Typically whenever a roadmap is announced, people assume that the company will stick to that roadmap and deliver on its promises and then some in a bit to keep up profit levels and attract more customers. AMD has been on track with 15% performance increases between processor families and Intel’s been trucking along with overall increases of 10% per family for years. A recently leaked product diagram, however, may be a blueprint for a future AMD product that could blow Intel out of the water, but it’s not intended for consumers – the internet exploded last weekend when people thought this was a new graphics card.
If the above graph looks like gibberish, don’t worry – I had no idea what I was looking at in the beginning when I first saw the image. But once you read a bit, it becomes clearer- this is either an APU or a future high-end GPU for the server market.
The theoretical product you see before you has 4096 stream processors divided into 16 separate Compute Unit arrays. It has Cache-coherent links that my link it up to the processor as per the hUMA technology we looked at a while ago. It supports Displayport 1.2 and uses PCI-Express 3.0.
As a further nod to the possibility of this being an APU is the inclusion of eight Serial Processing Modules, each consisting of two cores and a shared Floating-Point Unit, which is the design of Bulldozer and Piledriver acrhitecture. The sixteen-core processor is linked to the Parallel Compute Node with a shared North Bridge and both connect to the System controller hub. That could be a very low-clocked component or it could run at full speed, we just don’t know any of the details yet.
There are also eight 72-bit DDR3 ECC memory controllers. Error-Correcting memory isn’t normally found in consumer products and it if it, it’s normally in a workstation designed for specific purposes with increased reliability and accuracy. The hints of other core logic like the Security co-processor and AMD-V and AMD-P, both virtualisation technologies, point to this being a server product rather than anything from the consumer market. There’s also “Advanced Server RAS” logic promised if you were still in doubt.
If true, though, this could change the game for AMD and Intel. Currently Intel’s servers aim for a very high performance-per-watt ratio and if you take the performance overall, their Xeon lineup is tremendously efficient. But, there’s a catch – Intel doesn’t have the parallel computation abilities that Nvidia’s CUDA can offer, through the use of Quadro and Tesla graphics processors to accelerate application performance.
Intel’s Xeon Phi, a parallel processor made up of Pentium cores, comes close to offering the perfect all-Intel solution but it’s not that efficient. The developer kits for the Xeon Phi only just came out of beta and no-one’s really made a fuss about them since the launch. In performance terms a HD7970 would beat it in OpenCL workloads, but Xeon Phi is more versatile and not restricted to using a proprietary software standard, like CUDA.
AMD will be likewise looking for a way into the massively parallel server market and they actually bought Sea Micro a while ago. They also announced last year that they’ll be using ARM chips as security co-processors. What you’re seeing above is a server product that will likely be sold in separate units. The PCMs and the SPMs will all be able to work on the same data thanks to hUMA advancements and you’d have this huge chunk of compute muscle ready to do anything and everything thanks to OpenCL.
In addition, all that platform management logic means that you’ll be able to get racks or rooms full of these things with the ability to switch them on and off using software as the load increases or decreases. Remember, AMD’s new version of Turbo Boost already load-balances clock speed against the temperature and load on the GPU in Bonaire, so you could scale it to something like a mainframe server all running these APUs.
But I’m a gamer, how does this benefit me?
Good question. AMD’s next family after the HD7000 series is codenamed Sea Islands. Sea Islands will likely introduce some new features like advanced hUMA cache sharing and more on-board RAM to support higher resolutions. Sea Islands will probably only come onto the market in late 2013 or early 2014 – well after the Playstation 4, at any rate – but it will be GCN 2.0 and will incorporate all the previous technology as well as some new tricks AMD has learned while building the hardware for the next-gen consoles and the launch of Bonaire. Solar System, the update for the mobile market is already on its way – the glimpse of performance the HD8790M offered is enough to cement my theory that AMD will easily be competitive with Kepler this year.
Volcanic Islands is the family after Sea Islands, and will use codenames from Islands in and around the Pacific Ring of Fire. Its flagship will probably be similar to the PCM we’ve seen in the image above – 4096 stream processors, a 512-bit bus, anything up to 6GB of RAM and hUMA along with it. It’ll blow away the GTX Titan and the HD7990 and it’ll probably be one of the biggest, most complex cards AMD has ever created. And it’ll be sent into the sea filled with a new generation of gamers and hardware, tasked with driving 4K displays and Eyefinity setups or accelerating your Bitcoin mining.
Whatever the future holds, it’s going to be awesome.
Discuss this in the forums: Linky