+1
There are no applications in Windows out yet that can leverage HSA, and probably, the only applications that will ever leverage some of the capabilities of HSA, are going to be games that use the Mantle API.
In linux, there is no threshold for using scaling functionality in applications, but a lot of applications are not evolving that rapidly. For business and scientific users, the benefits of HSA are already there in applications, but they have to install the system tools to benefit from them manually. That will change with Fedora 21, which is now already available as Fedora rawhide, although it's still in an early integration stage because they're waiting for more merges and bug fixes by Intel and the open source AMD driver. As of right now, Intel has no OpenCL capabilities in its drivers, and hasn't merged all OpenGL functionality either. AMD's open source driver is still a couple of generations behind on OpenGL, but has OpenCL and since a couple of weeks, also full GPU audio support, which is very nice, and which is in my view the bonus feature that sets AMD solutions apart from Intel solutions. New features (OpenGL/OpenCL/OpenMP) are added almost every week, the integration is going really fast. On AMD, you can access all the functionality with the Catalyst driver, but that's a typical proprietary graphics driver like nVidia's, it's a pretty crappy piece of coding that also handles things like anti-cheating logic for certain games, patches for certain games, and a bunch of pretty unsafe code. The only benefit Catalyst has over the nVidia proprietary driver, is that it compiles fine with newer kernels as they are released, and with nVidia, there is always breakage, and the kernel modules need to be patched by third parties, which is just annoying.
A Kaveri is a 12-core chip in it's present configuration, in the HSA logic, it's basically a scaled system consisting of 4 full CPU cores, whereby every CPU core has 2 GP-GPU cores at it's disposal. All 12 cores are connected to the system RAM and each other through HUMA, which is a step up from IOMMU. The potential of such a configuration is huge, because in an HSA optimized system, if you add a discrete GP-GPU with many cores for instance, each CPU core can also directly use the cores of the discrete GP-GPU through IOMMU, and the GP-GPU (RHD7k and R-series) can directly access the system memory. The linux kernel has already been unlocked recently for handling more than 8000 cores, and that is necessary, because in a perfectly scaling system like linux, each GP-GPU core, providing the application is coded to do that, can execute jobs directly from and to system memory, without having to wait between cycles for a CPU instruction. Intel has made the Phi to do that, and the Phi goes a step further even, it can run it's own parallel linux operating system, but therefore it's not as efficient as the AMD GP-GPU's for the price of the hardware.
The thing is, right now, you can't fathom the impact of HSA and HSA-specific chips yet, because there is nothing in userspace - besides scientific and enterprise stuff, that's where Kabini comes in, with even more cores than Kaveri - that can really show you the benefit of the platform. Two years from now though, everyone will be substituting "normal" CPU's with APU's and iGPU-CPUs, because of the huge difference in performance. One thing to note is that people with an AMD Phenom or FX CPU and a RHD7k+ GP-GPU, will also be able to use the HSA functionality, and will be able to tag along with HSA for at least a couple of years, before the HSA-specific chips become so powerful that they clearly outperform traditional IOMMU-enabled discrete systems. People with a non-IOMMU-locked Intel iGPU-CPU, will probably be able to leverage the iGPU or a Phi, but not any discrete GP-GPU by AMD. It's unclear whether HSA scaling will even be possible on nVidia hardware, probably not to be honest, unless nVidia makes some fundamental changes in its designs (which would lower their margin because it would cost them more to make cards) and open sources those designs (which will probably also never happen).
In order to appreciate HSA, you have to forget about the present technologies, and think differently, HSA is about scaling, it's about removing the CPU bottleneck. Also, HSA will demonstrate that Windows is just a software console, because you'll only have certain benefits of the technology in Windows, but the huge performance boost will not be there. Instead, there will be a lot of face recognition and other spyware apps for windows that uses the HSA resources that remain inaccessible to the system for over-all performance. And AMD has done this is quite a clever way, they are not investing in these Windows applications, they are outsourcing it to commercial platforms that will mine data and pay money to AMD, and that leverages the commercial consumer exploitation aspect of Windows further.
So what are enthusiasts going to do with HSA chips? They are going to add a GP-GPU or Phi to their systems, and scale it up to exactly the performance point they can afford, and it's going to be much higher than the performance point they can afford in non-HSA systems. Does this performance point mean higher fps in games? Yes and no! A well optimized OpenGL game, even with stupid resolutions, already gets a really good fps in linux: most linux games, even with the less capable drivers that are available now, score a much higher fps in linux than in windows, and OpenGL makes the effects just look a whole lot better (especially things like blooming effects, diffuse lighting effects, etc...). On top of that, the CPU load will be a lot lower with HSA optimization, so more resources will be free for more sprites, more game logic, more bots, more players on an MMO map, etc... and not a few more, but exponentially more. That will enhance the capabilities of game devs to provide a better gaming experience, to make environments more lifelike, to let more players play together on huge maps. I think the focus is going to shift away from fps benchmarks, and towards a better gaming experience. What hardware features will become very important: lots of system memory (even though linux is efficient in terms of memory use, large textures and more game logic will require a bigger system RAM, and on discrete GP-GPU cards also a pretty large VRAM), fast system memory (as GP-GPU cores will directly access the system memory, it will need to be fast and have a low access time, there is definitely some more work to be done on the development of newer RAM technologies), and lots of very fast storage (the textures will have to be loaded from somewhere, and SATA3 right now is bottlenecking fast ssd's already, so there is definitely also some room for improvement there).
So HSA chips are only the first step, it's not a revolution, but an important evolution. Not everyone will be able to benefit from it in the same way right away, not everyone will be super enthusiastic about it right away, and software and hardware devs will need time to adapt to HSA.
So does it make sense to compare benchmarks between HSA and other chips? No, because there isn't much to benchmark for the moment. An APU is a good choice for a budget rig and for linux users that are into enterprise and scientific applications that already leverage scaling technologies. Whether Intel chips will perform on a similar level, remains to be seen, but I'm sure Intel is working on it like crazy, and it wouldn't suprise me if in as little as 5 years from now, Intel will offer chips that have something like 512 to 1024 "logical" cores, whereby different parts of the die can play different roles, depending on the kind of resources that are needed. I just think that - if the software development is not blocked by patents and shady business deals - AMD will probably rule for a few years until Intel catches up, especially for enterprises and scientific institutions that want bang for the buck.