Blackmagic eGPU: Additional graphics with Apple’s blessing

[vc_row][vc_column][vc_column_text]The Blackmagic eGPU is an integrated solution, unlike the eGPU housings offered to date, in which a normal graphics card is inserted into the slot of a computer. It is housed in a light metal case that can pass for “space grey” to some extent and looks good next to a Mac, even if its octagonal design fits in even better with the design language of an Alienware PC. It is surprisingly large, even slightly taller than a ‘current’ MacPro. Even the ventilation grille at the top and the solid base, on which you can easily carry the device, are made of metal. It is so robust that it is easy to imagine using it in the field, even if the elegant finish could suffer at some point. The powerful power supply unit is located in the housing.

There are good reasons for its size: after all, up to 315 watts are consumed and the whole thing is full of cooling fins and a large, slowly rotating fan in addition to the electronics. Apparently Apple’s engineers at Blackmagic could learn a thing or two about good cooling: even when rendering for hours on end, the device remained completely silent, which is not the case with the iMac or the MacBook Pro. There were also no rendering errors due to overheated graphics cards, as many MacPro users experienced with the D700. The included TB-3 cable is only 50 cm long, so there is not much freedom in terms of positioning. This does not necessarily have anything to do with stinginess: The high data rates of 40 Gigabit / second could be limited by considerably longer cables. A laptop can be charged with up to 85 watts via the TB-3 port, which is enough even for the 15-series MacBook Pro.

The device is “plug-and-play”, no drivers need to be installed or configurations set. Only High Sierra is provided as a system and the type of graphics chip is only recognised correctly from 10.13.6 onwards. Registration takes about 20 seconds and is automatic. An inconspicuous icon in the menu bar allows you to log off the eGPU again – simply pulling it out is not a good idea. Logging out takes approx. 10
Seconds.

Before disconnecting the TB cable, you should make sure to log off all memory devices connected to the eGPU, as these are not deactivated together with the graphics card. This can be up to 5 devices, as the eGPU provides four USB 3.1 ports including charging current and a second TB-3 port.

This is a first for Apple and Intel, as no eGPU with a second port has previously received eGFX certification. There is also a connection for HMDI 2.0 so that monitors with a high resolution and frame rate can also be connected directly.

The device has a dual UBS-C controller with the Mac-compatible TI83 chip. Devices with only one controller are less able to distribute the load between I/O traffic and graphics and tend to be unstable under high load. In addition, the eGPU is the first to use the new “Titan Ridge” TB-3 controller (Intel JHL7540), which supports the internal routing of display port signals from the GPU to a TB-3 output (predecessors always used the host computer). Accordingly, Apple advertises the sophisticated LG UltraFine 5K, which can be connected directly to the second port, as the ideal combination. It has the same technical values as the 5K Retina display and is presumably also based on the same panel – Apple itself no longer has pure monitors in its range. Incidentally, we did not disassemble the device ourselves and unscrew the 61 screws, but based this information on the photos of the innards at egpu.io.

On the right, an active SSD RAID reduced the throughput to the host, but the GPU still had enough.

General performance

With so many connections, the question arises as to whether the graphics throughput to the host is not slowed down by the peripherals. After all, graphics cards in the computer with 16x PCIe slots have significantly higher throughput rates available than with TB-3 with around 4x. We therefore tested whether the data throughput of a fast SSD array in the second port reduces the performance of the graphics. We ran the speed test from Blackmagic as a constant load and then measured the throughput with CL!ing from Bart Vanhaeren (note the spelling when googling). The proportion for the graphics is actually reduced by about 25%, but the performance of the GPU hardly decreased. Now this is an upper mid-range AMD chip, top models such as a Vega64 could possibly be slightly starved. It certainly doesn’t make sense to occupy all USB ports with fast drives. In our render tests with Resolve (see below), we had an SSD RAID on the USB-C port and also a USB 3.1 hard drive on the eGPU so as not to slow down the read/write processes.

An LG UltraFine 5K, which can even be operated with 10-bit, was not available to us. However, we tested an LG with UHD both via USB-C adapter and HDMI. It was no problem to use the full resolution at 60 Hz and according to the measurement with CL!ling, the monitor only consumes 3-5 per cent of the total power thanks to the support of Titan Ridge. You could therefore run two of the 5K monitors on an iMac Retina 5K and then have a total of 15,360 pixels horizontally – and a pretty full desk.

Performance under DaVinci Resolve

There is a good reason why the eGPU comes from Blackmagic: DaVinci Resolve (DR for short) utilises multiple GPUs in the computer extremely efficiently. Since no current Apple computer is available in which additional or better graphics cards can be installed, an eGPU is the only solution. Two particularly suitable candidates: A MacBook Pro without fast graphics or an iMac whose performance you want to expand. The expensive iMac Pro is available with high-performance graphics out of the box, but there is currently no MacBook Pro with more than 4 GB of VRAM. With DR, the usable VRAM always depends on the smaller card, so you would be wasting potential in terms of working resolution. The most suitable iMac is the model with the same GPU, as two identical GPUs run optimally under DR.
Resolve needs a restart to recognise the second GPU, and the first time you have to check whether both have been activated. When logging off the eGPU with the programme running, Resolve saves, exits and starts independently with only one GPU. That’s hardware and software from a single source! It would be nice if the eGPU could be registered again with a menu command. Instead, you have to disconnect it (log off all connected memory beforehand!) and restart it.

eGPU with the iMac

For the test, we put together a timeline in UHD with ProRes clips and, in addition to grading with several nodes, distributed a few computationally intensive filters such as temporal noise filtering, film grain and aperture diffraction (diffraction simulation) to some clips, in our opinion a relatively practical scenario. This resulted in a speed advantage of over 80 per cent in the final rendering on the iMac with an active eGPU. Due to the ‘light’ codec, the CPUs were only utilised between 50 and 75% (and the computer was still quite quiet). The source and target drives were separate, so the GPU performance was the main factor here. These are optimal conditions because the performance of all components is well balanced.

It also ran quite well with UHD clips in H.264 or H.265, but only as long as they were decoded by the hardware: Then the advantage with eGPU was still over 50%. With 8-bit and 4:2:0, these formats still run smoothly even at 60 fps in UHD. As soon as pure software decoding has to take over, as with 10-bit 4:2:2 (e.g. from the GH5), all CPU cores have to struggle and the two GPUs ‘starve’. Then the iMac only manages 24 or 25 fps with difficulty, even without any filters, and the eGPU becomes useless. Even with compressed DNG material from an Ursa Mini Pro, the CPUs were almost fully utilised, and a single GPU was already running at 50% during pure playback. Less than 30 fps also ran smoothly, 50 or even 60 were not possible with the iMac. With grades and filters, the eGPU only achieved an advantage of 30%.[/vc_column_text][/vc_column][/vc_row][vc_row][vc_column][vc_gallery type=”image_grid” images=”68169,68170,68171″ img_size=”200×200″][/vc_column][/vc_row][vc_row][vc_column][vc_column_text](Image1: This is what the configuration for the iMac with eGPU can look like in Resolve. Image2: If the CPUs are running at the limit for decompression, one GPU is also underutilised. Image3: Two GPUs can power two large screens much more easily than one alone)

A test with 5K material from a Red in the same timeline was even less favourable, as the CPUs were running at full capacity during decoding and the two GPUs were not challenged, even one GPU was only running at around 30%. An advantage from the second GPU is hardly measurable; at 8K the utilisation was even less favourable. This clearly calls for a computer with more CPU cores, which is only available from Apple in the form of the iMac Pro – which, however, is priced far above a similarly equipped PC. The combination of a 27 iMac and eGPU, on the other hand, is attractive in all cases where the CPU performance is sufficient. If you add the excellent screen, this combination is not easily beaten by a PC.

Metal was slightly faster than OpenCL for most filters in DR, but both can utilise the eGPU. Among the third-party filters, Boris Continuum does not utilise the second GPU with either Metal or OpenCL. Newer filters from Red Giant Universe, on the other hand, are up to 45% faster with Metal and eGPU, but the older ToonIt, for example, only uses one GPU. FilmConvert also requires a GPU, but does not benefit from the eGPU. Neatvideo recognises both GPUs, but is only about 10% faster. Effects from Resolve,
which are accelerated by the eGPU also benefit in Fusion. Fusion’s own functions, on the other hand, use both GPUs, but only half of them, so there is no recognisable speed advantage. We also measured a few demanding functions from DR separately. Noise reduction is certainly the most interesting, as the combination of temporal NR with 1 frame radius and “better” and spatial NR (better) almost achieves the quality of Neatvideo. Here the eGPU brought a massive advantage of 85%.

The “Automatic Dirt Removal” (better) filter is also computationally intensive and benefits by 70%. However, effects such as “Aperture Diffraction” (94%), “Film Grain” (82%), “Lens Blur” (40%) or “Deflicker” (14%) show how different the advantages can be. Optical Flow, Stabilisation and Tracking hardly benefit at all; only the Stabiliser showed a slight advantage. We were also interested in whether the GPUs are significantly involved in encoding. At best, there were slight advantages when encoding to Cineform with 6% and to DNxHR with 12%; the eGPU was not helpful with the common GOP formats or DCI encoding.

The second screen hardly slows things down as long as you don’t have a player with a UHD video or similar running there at the same time. It can be fully utilised in Resolve without noticeably reducing the gains made by the eGPU, and nothing happens on it during final rendering anyway. It was also largely irrelevant whether it was connected to the main computer or the eGPU. It is regrettable that the eGPU does not contain an I/O card to enable reliable monitoring on a calibrated screen. On the other hand, you can use the “Use Mac Display Colour Profiles for Viewers” function if you only work for the Internet and the screen has been calibrated.

eGPU and MacBook Pro 13

The simplest MacBook Pro from 2017 only has Intel graphics with 1.5 GB VRAM, a modest i5 CPU with two cores at 2.3 GHz and 8 GB RAM in our test computer. This is less about acceleration and more about the question of whether you can work seriously with Resolve at all. We first ran our UHD test sequence on the poor laptop without an eGPU and were amazed that the programme didn’t crash immediately. Blackmagic itself believes the laptop is capable of working with ProRes in HD at best. Needless to say, the sequence was as tough as chewing gum and serious work at 1-2 fps was out of the question. Thanks to the hardware decoding, UHD clips in H.264 or H.265 can even be played back in UHD at 8 bit, but you can’t add any filters or grading nodes, otherwise the weak GPU will slow everything down.

With the eGPU, on the other hand, you can work quite comfortably in an HD timeline as long as the codec and resolution of the original are not too demanding. ProRes and DNxHR in HD are unproblematic and can also tolerate complex grading, NR or other filters thanks to the eGPU. Even UHD/4K is possible to a limited extent in an HD timeline: Temporal noise reduction (faster) runs smoothly at 25 fps, without eGPU only at 8 fps, film grain is also real-time capable with eGPU, without it is only 3 fps. The very demanding Dust Remover (also temporal) only manages half that, but without eGPU it is only 3.5 fps. H.264 runs smoothly up to 60 fps, without eGPU the laptop barely manages 30. The stabilisation of UHD sources is 3.5 times faster with eGPU than without, the tracker is still 70% faster. It should be clear that clips in 4.6K DNG or even 8K R3D overtax the system, even playing compressed DNG in 4.6K in an HD sequence collapses to approx. 17 fps without any grading.

With the eGPU, you can also use the MacBook to optimise material or convert it to an offline codec with a rough correction (aka First Light) in order to work with it on the go. But you should treat the CPU to a laptop cooler or reduce the rendering speed a little under “Deliver”, otherwise the poor little computer will get very hot – the eGPU will remain very cool.[/vc_column_text][/vc_column][/vc_row][vc_row][vc_column][vc_gallery type=”image_grid” images=”68181,68182,68183″ img_size=”200×200″][/vc_column][/vc_row][vc_row][vc_column][vc_column_text]

Comment

The ideal combination for mobile filmmakers basically consists of all three devices: The iMac and the eGPU at the workplace quickly calculate offline material for the laptop, and creative work can be done almost anywhere. You come back, make a few fine corrections and have the project output in full quality. The compact and robust 4.5 kg eGPU can also be taken to a shoot to show the team how to make corrections to the material without having to wait. Only high-end formats such as Red are almost impossible to handle and unfortunately the GPU cannot be replaced. But if successful, perhaps a model with Vega64 on the iMac Pro will be available for cinema production[/vc_column_text][/vc_column][/vc_row]