By now many Architosh readers have spent some time checking out all the available information on Apple’s site for the new 2013 Mac Pro. There are also some good sources of additional photos and a few really good articles on additional details on the Mac Pro. Over at Arstechnica, veteran Mac 3D pro user, David Girard, has written a very comprehensive article.
After speaking with several insiders we thought we’d cover all the details about the graphics and expandability that we now know.
Details on the Graphics – The Choice
For the 2013 Mac Pros Apple has chosen to adopt the AMD FirePro W9000 as the basis for its dual-GPU setup. This is AMD’s fastest technology based on the all new Graphics Core Next (GCN) Architecture. Our contact at AMD reminded us yesterday that this is the most powerful workstation graphics card ever created.
Unlike the Nvidia Quadro 6000, the FirePro W9000 features a 384-bit memory interface delivering 264 GB/sec memory bandwidth as compared to the Quadro 6000 with its 384-bit memory interface delivering 144 GB/sec. This is a 1.8 x delta–advantage AMD. Additionally, the AMD card can calculate more geometry faster, delivering 1.95 billion triangles per second compared to Nvidia’s card at 1.3 billion per second. That is a 1.5 x delta–advantage AMD.
Yet while this AMD card is a performance monster, we believe Apple has partly chosen AMD to field the graphics due to its parallel compute performance based on the Graphics Core Next (GCN) Architecture. Our source has told us that the rivalry between Nvidia and AMD has swung a bit over the past few generations of technology. In the previous generation Nvidia was leading with “compute” (with its CUDA architecture) while AMD led with gaming. Now in the latest generation technology the pendulum has swung the other way. AMD is now looking like the leader in parallel compute capability. Its new GCN architecture is very strong and its hardware clearly shines in the compute realm.
The Detail on Graphics – Options
A really big question on people’s minds is will the new Mac Pro have upgradeable graphics? From the initial images and information online it seemed it would. But over at Arstechnica this excellent article says that the graphics are soldered onto the mother boards. If so, that is truly disappointing news for most pro users.
Does Apple have an upgrade plan? What about initial configuration options?
On the latter question it is not yet clear from our information what Apple has in mind for initial configuration options. They could conceivably lower the graphics memory to sell dual-GPU configs with say 3 and 6 GB of memory each GPU. We don’t think they will have a single GPU option and we have not learned otherwise (yet).
Options will be important for pro Mac users because the AMD FirePro W9000 retails for just one card at over 3,000.USD. If Apple actually installs the equivalent of two W9000 cards that’s 6,000.USD of firepower. That’s expensive.
This would be strange because in effect Apple would create a huge gulf between its top of the line iMac machine and the new Mac Pro. Dave Girard made the case that Apple is not going to charge the full retail amount for this level of GPU power.
Girard makes the point that typically “workstation GPUs” are inflated price-wise to begin with–which they are. He further argues that AMD doesn’t need to start from scratch in writing the drivers for it either or build an entirely new support structure for the FirePro card. He is correct about many of these things. The same Mac graphics division at AMD now is in charge of implementing the dual FirePro GPUs for the Mac Pro. They will be updating drivers and testing compatibility for all the GPUs they have issued in the past for the next OS X Mavericks.
For Apple to integrate a “workstation class” line from either AMD or Nvidia doesn’t change a whole lot within the graphics divisions of either of those companies today. Still, just how much Apple will essentially charge for these dual GPUs is anybody’d guess.
The Detail on Thunderbolt 2 for Graphics
Some folks are thinking that Apple’s future path for expandability for the new Mac Pro is all about Thunderbolt 2. They are likely very much correct in this assumption. But it doesn’t appear that 2013 is the year for Thunderbolt technology to start supporting advanced graphics on the future Mac Pro. This isn’t to say the idea is not in demand.
Over at Magma the hardware company notes that many customers are very hungry for a Thunderbolt-based PCI Express chassis for external graphics cards. PCI Express cards can work inside products like the Magma ExpressBox 3T for example. For a card to work, it must be supported on the OS side and come with a Thunderbolt Driver.
Over at Anandtech there is a great article on everything one needs to know about Intel’s Thunderbolt 2 technology, the technology inside of Apple’s 2013 Mac Pro. However, we want to point out that from this article we learn two key things. Firstly, Thunderbolt 2.0 is not yet PCI Express 3.0 based, and as Anand Lal Shimpi of Anandtech writes, this makes Thunderbolt 2 not quite ready for prime time when it comes to external GPUs.
Specifically, the bandwidth native to x16 (16 lane) PCI Express 2.1 and 3.0 cannot be met with Thunderbolt 2. We are not going to get into the technical explanations behind various spec levels for this technology but will point out that Apple’s new Mac Pro does support PCI Express 3.0 which means it supports 16 GB/sec data transfer from GPU to CPU. This is double that of PCI Express 2.0.
Dual GPUs and Crossfire
Some of you may be wondering how the dual GPUs work together, if they utilize AMD’s Crossfire technology. This too is not entirely clear but Crossfire was a technology created around gaming. For professional apps to tap the power of more than one GPU they need to alter their code base to distribute their workloads to both cards.
Mac OS X Mavericks may do some of this work in a future update. It should be pointed out that OpenCL doesn’t require application developers to tweak their code in order to take advantage of both GPUs or even more than two. This happens automatically as part of the OS’s implementation of OpenCL.