The story of AMD’s rise to glory on the back of Intel’s failures is well known. Intel, filled with the hubris that can only come from maintaining a dominate market position as long as they had, thought that the world could be brought into the 64bit world on the back of their brand new platform: Itanium. The cost for adopting this platform was high however as it made no attempts to be backwards compatible, forcing you to revamp your entire software stack to take advantage of it (the benefits of which were highly questionable). AMD, seeing the writing on the wall, instead developed their x86-64 architecture which not only promised 64bit compatibility but even went as far as to outclass then current generation Intel processors in 32bit performance. It was then an uphill battle for Intel to play catchup with AMD but the past few years have seen Intel dominate AMD in almost every metric with the one exception of performance per dollar at the low end.
That could be set to change however with AMD announcing their new processors, dubbed Kaveri:
On the surface Kaveri doesn’t seem too different from the regular processors you’ll see on the market today, sporting an on-die graphics card alongside the core compute units. As the above picture shows however the amount of on die space dedicated to said GPU is far more than any other chip currently on the market and indeed the transistor count, which is a cool 2.1 billion, is a testament to this. After that however it starts to look more and more like a traditional quad core CPU with an integrated graphics chip, something few would get excited about, but the real power of AMD’s new Kaveri chips comes from the architectural changes that underpin this insanely complex piece of silicon.
The integration of GPUs onto CPUs has been the standard for some years now with 90% of chips being shipped with an on-die graphics processor. For all intents and purposes the distinction between them and discrete units are their location within the computer as they’re essentially identical at the functional level. There is some advantages gained due to being so close to the CPU (usually to do with latency that’s eliminated by not having to communicate over the PCIe bus) but they’re still typically inferior due to the amount of die space that can be dedicated to them. This was especially true of generations previous to the current one which weren’t much better than the integrated graphics cards that shipped with many motherboards.
Kaveri, however, brings with it something that no other CPU has managed before: a unified memory architecture.
Under the hood under every computer is a whole cornucopia of different styles of memory, each with their own specific purpose. Traditionally the GPU and CPU would each have their own discrete pieces of memory, the CPU with its own pool of RAM (which is typically what people refer to) and the GPU with similar. Integrated graphics would typically take advantage of the system RAM, reserving part a section for its own use. In Kaveri the distinction between the CPU’s and GPUs memory is gone, replaced by a unified view where either processing unit is able to access the others. This might not sound particularly impressive but it’s by far one of the biggest changes to come to computing in recent memory and AMD is undoubtedly the pioneer in this realm.
GPUs power comes from their ability to rapidly process highly parallelizable tasks, examples being things like rendering or number crunching. Traditionally however they’re constrained by how fast they can talk with the more general purpose CPU which is responsible for giving it tasks and interpreting the results. Such activities usually involve costly copy operations that flow through slow interconnects in your PC, drastically reducing the effectiveness of a GPU’s power. Kaveri CPUs on the other hand suffer from no such limitations allowing for seamless communication between the GPU and the CPU enabling them both to perform tasks and share results without the traditional overhead.
The one caveat at this point however is that software needs to be explicitly coded to take advantage of this unified architecture. AMD is working extremely hard to get low level tools to support this, meaning that programs should eventually be able to take advantage of it without much hassle, however it does mean that the Kaveri hardware is arriving long before the software will be able to take advantage of it. It’s sounding a lot like an Itanium moment here, for sure, but as long as AMD holds good to their promises of working with tools developers to take advantage of this (whilst retaining the required backwards compatibility) this has the potential to be another coup for AMD.
If the results from the commercial units are anything to go by then Kaveri looks very promising. Sure it’s not a performance powerhouse but it certainly holds its own against the competition and I’m sure once the tools catch up you’ll start to see benchmarks demonstrating the power of a unified memory architecture. That may be a year or two out from now but rest assured this is likely the future for computing and every other chip manufacturer in the world will be rushing to replicate what AMD has created here.
