AMD Flopping Around - GPU'ing Intel or destined to be outflanked? Virtualizing your server in new and interesting ways.

So, just Monday, AMD started talking about its Firestream GPU off-loading again, in a big way ... breaking the 1 Teraflop barrier. What was missed in that Computerworld news story is the potential for CPU offload with the FireStream (given a little more attention in this older Computerworld piece as well as this piece at The Register). To give you some feel for what a Teraflop is, Sun's constellation supercomputing cluster cranks out 7.37 Teraflops with a rack of quad-core Xeon blades.

AMD's FireStream is an interesting technology that I've been following for a while (Nvidia's CUDA is similar, and has recently been talked about in depth over at Tom's Hardware for the truly curious), but I think the potential is overlooked by most folks in the market. Sure, FireStream alone is an interesting acceleration technology for single systems doing HPC-like processing or rendering. When you talk about the latest, the potential of a 10x increase or more for computationally intensive applications with a mere additional investment of 110 watts of power (practically only another CPU), things are interesting enough. But there are also real implications here for connectivity fabrics and how applications work within your infrastructure. In many ways, I consider this the real start of general virtualization 2.0, which reaches much deeper than the silos of storage virtualization and server virtualization we have today. Somewhere in here is the early hint of what is to come for grid computing and complete infrastructure virtualization.

What I mean by this is the distribution of processing resources across a fabric. Once you have some mechanism for assigning the processing of certain tasks to other systems/processors/resources, then you can do some truly flexible things. We're on the cusp of building a mainframe across an entire data center, where resources can be flexibly assigned and optimally utilized. In the potential grand realization of this scheme, it's a world where there's quality of service for everything, and it may be much easier to perform big picture management of all of your infrastructure.

AMD is tackling this with Hypertransport at the bus level, as well as PCIe, and InfiniBand is tackling this with a fabric outside the bus. In fact, in this regard, InfiniBand can essentially be an extension of the system bus, and InfiniBand and Hypertransport are not all that dissimilar. Is anybody paying attention to this potential? Sure. 3Leaf is one interesting small vendor with a grander vision, Xsigo is doing some interesting IO virtualization that hints at the future, and key InfiniBand vendors, such as Voltaire, Cisco, and Mellanox, have been working on this vision for a while. Most of these vendors are just looking at virtualizing IO adapters outside of server boxes today - a single InfiniBand adapter can use a gateway on the InfiniBand fabric to serve up multiple hardware managed virtual HBA's and virtual NICs to the host, or even virtual guests. But that's the precedent. Once you have what is essentially bus-level connectivity outside of single server box, you can potentially connect and share anything you might imagine - including processors or GPU's (General Processing GPU's, or GPGPU's) from other servers.

I personally think one of the killer opportunities here would be for consumer oriented gaming rigs, just because of ease of market entry and potential revenue. In a market spending hundreds or thousands of dollars on graphic cards, popping for a couple of InfiniBand adapters and slaving two high-end gaming machines together for double the performance and killer game frame rates is an interesting proposition. It is interesting precedent that AMD is demo'ing this product - an external graphics processor for laptops.

Then carry that vision forward into the datacenter... The most interesting implication today, is that you can get some incredible capabilities in the datacenter if you're willing to introduce "pockets" of different connectivity fabrics (e.g. InfiniBand) for selected systems. For example, if you have a rack of virtual servers, throw in an InfiniBand fabric just for that rack, and an InfiniBand gateway to connect it back to your existing FC, Ethernet, FCoE, or other fabrics. Without much more complexity, you can get some killer flexibility in that virtual environment. Pretty much, virtual everything. Caveat emptor, you'll have to invest some time in qualifying the architecture, or work with a trusted vendor. What do you think, is there pay-off in selecting special purpose fabrics and/or toolsets to increase infrastructure flexibility for specific pockets of servers in your enterprise?

I see a lot of AMD's/Nvidia's GPU co-processing as a fundamental impetus in this space. And it's happening in a market that is conveniently dealing with some fundamental challenges in parallel processing, both in x86 architectures as well as the cutting edge stuff happening down at the System-on-Chip (SoC) level (I'm still impressed by the old Aarohi SoC Asic accomplishments put together with the help of the guys over at ARC International - it's vendors like these guys, that are practically in stealth mode to the end user, that are the unsung heroes of making your stuff rock). So at the same time we're innovating around parallel processing, AMD/Nvidia GPU offload SDK's can be welcomed into mainstream computing with relatively open arms.

Now in the midst of this, you have Intel messaging that traditional sequential GPU processing is long in the tooth, and it's time for a new architecture. Fear, or vision? I think we'll have to wait until that Siggraph paper mentioned in the EETimes story on Intel's Larrabee is released in August. In my opinion, it's here at these deep architecture levels that things really have interesting implications for vendors and compute technologies alike. Your thoughts?

------------

Jeff Boles

Visit us at www.tanejagroup.com for the latest in analyst opinions.