Nvidia Turing Product Reviews and Previews: (Super, TI, 2080, 2070, 2060, 1660, etc)

Apparently, RTX Voice does not require tensor cores to function: https://www.computerbase.de/2020-04/nvidia-rtx-voice-ohne-rtx/

 

Initially I thought this RTX voice was a gimmick, but the GTX 1060 audio example is pretty telling.

 

Benchmarks: Premiere Pro with NVENC: Rendering videos in 20% of the time (
May 26, 2020

https://www.hardwareluxx.de/index.php/news/software/anwendungprogramme/53246-premiere-pro-mit-nvenc-videos-rendern-in-20-der-zeit.html

 

Premiere Pro 14.2 H.264/H.265 Hardware Encoding Performance
Written on June 4, 2020 by Matt Bach

In the 14.2 version of Premiere Pro, Adobe has added support for GPU-based H264/H.265 (HEVC) hardware encoding with both NVIDIA and AMD GPUs. How well does this feature work, and how much faster is it than the previous hardware encoding that utilized Intel Quicksync?

https://www.pugetsystems.com/labs/a...264-H-265-Hardware-Encoding-Performance-1778/

 

That only took them a decade or so.

 

Premiere Pro 14.2 GPU Roundup - NVIDIA GeForce SUPER vs AMD Radeon
Written on June 18, 2020 by Matt Bach

https://www.pugetsystems.com/labs/a...up---NVIDIA-GeForce-SUPER-vs-AMD-Radeon-1802/

 

Dual NVIDIA Quadro RTX 8000 Review with NVLink Performance
July 6, 2020

https://www.servethehome.com/dual-nvidia-quadro-rtx-8000-review-with-nvlink-performance/

 

Just in time for h.266 to enter the chat!

 

Does anyone know what Nvidia did with all the TU102's that couldn't cut it as a 2080 Ti? According to TPU all SKUs had at least 68 (94%) of the SMs enabled.

What did they do with all the others?

https://www.techpowerup.com/gpu-specs/nvidia-tu102.g813

 

When Turing appeared, its 12nm process tech was already in a very comfortable spot in the yield curve. Partly due to it's very close relationship to TSMCs 16nm (some say it's virtually identical, so 16nm++). With a salvage part like the 2080 Ti was, you could already afford up to 5 crippling errors in the SMs and Memory Controllers - maybe there were not so many fully dysfunctional dies after all.

 

Maybe. That would be pretty amazing for such a large chip even on a mature process.

 

Yeah. And maybe, knowing how large the chip would end up anyway, Nvidia did some more finegrained redundancy as well in less replicated areas of the chip.

 

18-Way NVIDIA GPU Performance With Blender 2.90 Using OptiX + CUDA
September 6, 2020

https://phoronix.com/scan.php?page=news_item&px=Blender-2.90-18-NVIDIA-GPUs

 

Some of the performance deltas there just don't make sense. 2070 vs. 2070S? Almost no improvement for 2080Ti over 2080S?

 

I think this has something to do with TU104 having 8SM's in a GPC compared to TU102 and TU106 having 12SM's in a GPC.

There are some TU104 based RTX 2060's out there which outperformed the regular TU106 variants in workstation tasks, but showed no difference in game benchmarks.

 

You're right, perf seems grouped by # of GPCs with Cuda, while in Optix, RT cores y/n seem to play a more dominant role, but yet 3-GPC 1660S is at the level of 4-GPC 1070/1080.
That's quite strange, given that cycles is a pathtracer.

 

A factor here could be that some GPUs (eg. 2080ti) are actually power limited in this usage scenario. Would be interesting to see power usage, clock rates, and utilization measurements during this tests and also in comparison to a gaming work load. The 2080ti for instance has much more hardware resources compared to the 2080 Super relative to the actual power available between the two (almost the same).

As an aside I have problem with how power consumption is tested (including for CPUs) which also influences how people look at power consumption. With how it's mostly currently done what you're really testing and showing is just the behavior of the GPU (or CPUs) power limiter. Which further just creates this illusion that all work loads utilize the same amount of power and behavior.

 

I checked a 2080 Ti FE under Windows 10 in the classroom scene. It averages at just shy of 200 Watts in Cuda and around 175 Watts via Optix, consistently boostig to north of 1900 MHz. Power budget should not be the main culprit here.

 

GPC and cash had some major improvments between 1660s and 1070. If you look at this list, it looks havely frontend bound.

Thats why also the 2080ti is in front of the Titan because the clockspeed is higher.

 

Not so sure about that. At least compared to the Founder's Edition (1635), the Titan RTX clocks higher (1770) on paper. And it has not only more ALUs, but more control logic for them too (72 vs. 68 SMs). Maybe though those were not