Let's start with the theoretical stuff, even though it's largely academic until more software comes along that can make use of it. The reason AMD calls the GPU cores inside Kaveri "compute cores" is that they're said to be fundamentally different to the GPU cores in other PC processors. This difference lies in the fact that they're able to function as equal citizens: Instead of relying on the CPU to orchestrate their workload, they can access system memory directly and take on tasks independently -- almost like a CPU core does. The only difference is that they can't take on the same types of tasks as a CPU, as they're better suited to simple parallel chores rather than complicated serial processing.
As things stand, software developers are already able to exploit the GPU for general computing using tools like OpenCL, which can be used to accelerate anything from Photoshop to big spreadsheets. But OpenCL requires reams of code and a lot of inefficient to-ing and fro-ing between the GPU and CPU -- all of which, AMD says, will be drastically reduced if developers latch onto HSA. That's a big "if," of course, but now that AMD has recruited a bunch of partners into its HSA Foundation, and now that it has managed push its silicon into millions of households via next-gen games consoles, developer interest looks more likely, and Kaveri's compute cores at least bring it some future-proofing as a result.
Bearing in mind that we're mostly reliant on AMD's in-house test results for now, until independent reviewers put their graphs online, let's look at that basic claim about Kaveri undercutting Intel as a gaming processor. The chart above shows a top-end Kaveri A10-7850K pitted against Intel's Core i5-4670K for games being played at 1080p with max settings (or at least close to max settings -- there's a bit of ambiguity there, but it doesn't affect the comparison). In each case, the processor is paired with a discrete graphics card, AMD's mid-range Radeon R9 270X, presumably because most enthusiasts would still avoid relying solely on integrated graphics. As you can see, Intel is slightly ahead in a number of games, but never by a significant margin, suggesting that spending $70 more on Intel's chip doesn't add much to the experience.
Power efficiency and onboard graphics
In addition to Kaveri's suitability for gaming when paired with a separate graphics card, the slide above suggests the chip also has an advantage over a Haswell Core i5 on certain synthetic benchmarks, likely due the fact that it has a bigger GPU than you'd find on an Intel processor. Kaveri's built-in GPU accounts for 47 percent of all transistors in the chip (more than a billion in total), and is potentially meaty enough for it to run games without the need for a discrete graphics card, thereby saving energy and money while also allowing for much smaller PCs. In practice, we played through a level of BioShock Infinite at 1080p with low settings, with Kaveri running beneath a little third-party cooler, and we experienced a steady frame rate of 30fps. This is something AMD claims is also possible in other big titles like Battlefield 4, which it's bundling free with high-end boxed Kaveri chips, but again, you have to be prepared to accept low detail settings.
For the sake of balance, it's important to point out that an Intel's chip is likely to be more power-efficient in its own right. Haswell has fewer transistors (1.4 billion instead of Kaveri's 2.3 billion) and its transistors are also significantly smaller (22nm instead of 28nm), which should equate to reduced power draw -- something that's especially significant when you think about notebook or hybrid/tablet versions of these chips, particularly ones that don't need to focus on 3D graphics (or, equally, which delegate all such tasks to a separate GPU).
Mantle and TrueAudio
Speaking of Battlefield 4, we arrive neatly at Kaveri's other big claim to fame -- and it's a claim that requires a much smaller leap of faith than HSA does. You see, Battlefield 4 is one of a growing number of games that will take advantage of an AMD-tailored programming tool called Mantle, which promises big boosts in performance even on lower-power (e.g., HTPC and laptop) versions of the chip. Mantle runs on any AMD graphics card that contains the newer Graphics Core Next (GCN) architecture, and since Kaveri's graphics processor is based on GCN, it can run Mantle-optimized games and applications too, resulting in claimed performance increases of up to 45 percent in BF4 (once it gets its Mantle update later this month) and as much as 300 percent in real-time strategy games running on the new Star Swarm game engine. (For more on Mantle, read this.)
Finally, in addition to Mantle, Kaveri also brings another feature across from AMD's latest graphics cards: TrueAudio. This is a dedicated, programmable audio processor that sits on the chip and helps to improve the audio in games by decoding data about location (giving sounds a feeling of directionality and distance) and also increasing the total number of voices and effects that can be heard at one time.
Kaveri apparently took four years to develop, due to all the extra gubbins AMD has squeezed onto it, including HSA, Mantle and TrueAudio. This also explains why Kaveri chips are priced significantly higher than their predecessor, Richland: The lower-specced A8-7600 will start at $119, rising to $152 for the A10-7700K and, as we've mentioned, $173 for the flagship A10. Will they be worth the money? We'll wait to round-up independent reviews from specialist sites before we make any final judgment, but it certainly looks like AMD has brought some clever additions to this generation that could boost its value. It looks good as a traditional gaming processor right now, especially if you intend to pair it with a Radeon graphics card in order to enable Dual Graphics (with the GCN cores in Kaveri's GPU and in the discrete GPU effectively being added together), but we'll need to see more Mantle- and HSA-enabled software before we're ready to believe it can tackle Intel on general computing.