GPU Choices

The modern Apple is a big fan of GPU power. This is true regardless of whether we’re talking about phones, tablets, notebooks or, more recently, desktops. The new Mac Pro is no exception as it is the first Mac in Apple history to ship with two GPUs by default.

AMD won the contract this time around. The new Mac Pro comes outfitted with a pair of identical Pitcairn, Tahiti LE or Tahiti XT derived FirePro branded GPUs. These are 28nm Graphics Core Next 1.0 based GPUs, so not the absolute latest tech from AMD but the latest of what you’d find carrying a FirePro name.

The model numbers are unique to Apple. FirePro D300, D500 and D700 are the only three options available on the new Mac Pro. The D300 is Pitcairn based, D500 appears to use a Tahiti LE with a wider 384-bit memory bus while D700 is a full blown Tahiti XT. I’ve tossed the specs into the table below:

Mac Pro (Late 2013) GPU Options
  AMD FirePro D300 AMD FirePro D500 AMD FirePro D700
SPs 1280 1536 2048
GPU Clock (base) 800MHz 650MHz 650MHz
GPU Clock (boost) 850MHz 725MHz 850MHz
Single Precision GFLOPS 2176 GFLOPS 2227 GFLOPS 3481 GFLOPS
Double Precision GFLOPS 136 GFLOPS 556.8 GFLOPS 870.4 GFLOPS
Texture Units 80 96 128
ROPs 32 32 32
Transistor Count 2.8 Billion 4.3 Billion 4.3 Billion
Memory Interface 256-bit GDDR5 384-bit GDDR5 384-bit GDDR5
Memory Datarate 5080MHz 5080MHz 5480MHz
Peak GPU Memory Bandwidth 160 GB/s 240 GB/s 264 GB/s
GPU Memory 2GB 3GB 6GB
Apple Upgrade Cost (Base Config) - +$400 +$1000
Apple Upgrade Cost (High End Config) - - +$600

Despite the FirePro brand, these GPUs have at least some features in common with their desktop Radeon counterparts. FirePro GPUs ship with ECC memory, however in the case of the FirePro D300/D500/D700, ECC isn’t enabled on the GPU memories. Similarly, CrossFire X isn’t supported by FirePro (instead you get CrossFire Pro) but in the case of the Dx00 cards you do get CrossFire X support under Windows. 

Each GPU gets a full PCIe 3.0 x16 interface to the Xeon CPU via a custom high density connector and flex cable on the bottom of each GPU card in the Mac Pro. I believe Apple also integrated CrossFire X bridge support over this cable. 

With two GPUs standard in every Mac Pro configuration, there’s obviously OS support for the configuration. Under Windows, that amounts to basic CrossFire X support. Apple’s Boot Camp drivers ship with CFX support, and you can download the latest Catalyst drivers directly from AMD and enable CFX under Windows as well. I did the latter and found that despite the option being there I couldn’t actually disable CrossFire X under Windows. Disabling CFX would drop power consumption, but I didn't always see a corresponding decrease in performance.

Under OS X the situation is a bit more complicated. There is no system-wide CrossFire X equivalent that will automatically split up rendering tasks across both GPUs. By default, one GPU is setup for display duties while the other is used exclusively for GPU compute workloads. GPUs are notoriously bad at context switching, which can severely limit compute performance if the GPU also has to deal with the rendering workloads associated with display in a modern OS. NVIDIA sought to address a similar problem with their Maximus technology, combining Quadro and Tesla cards into a single system for display and compute.

Due to the nature of the default GPU division under OS X, all games by default will only use a single GPU. It is up to the game developer to recognize and split rendering across both GPUs, which no one is doing at present. Unfortunately firing up two instances of a 3D workload won’t load balance across the two GPUs by default. I ran Unigine Heaven and Valley benchmarks in parallel, unfortunately both were scheduled on the display GPU leaving the compute GPU completely idle.

The same is true for professional applications. By default you will see only one GPU used for compute workloads. Just like the gaming example however, applications may be written to spread compute workloads out across both GPUs if they need the horsepower. The latest update to Final Cut Pro (10.1) is one example of an app that has been specifically written to take advantage of both GPUs in compute tasks.

The question of which GPU to choose is a difficult one. There are substantial differences in performance between all of the options. The D700 for example offers 75% more single precision compute than the D300 and 56% more than the D500. All of the GPUs have the same number of render backends however, so all of them should be equally capable of driving a 4K display. In many professional apps, the bigger driver for the higher end GPU options will likely be the larger VRAM configurations.

I was particularly surprised by how much video memory Final Cut Pro appeared to take up on the primary (non-compute) GPU. I measured over 3GB of video memory usage while on a 1080p display, editing 4K content. The D700 is the only configuration Apple offers with more than 3GB of video memory. I’m not exactly sure how the experience would degrade if you had less, but throwing more VRAM at the problem doesn’t seem to be a bad idea. The compute GPU’s memory usage is very limited (obviously) until the GPU is actually in use. OS X reported ~8GB of usage when idle, which I can only assume is a bug and a backwards way of saying that none of the memory was in use. Under a GPU compute load (effects rendering in FCP), I saw around 2GB of memory usage on the compute GPU.

Since Final Cut Pro 10.1 appears to be a flagship app for the Mac Pro’s CPU + GPU configuration, I did some poking around to see how the three separate processors are used in the application. Basic rendering still happens on the CPU. With 4K content and the right effects I see 20 - 21 threads in use, maxing out nearly all available cores and threads. I still believe the 8-core version may be a slightly better choice if you're concerned about cost, but that's a guess on my part since I don't have a ton of 4K FCP 10.1 projects to profile. The obvious benefit to the 12-core version is you get more performance when the workload allows it, and when it doesn't you get a more responsive system.

Live preview of content that has yet to be rendered is also CPU bound. I don’t see substantial GPU compute use here, and the same is actually true for the CPU. Scrubbing through and playing back non-rendered content seems to use between 1 - 3 CPU cores. Even if you apply video effects to the project, prior to rendering this ends up being a predominantly CPU workload with the non-compute (display) GPU spending some cycles.

It’s when you actually go to render visual effects that the compute GPU kicks in. Video rendering/transcoding, as I mentioned earlier, is still a CPU bound affair but all effects rendering takes place on the GPUs. The GPU workload increases depending on the number of effects layered upon one another. Effects rendering appears to be spread over both GPUs, with the compute GPU taking the brunt of the workload in some cases and in others the two appear more balanced.


GPU load while running my 4K CPU+GPU FCP 10.1 workload

Final Cut Pro’s division of labor between CPU and GPUs exemplifies what you’ll need to see happen across the board if you want big performance gains going forward. If you’re not bound by storage performance and want more than double digit increases in performance, your applications will have to take advantage of GPU computing to get significant speedups. There are some exceptions (e.g. leveraging AVX hardware in the CPU cores), but for the most part this heterogeneous approach is what needs to happen. What we’ve seen from FCP shows us that the solution won’t come in the form of CPU performance no longer mattering and GPU performance being all we care about. A huge portion of my workflow in Final Cut Pro is still CPU bound, the GPU is used to accelerate certain components within the application. You need the best of both to build good, high performance systems going forward.

The PCIe Layout Gaming Performance
POST A COMMENT

267 Comments

View All Comments

  • Anand Lal Shimpi - Tuesday, December 31, 2013 - link

    Behind the multithreaded curve, the two are almost identical :) Reply
  • Calista - Tuesday, December 31, 2013 - link

    It seem to follow the multithreaded graph perfectly, a tiny bit of a blue graph can be seen in the upper right corner. So it's actually hidden by the second graph. Reply
  • japtor - Tuesday, December 31, 2013 - link

    One thing I keep forgetting to ask since it hasn't been mentioned anywhere, does AirPlay display spanning/mirroring work? I figured it used QSV on the other Macs which this machine doesn't have, so just curious if they just left it out completely. Reply
  • skiboysteve - Tuesday, December 31, 2013 - link

    I wonder if they changed it to AMD hardware encode... Reply
  • Calista - Tuesday, December 31, 2013 - link

    So not really a proper upgrade for anyone owning a Mac Pro from the last few years unless Thunderbolt, faster GPU or a small form factor is needed.

    Anyway, it's an impressive package and it's clear Apple have brought with them a lot of the knowledge they have gained over the years building laptops.
    Reply
  • Lonyo - Tuesday, December 31, 2013 - link

    Indeed. As was the question with the "super thin" iMac... what's the point?

    It's all very well having a super small computer blah blah, but in this instance, for this type of machine and end user, what's the real benefit?
    The cost has gone up for a base model, performance per dollar has gone down compared to the previous one, there's no ability to upgrade GPUs.
    As soon as you start plugging in Thunderbolt devices, there goes your "sleek looks" etc. Plus it's more expensive to get a Thunderbolt HDD/etc than just stick one inside the case, further increasing costs.

    Yes, it looks nice, and from an engineering standpoint it's very well done, but... is it really the right product for the market?
    Reply
  • Calista - Wednesday, January 1, 2014 - link

    For anyone not planning on bringing the computer with them from time to time it's certainly not a very practical design. Desktop-space is often more highly valued than floorspace, and lack of upgrade paths are obviously a con.

    But for those with a need to bring a powerful computer with them on a set or similar it's a much more practical solution as compared to the previous design. I think Apple was quite aware what they were doing. A complete field setup with a 27" monitor, the Mac Pro, cables, keyboard and mouse is less than 20 kg. Much more than a laptop for sure, but still a fairly acceptable weight.
    Reply
  • Lonyo - Wednesday, January 1, 2014 - link

    Wouldn't it make more sense to have TWO designs then? A Mac Pro for people who need portability, and a Mac Pro for standard single location users...?

    I mean, I know Apple tends to be all about deciding what the consumer wants for them and removing choice as much as possible, but sometimes that's not the best way.
    Reply
  • akdj - Wednesday, January 1, 2014 - link

    Why doesn't this model fit that mold? For the stay at home/office/studio...one can easily AND reasonably tie thunderbolt storage together in a very acceptable and aesthetic way. Whether it be a drive enclosure...set of enclosures, TB docks that are now available adding more USB 3/HDMI/audio/et al I/O....who needs a huge box for slow internal 3.5" HDDs anymore? These PCIe SSDs tear the 2.5" models apart. Inside the 'old' style MP, a 'new' GPU on X16 takes up two slots! Sure doesn't leave much room for your MIDI, PCIe SSD or external pro sound card!
    I'm amazed at how few 'get it' here anymore. Especially after such an exhaustive review. I'm a bit biased as I make my mortgage and have for 22 years doing audio and video production. From hauling reel to reels, vinyl, film and racks of around and lighting gear to rMBPs, iPads (now with fill 64 channel wifi front of house control with Mackie) and this new Pro....I've shaved thousands of pounds from load ins and outs. Same in the camera realm. Working the last seven years with Discovery and it's subsidiaries in Alaska, I can't put into words what this machine means to us. And it's ability to pay itself off many times over just in the course of a year. Exciting times. Hopefully an evolution Win OEMs will consider as well. Shouldn't be any moving parts any longer. Wait and space are ALWAYS an issue. As is the price of power....the advantage of speed, and software developers following suit to unload computational crunching to the GPU
    Reply
  • nunomoreira10 - Tuesday, December 31, 2013 - link

    Now there just need to be a dual cpu single gpu offer to please everybody. Reply

Log in

Don't have an account? Sign up now