The most recent foray into improving the usability and time:money ratio of my system was to try an external GPU. Thunderbolt 3 has finally removed the bottleneck of being able to shove the polygon crunching power of an external GPU down the pipe and into the motherboard/logic board and so coupling a cheap enough AkiTio Node and as expensive as you’d like to spend GPU, you can improve your image processing times by quite a significant degree. Rhetoric? Waxing poetic? Here’s the numbers;
Note: OPENCL IMPLEMENTATIONS ARE AGNOSTIC IN NATURE. I.E. IF THE SOFTWARE REQUIRES OPENCL, IT WILL UTILIZE WHATEVER CARDS ARE AVAILABLE.
Proof in the pudding
Common tasks for a Photography DIT is processing baked looks into 16 Bit TIFF’s for the client/post house. We processed 200 5DSr photographs and compare the time difference from the first and last image. We did this on the available setups we have starting with using the laptop and its built in GPU’s. Below is a rundown of the setups used.
Laptop – MacBook Pro 15″ Late 2016 – 2.9Ghz Intel Core i7 / 16GB 2133Mhz LPDDR3 / Radeon Pro 460 4096MB & Intel HD Graphics 530 1536MB
Laptop & AkiTio – Same as above, including an eVGA 1080Ti FTW3 in an AkiTio Node enclosure running through Thunderbolt 3
Rackintosh – Custom built rack mounted OS X build. 4.6Ghz Intel Core i7 / 32Gb 2666Mhz DDR4 / Zotak 980Ti Amp Extreme & eVGA 1080Ti FTW3 / nVME 2GB/s SSD
There is a strong lead with the rackintosh, but it is far from mobile and requires a more studio based setting. It is often overkill for many jobs. The laptop & external GPU performs very nicely in comparison to the laptop solo. Final processed tiff’s can often be in the thousands so this is a short window into the potential time saved. If you think this is the route for you, read on and see how easy it is to get going.
How do we get there?
Install, cable up with the sadly extremely short TB3 cable, and run the following commands in the Terminal;
curl -o ~/Desktop/metaltest.swift https://raw.githubusercontent.com/goalque/automate-eGPU/master/metaltest.swift cd ~/Desktop swiftc -sdk $(xcrun --show-sdk-path) -target x86_64-apple-macosx10.11 -o metaltest metaltest.swift ./metaltest
Note: SIP Needs to be disabled. If you do not know how to do that, the AkiTio Node might not be for you.
Reboot, and hey presto — you’re in. You can hot plug your monitors no problem. At least, we could — your mileage may vary!
The question is, is it worth it? The AkiTio Node comes in around $300, the best GPU you can buy right now besides Titan money is the 1080Ti at $800. This performance increase is definitely significant enough that there isn’t a need to upgrade the CPU or RAM on a laptop to squeeze out the last remaining life out of it. An external GPU is breathing life into these machines the same way SSD’s did 5 years ago.
In fact, if you’re creating dailies and aren’t dealing with more than a single-pass h.264 deliverable, the laptop & eGPU will actually far outperform any current Mac Pro offerings and would even surpass the 5K iMac. Interesting food for thought. Read on if you’re curious why.
Now playing around with the AkiTio has had us delving pretty deeply into the inner workings of performance benchmarking and it brought us to an interesting head-scratch with regards to a benchmark we were performing in DaVinci Resolve. The benchmarks above show just how much more powerful the Rackintosh is in comparison to the laptop. The laptop is no slouch, either. It’s the most expensive current MacBook Pro offering, and even rocks an SSD fast enough that writing to the disk wasn’t a bottleneck either. The curious problem was that the laptop was far outperforming the Rackintosh on a h.264 compression benchmark. Why?
The answer is a little CPU embedded hardware based encoding/decoding utility called Intel Quick Sync. That’s right. A feature that is standard on a lot of consumer chip offerings by Intel will improve your (single-pass only) h.264 encoding times by double, if not more. So your Core i5 might even make a Mac Pro eat shit due to the Mac Pro running professional grade Xeon chips that are inherently much more expensive, and also lacking… Intel Quick Sync. So if you wonder why so many Video Content Creator’s love the iMac 5k, it’s because it actually IS faster for their application than a Mac Pro.
I’ll be following up about Intel Quick Sync and benchmarking the latest Apple offerings further down the line. It’s an interesting gotcha that many overlook.