Blackmagic eGPU and IMac Pro

I am wondering if the new eGPU does make any sense with an iMac Pro with Vega 64, since the eGPU has a less powerful graphic card. Are there speed advantages with Resolve 15 Studio?

Thanks

I believe so yes Resolve can use multiple GPUs. Paul Saccone uses the eGPU with this iMac Pro

since the eGPU has a less powerful graphic card. Are there speed advantages with Resolve 15 Studio?

Two is better than one, but if you pair them and share the UI with compute, the combined system will be as if you had two of the weaker GPUs (the 580) - total VRAM is not cumulative, and since these are not nVidia cards, there is no CUDA leveraging obviously. OpenCL works very nicely now, so Resolve won't suffer particularly because of the AMD GPU configuration.

jPo, CSI

While this eGPU is perfect for a non-Pro iMac with the same card, in your case I’d get me a third party eGFX box with TB-3 and put another Vega64 into it.
Should be better bang for the buck.

While this eGPU is perfect for a non-Pro iMac with the same card, in your case I’d get me a third party eGFX box with TB-3 and put another Vega64 into it.
Should be better bang for the buck.

That's what I thought...
Thanks!

Two is better than one, but if you pair them and share the UI with compute, the combined system will be as if you had two of the weaker GPUs (the 580) - total VRAM is not cumulative...

Didn’t realize that VRAM would not be cumulative. It is very unfortunate for those with projects pushing the limits of their VRAM that adding eGPU only gives you more power.


Sent from my iPhone using Tapatalk

Rick, the best way to get rid of never-ending concerns about that is to have similar cards. We've been discussing a lot here these 3 past years about the potential benefits of the GUI dedicated GPU if the cards are not the same, but in the end, the best way (in case of multiple cards) is to have similar ones + studio version + gui gpu to compute ticked in the pref...

Because yes, 2 x 8 Go vram = 8 Go vram for Resolve... Very good to share computational jobs (and therefore better than one gpu), but 8 and not 16. Resulting btw in 6 Go + 8Go GPUs = 6 Go used by Resolve, hence the recommendation to use similar ones...

Didn’t realize that VRAM would not be cumulative. It is very unfortunate for those with projects pushing the limits of their VRAM that adding eGPU only gives you more power.

Maybe some day BMD will revamp its engine so that it can better use asymmetric GPUs.

It would add significant complexity though, which is probably why the devs over there haven't done it yet.

Don't hold your breath, they got a few other things to fix.

BTW, external GPUs will still be restricted by the speed of Thunderbolt, even version 3. While the Radeon 580 gave me about 80% gain over a single Radeon Pro 580, which might be a explained by general computing overhead, a guy who tried two in a chain only got something like another 30%.

If you have the option, get one really strong GPU inside your machine. For Mac users, I'd hold out until Apple finally get's their new Mac Pro out (if ever), go with a Hackintosh or finally make the switch. I say that as a user of both systems and my experience is this: Resolve 15 is more stable under MacOS, but you get much more power for your money in a PC. Choose your poison!

Personally, I decided not to waste my money on an overpriced iMac Pro, but got me the iMac with the 580 as an intermediate solution. But if Apple doesn't deliver by the end of this year, I'm gone or I may add the eGPU and live with it. Depends on the projects coming in.

Don't hold your breath, they got a few other things to fix... For Mac users, I'd hold out until Apple finally get's their new Mac Pro out (if ever)... got me the iMac with the 580 as an intermediate solution. But if Apple doesn't deliver by the end of this year, I'm gone or I may add the eGPU and live with it...

Apple, in one of their recent intimate fireside chats with the chosen few, has said that the newer Mac Pro will not be ready in 2018. The cognoscenti felt Apple’s implication was targeting 2019, but even that wasn’t explicitly promised.

From what I’ve seen, the iMac you have with internal and external AMD 580 should serve you fairly well until 2020 and is reasonably good value.

Apple’s very long development cycle for the new Mac Pro late 2013 and the newer Mac Pro which might be late 2019 suggests they may be repeating history. I have the feeling that they’re waiting on some technology that isn’t quite ready yet and that depends on third parties so that Apple isn’t in control of the schedule.

You know they want to knock your socks off again, but the longer it takes the more likelihood that they’ll miss the boat on the next great thing and your newer Mac Pro will be dated within two years. The new Mac Pro was pretty exciting when we had the sneak preview but by the time significant numbers of people were using those dual D700 GPUs, the cracks in the closed design were evident.

Apple said the newer Mac Pro won’t have that problem because it’s modular. If something better comes along, just pop in the newest module. Sure, we all know, it’s that easy... not. The price may be into five figures and it may buy you a few more years, but you’d need a crystal ball to predict today where various technologies will be by 2025 and I expect they’ll only be applied by the newest Mac Pro that will begin to be designed in 2022 for release in 2027.


Sent from my iPad using Tapatalk

Apart from some more CPU power I'm not unhappy with that iMac.

I have a similar feeling like you, Rick, about Apple waiting for someone else. I suppose it's Intel, who have postponed their schedules for newer processor technologies recently (again).

The end is near (for Moore's 'law').

Indeed... the trashcan was a repeat of one of SGI's classic blunders.

I bet hardly anyone here is aware that SGI actually made an x86 workstation that ran the then current version of Windows.

It had some pretty amazing technology in it; SGI built it around a very high throughput crossbar that provided gobs of bandwidth. It included an SGI designed GPU that was in theory spectacularly fast and delivered excellent image quality worthy of the Silicon Graphics legacy.

The problem?

Due to being all custom, it was late. A year late. By then, Intel had newer processors that required new sockets, and nVidia and ATI had newer graphics processors that outperformed the one in the SGI box, which also due to being all custom, was significantly more expensive.

The iMac Pro reflects similar blunders also, though it was at least reasonably current; the reason that its GPU is however not competitive with even the aging nVidia Pascal GPUs out there is that it's custom made for Apple.

Lisa Su has been focussing the Radeon Technologies Group on semi-custom GPU design for the revenue, rather than on competing with nVidia... because AMD can't afford to push the envelope on CPUs and GPUs at the same time. So we get the monster ThreadRipper2 dominating benchmarks, and the Vega coming in behind the previous generation Pascal and steamrolled by Turing.

And we have the former head of the Radeon Technologies Group now at Intel, seeking to do there what AMD couldn't afford to let him do while at Radeon.

Semi-custom works for consoles and for products like Kaby Lake-G (the one with an integrated "Vega" GPU -- it's actually a Fiji core with some enhancements, so while it's far ahead of the Iris Pro, it's not competitive outside of the ultrabook market). That's also why we see AMD processors and GPUs in every major console these days, but no hint of nVidia and Intel.

If Apple is going down the road of making a modular Mac Pro based on custom hardware, it will likely be out of date BEFORE it's available, kind of like the G5... competitive when announced, shipped a year late, ended up being barely a blip in the annals of computing history...

With Tim Cook at the helm, if the ‘newer’ Mac Pro doesn’t sell well in 2019/20... well there may never be a ‘newest’ Mac Pro in 2025.

By the way, if anyone is confused or annoyed that I call the 2019 Mac Pro, the newer Mac Pro and the 2025 Mac Pro, the newest Mac Pro, it’s pointing out the absurdity of the 2013 Mac Pro that was officially marketed as the new Mac Pro. How bereft of ideas Apple Marketing has become to name any device ‘new.’ Ergo new begets newer begets newest... and there isn’t another level of comparison in the English language so that must be the last Mac Pro.


Sent from my iPad using Tapatalk

Apart from some more CPU power I'm not unhappy with that iMac.

I have a similar feeling like you, Rick, about Apple waiting for someone else. I suppose it's Intel, who have postponed their schedules for newer processor technologies recently (again).

The end is near (for Moore's 'law').

Apple also appears to be developing their own chipsets:
https://www.theverge.com/2018/4/3/17191 ... ompetition
Support reading: https://en.wikipedia.org/wiki/ARM_architecture

The Bloomberg article suggests that Moore's Law was done 3 years ago.
The reason is that the silicon chip is at the end of its 14/10 nM physical constraints... and there is the limitation of "the speed of causality" or c. The speed of the universe at which the speed of light is defined. Look for a terrific little PBS short with those keywords and the Lorentz Transform sometime if you've got 15 minutes to watch and a day to think about it.

Moore's Law would still be in effect if exotic materials (diamond substrate, gallium arsenide, etc.) were in play, but the other limiting factor "economics" dwarfs those in terms of cost of production, rejection rates and amortization, aka obsolescence. As super-using professionals trying to run a piece of software that over-demands what is currently on the market as flashy consumer goods, we are only too aware that making a profit is marginal when you have to keep re-investing serious money on equipment that goes over the cliff in a couple of months.

jPo, CSI

BTW, external GPUs will still be restricted by the speed of Thunderbolt, even version 3. While the Radeon 580 gave me about 80% gain over a single Radeon Pro 580, which might be a explained by general computing overhead, a guy who tried two in a chain only got something like another 30%.

If he is literally daisy chaining, (ie plugging eGPU #2 into the 2nd thunderbolt port on eGPU #1) that will explain the 2nd eGPU only adding a 30% improvement. Each eGPU should be on its own TB3 port of the host computer, and ideally each eGPU on its own bus. MacBook Pro and iMac Pro each have 2 Thunderbolt Busses. The iMac has just 2 thunderbolt 3 ports so I would guess they are on the same bus.

Re: Moore’s Law at its limit
That seems to true with the current level of 7nm by TSMC Taiwan possibly being the final shrink on silicon. Although there are activities in the lab storing ‘data’ at the atomic level, it does sound too difficult to commercialize. We’ve heard of gallium arsenide for a decade or more. We’ve heard of memory that is 1,000x faster than current technology for half a decade. But getting these things in integrated function commercially viable forms hasn’t happened. Data paths get wider, the number of cores on a chip continue to grow, but for virtually all computing needs we’re likely at one practical limit as well with 64bit computing for 99.999% of our needs. Never say never, but the idea of doubling something every two years may not be sustainable. It takes leaps in technology and not the tweaking we’ve been subjected to on desktops and workstations.

Warning: tangent fast approaching!

I could go out on a limb and say a CEO may no longer expect to receive $120,000,000 in incentivized stock awards beyond normal compensation for simply doing his job or a tennis star signs a single endorsement for $300,000,000. It’s about time Moore’s Law ends in that arena too.

Remember when some folks used to say “What’s a million?” to minimize the impact of spending or losing $1,000,000? That’s when I was young. And ordinary folks were upset to hear that. We’ve already got to the point where the expression is now “What’s a billion?” and no one bats an eye when only a trillion gets our attention.

Back to your regular programming. Apologies.

Apple also appears to be developing their own chipsets:
https://www.theverge.com/2018/4/3/17191 ... ompetition
Support reading: https://en.wikipedia.org/wiki/ARM_architecture

Apple's current ARM efforts aren't even close to i7 or i9 performance.

The Bloomberg article suggests that Moore's Law was done 3 years ago.

Not quite done, but it's near enough to be slowing down quite a bit.

The reason is that the silicon chip is at the end of its 14/10 nM physical constraints... and there is the limitation of "the speed of causality" or c. The speed of the universe at which the speed of light is defined. Look for a terrific little PBS short with those keywords and the Lorentz Transform sometime if you've got 15 minutes to watch and a day to think about it.

The speed of light isn't the limiting factor presently. The limiting factor is that we're toeing the line where the size of the artifacts on a chip are so small that the dieletric walls between the wire are so small that electrons can simply hop from one to the other at random. The dieletrics aren't current opaque enough to stop them any longer. Also, the transistors are so small that without exotic 3-dimensional manufacturing techniques like FinFETs, the gates leak too much current when in the "off" state to be effective switches for the same reason -- the gates aren't much bigger then electron's Gaussian wave packets.

Moore's Law would still be in effect if exotic materials (diamond substrate, gallium arsenide, etc.) were in play, but the other limiting factor "economics" dwarfs those in terms of cost of production, rejection rates and amortization, aka obsolescence. As super-using professionals trying to run a piece of software that over-demands what is currently on the market as flashy consumer goods, we are only too aware that making a profit is marginal when you have to keep re-investing serious money on equipment that goes over the cliff in a couple of months.

jPo, CSI

Other exotics that have been in R&D for a long time include carbon nanotobes and photonics. There have been some optical computers built, but they were one-off proof of concept deals, not manufacturable ones.

In the near term, we're probably going to see more approaches like what AMD and Fujitsu are doing in parallel with more of what nVidia is doing. AMD and Fujitsu are using multiple dies arranged in a sophisticated chip carrier with a high-speed, high-scalability bus like what AMD is calling the Infinity Fabric, while nVidia is going for truly massive parallelism and combining specialized "cores" (Tensor, RT) along with simple but general "cores" (the CUDA cores). Those cores are arrange hierarchically with controllers that feed the cores data and retrieve the stuff they're done with, etc.

It's fun to speculate sometimes

The iMac has one bus, so it doesn't matter where you plug. AFAIK, so do the MBPs.

But even on the iMac Pro you may want to connect quite a few fast storage devices, and that should not be on the same bus with the eGPU.
I doubt that in practical terms more than one eGPU makes much sense with either iMac – but it should be the right one.

My question is a simple layman’s one. Not so technical. I have a Blackmagic egpu, but can’t use it currently. I can sell it. That’s one option. A second option is that I’m contemplating buying a MacBook Pro 2018 model or an iMac Pro. I know the Blackmagic egpu will work with the MacBook Pro 2018 model. My question is will it work with the iMac Pro?