New high-performance GPUs now available on RenderStreet

At the beginning of 2019, we were announcing an upgrade in our GPU line-up. The new GPUs at the time (Nvidia® Tesla K80) came with almost double speed compared to the old ones, and three times more memory.

Now it’s time for a new upgrade. We are jumping forward three hardware architecture generations, from Kepler to Volta, and we’re powering new machines with Nvidia® V100 Tensor Core GPUs. The new GPUs are, as described by Nvidia, “The Most Advanced Data Center GPU Ever Built” and bring even more speed and 16 GB of RAM for your renders. As with the previous upgrade, there is no increase in the price you are paying for our services – this is a free performance upgrade. The new machines are live at this moment (in fact, they have been live for a while now, introduced gradually in production). And we’ve run the numbers for you, to see exactly how much better they are.

We started with the standard battery of tests from the official Blender benchmark. As those scenes are quite old now, we added a couple of new ones as well: the shots from the latest Blender movies: Agent 327 (Barbershop) and Spring. Check out the render times with Blender 2.82a below, with the new GPUs being shown in green.

1. BMW27, 960 x 540 px, 1,225 Samples (35 squared)

RenderStreet Nvidia V100 Benchmark 1 – BMW. Render time in h:mm:ss, lower is better . Blender 2.82a — Benchmark 1 – BMW. Render time in h:mm:ss, lower is better . Blender 2.82a

The simplest scene in this batch, the BMW, shows a 25% speed increase with the V100 GPUs. The render time gets close to half a minute with the standard settings now.

2. Classroom, 1,920 x 1,080 px, 300 Samples

RenderStreet Nvidia V100 Benchmark 2 – Classroom. Render time in h:mm:ss, lower is better. Blender 2.82a — Benchmark 2 – Classroom. Render time in h:mm:ss, lower is better. Blender 2.82a

The same kind of speed-up (25%) can be observed in the Classroom scene as well. The new V100 machines bring the render time under the two-minute mark for this scene.

3. Fishy Cat, 1,002 x 460 px, 1,000 Samples

RenderStreet Nvidia V100 Benchmark 3 – Fishy Cat. Render time in h:mm:ss, lower is better . Blender 2.82a — Benchmark 3 – Fishy Cat. Render time in h:mm:ss, lower is better . Blender 2.82a

The Fishy Cat benchmark shows a 27% improvement with the new GPU configuration. The results are quite consistent so far, with all the scenes showing a similar decrease in the render times.

4. Koro, 720 x 1,280 px, 500 Samples

RenderStreet Nvidia V100 Benchmark 4 – Koro. Render time in h:mm:ss, lower is better . Blender 2.82a — Benchmark 4 – Koro. Render time in h:mm:ss, lower is better . Blender 2.82a

The cute llama sports a speed increase in the same range as the other three benchmarks: 23% decrease in render time with the new V100 boards. It’s interesting to see that the fur-heavy scenes behave the same way as scenes that do not contain hair. Traditionally, hair rendering has been a tough nut to crack, and it’s nice to see consistent results now.

5. Pabellon Barcelona, 1,280 x 720 px, 1,000 Samples

RenderStreet Nvidia V100 Benchmark 5 – Pabellon Barcelona. Render time in h:mm:ss, lower is better . Blender 2.82a — Benchmark 5 – Pabellon Barcelona. Render time in h:mm:ss, lower is better . Blender 2.82a

The Pavillion Barcelone scene only benefits from a 20% increase in speed. The new render time is under three minutes, quite fast for a complex project.

6. Victor, 2,048 x 858 px, 600 Samples

RenderStreet Nvidia V100 Benchmark 6 – Victor. Render time in h:mm:ss, lower is better . Blender 2.82a — Benchmark 6 – Victor. Render time in h:mm:ss, lower is better . Blender 2.82a

The most complex and the most memory-hungry scene in the original benchmark set receives a 19% increase in speed. While no further than 14 months ago this scene couldn’t render on our GPUs because of the memory requirements, now it can be completed on these new machines in just a hair over 5 minutes.

7. Barbershop, 2,048 x 858, 32 AA samples

RenderStreet Nvidia V100 Benchmark 7 – Barbershop. Render time in h:mm:ss, lower is better . Blender 2.82a — Benchmark 7 – Barbershop. Render time in h:mm:ss, lower is better . Blender 2.82a

The first one of the new scenes used for testing the GPU upgrade is the Barbershop one from Agent 327. It’s also the one with the most interesting result: the new GPUs are a bit slower than the old ones. My educated guess is that the anomaly is caused by the way the rendering is optimized internally for this scene. There are probably quite a few tiles that are taking a long time to complete. And this may allow Blender to better parallelize the processing in the previous setup with 4 GPUs. Hopefully, this is something that can get addressed in future optimizations that are coming for Cycles, as most people only have a single GPU in their machines.

8. Spring, 2,004 x 1,000, 300 samples

RenderStreet Nvidia V100 Benchmark 8 – Spring. Render time in h:mm:ss, lower is better . Blender 2.82a — Benchmark 8 – Spring. Render time in h:mm:ss, lower is better . Blender 2.82a

The last benchmark is a scene from the most recent movie released by the Blender Institute – Spring. This scene shows the most significant speed improvement as well: 36%. The bigger than average speed increase may be caused by the fact that Spring is the only scene from this set which has been created using Blender 2.80. It may also be a matter of how the scene was optimized by the Blender team – or a combination of the two. The fact remains though, that this render benefits the most from the hardware upgrade.

I know that you are always looking for the best possible solution for rendering your projects. This is why we are constantly working on providing it – from our periodic hardware performance upgrades to the unique monthly plan RenderStreet One and to the constant software updates to our platform. We are working as we speak on a couple more surprises that are to be released soon, which will make your life even easier. Happy rendering, keep close and be safe!