Theia 2023 Release Part 2: 4090 Unboxed

Summary: This is likely the fastest and most impressive consumer grade piece of hardware we have tested. The build quality itself is great, and the results show in the analysis speed. With Theia 2023, this new hardware is supported.

I have to admit, when we first received our 4090 build, I was a little hesitant. Not only was the radiator quite a bit bigger than any I had seen, I had also read the reports on them melting wires and literally catching on fire. So, in a household with humans and dogs, this frightened me a bit, and I think reasonably so.

After reading that those occurrences were rare (and due to wire bending - though should that really cause a fire?), my fears subsided and we opened the box. And we were very impressed.

The build itself is, to say the least, impressive. The components are as follows:

1. CPU: 13900K

2. RAM: 128 GB @ 6000 MHz

3. PSU: 1600 Watt TITANIUM

4. GPU: Dual 4090

5. Dual m.2 2TB drives

Here is an inside look at the PC.

The GPUs themselves were selected like this intentionally: the bottom one has the radiator attached, and the other has the radiator separate and attached to a different portion of the case and is liquid cooled. The rationale here is that two full size cards just wouldn’t fit. And we were right; the second card was very hard to slot in the bottom slot. In hindsight, I think I would have preferred going with two liquid cooled cards if we could figure out where to mount the second radiator. I guess this is a problem we will address in the future.

Now to the build; it hums. I haven’t had a new machine in some time, but I was really surprised with how fast it booted and how quiet it was. Even when we started running workloads on it, it handled it really well, didn’t overheat, and didn’t sound like an aircraft carrier.

Benchmarks:

From a speed perspective, this is really where we were interested. The benchmarks on these cards are impressive, but they aren’t always applicable to our software. For testing, we looked at a four-second trial collected at 85 frames per second using eight cameras, and this is what we found: total analysis time was 14 seconds. In other words, we analyzed 32 seconds of video (8 videos x 4 seconds) (at 85 fps) data in 14 seconds, which is around 200 frames per second.

Here is some real-time screen capture demonstrating the performance. I promise it's not sped up!

In addition to the hardware component, Theia 2023 has better load balancing between the GPUs, and more efficient CPU computations. Though this is not our primary focus, as the main driver for our software is accuracy, having it run a bit faster is always better for the user. For that trial, the GPU load was modest at around 65%, (we do have some options to push that, and my guess is that for longer trials it would be higher) but no matter what you do, the GPU loads for two cards will be different, and they will take different amounts of time to finish. It’s a bit technical, but in this version we dynamically assign blocks of analysis to each GPU (and each GPU thread) so they are always working, and finished at the same time. Previously, the slower GPU, or longest file, would be the rate limiting step in analysis. For this analysis, the cards finish within 0.3 s despite having different speeds. Better yet, all of this is now available in the log so we are better at recording our performance.

The dual m.2 drives in this machine are super convenient. Reliable SSD with a lot of storage, that isn't prohibitively expensive. We store most of our data on remote drives, so we don’t actually require a lot of space locally, but even then, it’s rare that any developer has more than 40 GB available on their C: drive. With this setup, local storage just isn’t a problem.

The verdict? Two thumbs up! 10/10! If I had known this information beforehand, I would have purchased this machine as soon as the 4000 series GPUs were released. Quiet, reliable, and super fast, making quasi real-time analysis possible!

For most researchers, knowing that the analysis is going to work before the participant leaves is just another weight off of their shoulders. With this computer, and Theia 2023, it’s possible. If this or anything else in this blog is something that resonates with you, shoot us an email and we can provide more information.

To learn more about Theia, click here to book a demo.