Puget Systems print logo

https://www.pugetsystems.com

Read this article at https://www.pugetsystems.com/guides/1083
Matt Bach (Senior Puget Labs Technician)

A quick look at Titan V rendering performance

Written on December 12, 2017 by Matt Bach
Share:

Introduction

If you are not familiar with the new Titan V ("V" for the Volta architecture), this is the new top-end card in NVIDIA's Titan line of GPUs. With a $3,000 price tag, however, this card is most likely out of the reach of the majority of consumers. But for those with the budget, this card should give terrific performance in certain workloads.

There are four aspects to this card that makes it interesting: more CUDA cores than any GTX/Titan GPU (although they run at a lower speed), HBM2 memory, terrific FP64 performance, and the addition of Tensor cores. If you want more information on these features, we recommend reading William's NVIDIA Titan V Surprise blog post that was posted a few days ago. For GPU-based rendering, however, the FP64 performance and Tensor cores should not have any impact on performance. The HBM2 memory may help a bit, but the simple addition of more cores is what is likely to make the biggest difference.

We will have more in-depth articles in the future looking at this GPU, but we thought it would be useful to get up some preliminary results for a number of GPU-based rendering engines as soon as possible.

V-Ray

NVIDIA Titan V V-Ray Benchmark
While the V-Ray benchmark does not make use of the new "Hybrid" mode available in V-Ray RT 3.6, it is still an excellent indicator of GPU performance when using the V-Ray rendering engine. As a point of reference, in this benchmark the older Titan XP performs about 18% faster than the GTX 1080 Ti. The new Titan V, however, comes in at 60% faster! That is still only a 60% performance gain for more than 4x the cost, but in terms of single GPU performance this is extremely good.

FurryBall RT

NVIDIA Titan V FurryBall RT Benchmark
We thought a 60% performance gain over the GTX 1080 Ti was pretty good in V-Ray, but the Titan V's performance in FurryBall RT is even better. In this rendering engine, the older Titan XP comes in at about 19% faster than the GTX 1080 Ti, but the new Titan V is just shy of twice as fast as the GTX 1080 Ti! Cost aside, it is very unusual to see this much of a performance boost from new GPU. We are not sure if it is the HBM2 memory or simply the fact that it has more CUDA cores, but the Titan V definitely gives amazing performance in FurryBall RT.

OctaneRender

Unfortunately, OctaneRender does not currently work with the Titan V. It is detected as an available CUDA device, but Octane is not able to actually use it for rendering. We have reported the issue to OTOY and will update our OctaneRender workstation FAQ when it becomes supported.

Conclusion

From a performance standpoint, the new Titan V is a beast for rendering. It is not yet supported in OctaneRender, but in V-Ray we saw a 60% performance gain over a GTX 1080 Ti while FurryBall RT saw almost a 2x increase in performance.

  GTX 1080 Ti Titan XP Titan V
MSRP $699 $1,200 $2,999
CUDA Cores 3584 3840 5120
Boost Clock 1582 MHz 1582 MHz 1455 MHz
VRAM Size 11GB 12GB 12GB
VRAM Type GDDR5X GDDR5X HBM2
VRAM Bandwidth 484 GB/s 547.7 GB/s 652.8 GB/s
TDP 250 watts 250 watts 250 watts
SLI Support Yes Yes No

For most people, however, the cost of the Titan V is going to be a hard barrier to overcome. Coming in at four times the cost of a GTX 1080 Ti while "only" being at most twice as fast, most users will simply use more GTX 1080 Ti GPUs to get higher performance at a similar price point. Of course, raw price-to-performance is only one of many considerations. From a power draw standpoint, you should be able to get much higher performance at the same wattage using the Titan V over the GTX 1080 Ti or Titan XP. This is a pretty major consideration when getting into high-end rendering workstations, as a single quad GPU rendering box can easily use ~11 amps of power. This means that a standard 15 amp power circuit is simply not able to handle more than one quad GPU rendering box.

Density is another major consideration, as you could get the same performance from a much smaller footprint. This is mostly a factor for network rendering nodes, but even in a desktop workstation this lets you get the same performance from a MicroATX chassis with two Titan V GPUs as you would from a large full tower chassis with four GTX 1080 Ti GPUs. Alternatively, it allows you to use a single full tower workstation with four Titan V GPUs to get the same performance that would previously require the use of network rendering with multiple rendering nodes.

The Titan V is definitely going to be a niche product in the content creation industry. It is more focused on machine learning and high performance computing, where its price point and features make it a great value, but most people doing GPU-based rendering are going to find it hard to justify the much higher cost. From a raw performance standpoint, however, the Titan V is absolutely terrific for GPU-based rendering - assuming your rendering engine supports it, of course.

Titan V coming soon to our GPU rendering workstations:

Tags: NVIDIA, Titan, V, rendering, V-Ray, OctaneRender, Furryball
August Black

Hi there, that's a nice article!

Can you compare the test results to the quadro GP100 in some benchmarks? (Mainly V-Ray & Octane, though...)

Posted on 2017-12-13 02:32:14
Douglas Thais Rigueira

In the comparison chart, it is a Titan XP (2016) or Titan Xp (2017)?

Posted on 2017-12-13 02:59:16
veggiemuncher

Hi,

Would you guys be able to test out the NVLink bridge on your Titan Vs by removing the backplates? I know nvidia says NVLink isn't supported, but the electrical contacts seem to be fully intact on the Titan Vs, they are just blocked by the backplate.

Posted on 2017-12-14 18:58:19