- Rage3D Home Page
- Discuss This Article
State of the HSA
Author: James Prior
Editor: Charles Oliver
Date: October 28th, 2012
State of the HSA - An Interview with Manju Hegde
On October 3rd, the seventh founder member of the HSA Foundation was announced, and it was a biggie - Qualcomm. This comes just weeks after Samsung announced their membership, also as founder member, sustaining a rapid growth for the non-profit foundation since its initial inception at the AMD Fusion Developer Summit in June 2012. The initial founder members of AMD, ARM, TI, Imagination, and Media Tek bring a diverse background of experience to the table, all around the custom silicon area and system on chip products.
Heterogeneous System Architecture Foundation
We had the opportunity to speak with AMD's Manju Hegde, Corporate Vice President of the Fusion Experience program and now also HSA Foundation director. We covered a broad range of topics in our conversation, and Qualcomm's decision to join the HSA Foundation was first. Qualcomm didn't join the HSA Foundation as a response to Samsung's involvement, but instead as the result of a long discussion behind the scenes about what the HSA is and why Qualcomm would benefit.
The current level of HSA integration centers around the CPU and the GPU, bringing those devices together into a single device; well beyond two dies on a single package or even AMD's latest APU, Trinity. The capabilities seen in Trinity for coherent address space addressing from both CPU and GPU sides of the processor, ability to use pointers and move the compute instead of the data show real world speeds up for OpenCL code now. In looking at Trinity you can see that the biggest improvements over previous AMD platforms happen at lower power points, especially for mobile platforms like ultraportable notebooks.
Attractive to Qualcomm (and other HSA Foundation members) is the performance/watt possibilities. We've all seen the public press coverage of AMD bemoaning the lack of mobile strategy, and we've all seen AMD's branding as a leading innovator in technology. Well, HSA is how AMD is planning to survive the PC market transformation, something that they have not anticipated well enough, as evidenced in their Q3 results call.
AMD is transforming, looking to move into different growing markets to survive as the PC market changes from desktops, workstations and commodity servers into a much more differentiated landscape. As our resident blogger Panagiotis explains so eloquently in his October 19th Kombuting article, AMD's acquisition of ATI brought with it all the tools to excel in the post-Apple iPod-iPhone-iPad world, but AMD management had other ideas. Whether the sale of BitBoys Oy founded Imageon division to Qualcomm was from lack of resources to exploit the technology and a desire to capitalize on even a small amount of IP sale, or just it-seemed-like-a-good-idea-at-the-time short sightedness we'll likely never know.
Instead AMD is looking to capitalize on the internet of things ecosystem with three things: one is x86, the next is ATI graphics, and the last is their design collaboration skills. Currently the focus is on big cores, the Bulldozer family and the unfavorable comparison to Intel's bridgey bits and next generation Haswell. If you hadn't noticed, while AMD was working on branding Fusion, acquiring ATI and messing around integrated sizeable discrete GPUs into their processor packages, Intel has significantly increased their GPU investment in both performance and capabilities. Intel remains a long way back from where AMD and NVIDIA are, but the Xeon Phi has shown everyone that when Intel needs to be at a certain point, they will invariably be there no matter how it looks now.
For all the negative attention FX has received in the desktop consumer space, it has done well in the growth areas for the server market. The Bulldozer big core design is aimed for different use cases than the traditional and historical uses the commodity server and workstation market uses them for and, while AMD is pushing it into that niche (and enthusiast desktop), it is an also-ran not a first choice. Largely this is due to OEM support, for each Bulldozer equipped server design there are half a dozen Sandy Bridge E models, which gives the appearance of getting the right fit for your operations easier with the blue badged box. HPC, Cloud and Virtualization like Opteron 6200 series though, thanks to AMD's custom instruction set, throughput biased core density and strong hardware virtualization features.
What's going largely under the radar, though, is AMD's small core. The Bobcat architecture is the basis for the little APU that could, as while the Llano APU went from selling out twice in the first two quarters to sitting on the shelf and becoming a $100 million dollar write down, the Brazos platform and derivatives have done very well. Brazos is broader than HTPCs and netbooks, it's a hugely popular in embedded applications (think digital signage, appliance computing devices like specialty cameras, controls devices) and is being trialed in some manufacturers smart TVs and set top boxes. AMD did get a few tablet design wins using the C and Z series variants, and the new Z-60 revision launched with lots of potential but no announced designs.
For some time now, people who were buying desktop and notebook PCs have switched over to smart phones and tablets, sometimes augmenting that with a notebook too. Untethered computing is very personal, we've gone past the phase of allowing people to borrow smart phones - it's your gateway to your life: banking, contacts, confidential correspondence, photos/videos; you don't let people borrow your wallet - no, you can't borrow my phone.
Despite the acclaim for Trinity and it's competitive nature against Sandy and Ivy Bridge processors for ultra-portable notebooks, design wins are fewer than would be expected. Intel's Ultrabook marketing fund created the Ultrabook market and Intel specifies in their funding to participating OEMs what products similar to ones created out of Intel resource contributions can be permitted, which makes a Trinity based ultra portable hard to implement. Windows 8 changes the needs for x86 tablets, Surface RT attacks the ARM media consumption devices (you can get a feel for how that's going to go down from Pete's column) with Windows 8 x86 surface devices covering higher end functionality, being capable of being used with your home PC applications and also in the enterprise. This is where AMD's tablet market is at the moment, and the new 28nm Temash APU can't come soon enough for AMD. Using the new Jaguar microarchitecture revision of the Bobcat family, Temash and Kabini is where all the action is. But will the Intel powered Surface designs lock it out, too?
As AMD continues on the path of HSA, the plan is to bring major competitors into the fold and coexist with them, rather than fight against restrictive competitor design fund contracts. To do this, OS support will be key and that must extend beyond Microsoft; AMD's reticence to talk about Z-60 Linux support is at odds with our conversation with Manju Hegde, where he talked about how Linux is pervasive, Brazos appliances are running Linux and HSA will could help to unify the fragmented forks. Somebody is playing coy, and this might mean exciting things are happening behind the scenes for APU powered Linux (x86 Android?) or it might mean the opposite.
With the HSA platform and intermediate language, an ARM license, willingness to perform custom design work for customers, AMD need to make an ecosystem where partners and they don't need to be scared of each other. Perhaps if they all entered into a foundation together and agreed upon common standards for design - and perhaps behavior - there is a broad spectrum of opportunity for AMD to execute in. AMD's HSA plans are going to come to fruition around 2014 with the APU's offering GPU compute context switching, unified address space and fully integrated CPU and GPU workloads. 2014 is also looking to be when ARM gets seriously competitive with 64-bit designs and their own post-DX11 GPU cores, and Intel's small core design will be a viable option, too. Viable for what? Servers. The little core design is pushing into the datacenter, and despite AMD's previous protestations to the contrary when the first talked about Bobcat, they're going there too with Bobcat, Jaguar and the full HSA designs in 2014 using both big core and little core APU plus ARM security zone processing baked in, too.
The final mystery about the HSA Foundation is who will fill the final tile space on the web page, it's likely already decided with each member choosing when they announce their membership. I can think of a few candidates that might be interested in being part of the HSA, like Broadcom, IBM, Microsoft and Sony. Microsoft and Sony? Well, they both design their own hardware, in Microsoft's case have successfully worked with AMD in the past, and AMD is rumored to be the platform provider for both next gen consoles with custom IP designs. Will the next generation of Sony Playstation and XBOX be powered by AMD APUs, GCN architecture variant, and will they be Bulldozer family architecture - or Jaguar family?