At the beginning of April, AMD announced that Llano (the codename for the A-series APUs) parts were shipping for revenue. Thus we have two generations of AMD discrete GPUs and two first-generation APUs using the same hardware gubbins to provide compute power. AMD recently released OpenCL SDK 2.4, which significantly improved performance and access to hardware abilities, in preparation for AFDS. Rage3D had the chance to talk with John Taylor (Director, Client Product & Software Marketing) today about the announcement and OpenCL with AMD products. What follows is our speculation and thoughts based on that conversation.
A-series APUs (codenamed Llano) are made using Global Foundries (or GloFlo as we like to call it) 32nm process, and combine current AMD Athlon II/Phenom II x86 cores (up to 4) with DirectX 11 graphics (up to 480 stream cores). Currently, AMD's Athlon II processors use GloFo's 45nm SOI process, so the shrink obviously gave AMD's engineering boffins an opportunity to apply timely and well-placed tweaks. Additionally, all DirectX 11 graphics parts were previously made on TSMC's 40nm process. Combining two different process technology designs onto a new one isn't copy-paste, especially when one of the two is from a completely different foundry.
Therefore, you might think it's likely that AMD did more than just import a design file into their process tool and click the 'process node shrink' button. Whatever tweaks were applied, it was probably to make room in the package budget (thermal and transistor) for the graphics core bits to be integrated, rather than ramping up IPC. The CPU cores are likely based on the most recent CPU architecture from the STARS family, as seen in Thuban products like the Phenom II X6. This implies Turbo CORE support is a possibility, although the L3 cache will be sacrificed. Thinking about the 'up to 480 stream core' design, it sounds suspiciously like a Turks family based design, which was recently introduced as the AMD Radeon HD 6600 and 6500 series. This bodes well for Llano APUs, as they will provide a truly useful amount of gaming performance, more than enough for most - and lots of compute performance too. AMD are notionally estimating around the 400-500GFLOP range for single precision, which compares very favorably to any traditional CPU currently available, especially for the guesstimated 65W-100W TDP range of the desktop Llano variants.
VLIW4 vs. VLIW-5
You might also consider VLIW-4 as the basis for Llano's GPU cores. AMD's Radeon HD 6900 series use a different GPU architecture from the rest of the Northern Islands (HD 6000 series) and Evergreen (HD 5000 series) range of discrete GPUs. Instead of arranging the stream cores in fours and adding a special purpose stream core, all four stream cores are equally capable and make up a thread processor. 16 thread processors make a SIMD as usual for AMD's designs, but in VLIW-4 GPUs a SIMD is 64 stream cores instead of 80. AMD made this architectural change to address the need for more general purpose compute performance. VLIW-5 is a great architecture arrangement for gaming and light GPGPU, but more serious compute needs a more serious architecture - and at this time that is VLIW-4. Basic math indicates that Llano isn't using a VLIW-4 based SIMD - 480 divided by 64 doesn't result in a whole number, while 80 stream cores per SIMD does. Up to 6 SIMDs in Llano, and basically the same GPU architecture as found in C-series and E-series APUs. This simplifies things for developers as they have only a single underlying architecture design for which to optimize, which happens to be the majority of AMD OpenCL compute devices. How the transition to mixing in VLIW-4 will be handled is going to be interesting.
AMD Fusion Developer Summit
So, June is the target for Fusion APUs to go mainstream - and AMD is hosting their first ever Fusion Developer Summit on the doorstep of one of the largest congregations of Software Developers in the world. Being an x86 licensee and selling product designs that power the world's most popular Operating Systems, AMD is naturally working with Microsoft for their APUs. Currently AMD has been working very closely with several key independent software vendors (ISVs) to encourage OpenCL adoption into mainstream products like Adobe, Cyberlink and Arcsoft media processing products. AMD is opening up to more of the software ecosphere to help more and more developers leverage the power of OpenCL running on AMD GPUs. This includes a big focus on University programs, where AMD has strategic partnerships with key Universities, including creating Innovation centers, providing silicon, software and course materials. Another generation of Microsoft .NET Framework and Visual Studio developers is born, and able to leverage AMD OpenCL processing power.
At CES 2011 we heard how Microsoft is widening their hardware platform support to include ARM based devices. The fast growing and very popular smart/super phone and tablet market runs on ARM, not x86-licensed hardware. ARM will be speaking at the AMD Fusion Developer Summit in the same terms as Microsoft, about heterogenous computing. Windows 8 will run on ARM and x86, everybody supports OpenCL - Intel, NVIDIA, AMD - and as ARM expands to support OpenCL it makes the set. John Taylor was very specific about how there is no announcement of ARM and AMD products, and that there was no comment to be made about possible collaborations - except to say, look at ARM products now, and the GPU cores attached to them. The GPU cores are very important parts of what makes the ARM products successful.
Divining AMD's Fusion Strategy
Putting together the pieces of the puzzle, it seems that you can just make out the outline of AMD's success strategy for Fusion on the horizon. While the x86 market seems destined to slow in growth, or even plateau, AMD doesn't seem like they're going to keep all their eggs in that basket. AMD is a Design Company, and Design Companies work with lots of different partners to provide the products that they want, and that consumers want to buy. There's no inherent reason for Accelerated Processing Units to be based on x86 processor cores alone, and AMD has two aces to offer ARM licensees with UVD technologies and their GPU prowess. Oh, plus a little bit of experience working well with others; IBM, Nintendo, Microsoft, Sony, spring to mind.
AMD is looking to make an inflection point with combining GPUs and CPUs, and they're going to do so by looking at a bigger market than just desktop PCs and servers. The computing world is looking to multiple hardware platforms to provide the best consumer experience, regardless of form factor, input method or even operating system. The first generation Fusion APUs are selling very well, and if AMD sees the same adoption rate for their Llano APUs there will be a wealth of extraordinarily powerful computers out there very soon.
Certain companies, who like OpenCL so much they let the Khronos Group use their trademark, might like ARM AMD APUs as they try to move towards an mobile OS with internet media and tunes stores, GPU accelerated interface and programs. A design company, who doesn't manufacture products that direct compete with the aforementioned certain company (but still have a close partnership with an independent foundry with experience in making ARM, x86 and various GPU cores, working with said design company) would be in a position to offer tailor made design solutions to power a certain company's PMP, smart phone, tablet, ultra-portable, notebook, AIO and high performance workstation/server products quite easily. Especially if that certain company is already adopting products said design company, delivering where another company who has an ARM+GPU strategy had a bumpy path.
content not found