni2k2]

Don't even know how you would offload anything to the CPU. It would be ridiculously slow.

It just means there is no measurable impact on the CPU for whatever driver sorcery Nvidia is doing.

 

[Erenhardt]

Just tested async vs async disabled on a single GTX 1070.

First peak is with Async, second peak is without. GPU is clearly utilized less with Async off.

Actually, not. I edited your screengrab:

See how in the right yellow box the cpu is slightly below and later little bit above the red line of the box (used boxing tool to have a perfectly vertical line), and on the left the values are higher, and do not go below the red line?

@Termie. Looks like your results match mine as far as cpu usage goes.

I just used the custom setting and didn't test cpu as I was just looking for the data on the cpu usage with and without Async support.

Here is with it enabled.

And this is with it disabled.

HWINFO was reset just before the run was started. Image is from directly after the benchmark finished. Guess you can compare the max and average cpu usages to get an idea.

Zotac GTX 1070 Amp max boost 2025 memory 9558

Not really. There is a big flaw or two in your methodology.
First thing disable CPU downclocking. When it is running at 800mhz each second screws the average outcome. Second thing, you will not be resetting each time in the same moment. Just look at the differences between average core clocks. In your second run each core runs 100mhz slower than in the first run.

It really is not that simple.

 

[Hitman928]

Thanks, looks like the benefit of async on the 1070 drops to about 3.5% at 4k. I'll do the same on my 290 tomorrow to compare.

Looks like on the 290 the async benefit goes from about 15 percent at 1440p to about 7.5 percent at 4k. So it is still getting more benefit but the effect of the async benefit getting cut in about half is the same.

 

[3DVagabond]

Looks like on the 290 the async benefit goes from about 15 percent at 1440p to about 7.5 percent at 4k. So it is still getting more benefit but the effect of the async benefit getting cut in about half is the same.

I don't understand why async compute would be resolution dependent at all?

 

[Flapdrol1337]

I don't understand why async compute would be resolution dependent at all?

My guess is it's because games and hardware are not designed for extreme resolution.

If you increase one part of the rendering pipeline (here the resolution) the other tasks become more insignificant. The performance would be more determined by the finite capacity of the gpu to calculate all those pixels. Doing the other tasks asynchronously still gives you a benefit, but a smaller benefit since the gpu wouldn't spend much time on those tasks anyway.

 

[Kenmitch]

Not really. There is a big flaw or two in your methodology.
First thing disable CPU downclocking. When it is running at 800mhz each second screws the average outcome. Second thing, you will not be resetting each time in the same moment. Just look at the differences between average core clocks. In your second run each core runs 100mhz slower than in the first run.

It really is not that simple.

Used a different method this AM to take another look.

I'd say overall the cpu usage is greater with Async On vs Off at this time. Not a huge difference but is visibly greater.

 

[Raduque]

http://www.3dmark.com/spy/74276

I'm using an older driver so it's upset about that.

 

[dogen1]

I don't understand why async compute would be resolution dependent at all?

Because bottlenecks shift when you change resolution.
Lower resolutions stress geometry processors relatively more, leaving more opportunity for performance improvement from async compute.

 

[Termie]

Used a different method this AM to take another look.

I'd say overall the cpu usage is greater with Async On vs Off at this time. Not a huge difference but is visibly greater.

It's a bit tricky to come to this conclusion, as the GPU is also working harder as I found in my graph, which pushes the whole system harder. But given that this is a GPU-intensive benchmark, and really isn't designed to tax the CPU, perhaps there is something unusual going on. Would have to see identical graphs on an AMD GPU to know...

Maybe I'll run Time Spy on my 390X later today.

 

[Deders]

It would make sense that if the CPU was helping thread the gpu better, then it would be able to do more work concurrently.

 

[Azix]

Don't even know how you would offload anything to the CPU. It would be ridiculously slow.

It just means there is no measurable impact on the CPU for whatever driver sorcery Nvidia is doing.

not offloading, just telling the GPU how to handle different async tasks. IIRC nvidia does shader substitutions in dx11, something like that. what it is doing might not be that demanding actually.

Why I was wondering about CPU usage is because actual games would load the CPU and could mess up the benefits for pascal if the driver is doing some heavy lifting getting the GPU to perform the dynamic load balancing.

But maybe the load balancing occurs in hardware

 

[Yakk]

I was going to try TimeSPY but then I started reading...

The more things are evolving over at Overclock.net, the more TimeSpy looks to be really just a nvidia benchmark which is disappointing. Steam forums are also a mess now.

Looks like TimeSpy is not really working Asynchronously.

http://www.overclock.net/t/1605674/computerbase-de-doom-vulkan-benchmarked/210#post_25351958

All of the current games supporting Asynchronous Compute make use of parallel execution of compute and graphics tasks. 3D Mark Time Fly support concurrent. It is not the same Asynchronous Compute.

Concurrency fills in gaps which are in the execution pipeline. Parallelism executes two tasks at the same time.


Notice the context switch involved?

If 3D Mark Time fly were using Parallel executions then there would be synchronization points between the two contexts (Graphics and Compute). There would also be pipeline stalls on Maxwell GPUs. Both the pipeline stalls and the flush required for a synchronization point would add latency thus leading to Maxwell losing performance when running this variant of Asynchronous compute. We do not see Maxwell losing performance under 3D Mark Time Spy. We see a tiny performance boost. Thus 3D Mark Time Spy is not running Asynchronous Compute + graphics. You see parallel executions = Asynchronous Compute + Graphics. Concurrent execution = Asynchronous Compute. They are not the same thing.

Mahigan has been correct in the past on other issues, very curious to see how 3DMark will be responding to all this.

 

[Thala]

Lower resolutions stress geometry processors relatively more, leaving more opportunity for performance improvement from async compute.

While i agree with the first statement, the conclusion would be, that the opportunity for async compute would increase with resolution.