IBM Goes 'GPU-riffic' with new blade
GPU goodness — not just for HPC anymore
GTC Video Blog IBM made big news on the first day of last week's GPU Technology Conference by announcing that it'll roll out an Nvidia Fermi–based expansion blade. While it's not formally announced yet (the plan is to do so in Q4), IBM had one at the show and walked me through it for the video below.
It isn't a standalone server; it's a single Fermi GPU with 6GB memory that clicks onto a host – initially an IBM HS22 two-socket blade with 12 DDR3 DIMM slots and all the typical IBM bladey trimmings. It attaches to the host blade via a 16x Gen-2 PCIe port.
The ports on the Fermi blade extension have a pass-through feature so that up to four expansion blades can be attached to a single host. That's a hell of a lot of GPU goodness — and it's certainly GPU-riffic.
Also GPU-riffic is the fact that these expansion blades fit into the most popular IBM BladeCenter chassis (the E, H, and S varieties). This is significant in that it marks a big step for GPU computing. There are a lot of these blade chassis (and H22 blades) out in Customer-Land, and these new GPU blades really open up the market for non-HPC centric buyers to give GPUs a try.
These products, and others like them from HP and Dell, make GPUs a standard IT item rather than some exotic technology that's suitable only for labs, Hadron Colliders, or data centers supporting super villains.
IBM also showed the Fermi offering for their ultra-dense iDataplex server, as you'll see on the video. I remember getting an early briefing back when iDataplex was just a vision in Jimmy the Bull's mind. At that time, it was firmly aimed only at the mega Net 2.0-type data centers (think Yahoo!, Google, and folks like that). It was supposed to be extremely dense and efficient, with servers that had only what was absolutely necessary to function — with no additional RAS or management hardware features.
I had two initial thoughts during those discussions. The first was that I didn't totally believe that IBM could stay true to the vision of providing a stripped-down server. I figured it would be hard for them to resist the urge to add on additional hardware features that might seem "nice to have", but which would pump up the cost, size, and power draw. I was wrong on this score — they successfully delivered what they promised.
My second thought was, "It's dense, cheap to buy and operate, and you're looking to sell it in big chunks, why not target HPC too? It's a good fit."
They didn't agree with my reasoning, at least initially, but the market said differently, and iDataplex has become a standard IBM HPC offering. ®
I'm still working on making the term "GPU-riffic" a part of the industry lexicon. Anyone who is supporting "GPU-tastic" as an alternative should reconsider their stance.
I can see these as a great use to major operators as they computing provide the power for their custom data processing applications to run on, however I have to agree that i'll certainly see no benefit from plugging a GPU in to my server which runs headless with standard workloads on, because i have no software that can make any use at all of GPU.
This is probably the case for the vast majority of people, however I can see the potential for database servers to start supporting GPU offloading for their enterprise database users now that there are products like this coming to the market, and for this to then trickle down to become another standard feature. This is a software problem, cheap high power hardware is everywhere as a result of the gaming industry, now we just need software to take advantage or it.
the real question is why can i still not use those 64 400MHz cores in my bottom end graphics card to run a pretty beefy database server?
nVidia and ATI thought outside the box.
What they found was that parallel computation has plenty of uses outside gaming (look at Folding@home; their strongest contributions come from GPUs). Where GPUs excel is what is called "stream computing": essentially, repetitive independent calculations which are ripe for parallelization (if you think the 4-way multitasking of modern quad-core CPUs is hot, GPUs can divide tasks 32 ways or more). Plus in terms of raw computational horsepower, GPUs have CPUs licked (the top-end cards are reaching teraFLOP levels in single-precision--a double-precision teraFLOP GPU device is only a matter of time).
BTW, there's serious uses for graphical computation, too. Think climate and weather modeling, physics simulations, and other forms of "what if" modeling. GPU computations can even help with professional ray tracing and similar forms of advanced 3D modeling (not to the realtime level yet, but they can still seriously cut down rendering times).
that's what I thought it meant
but it just confuses me more - why would a server blade need better graphics?
I can understand someone bumping up the graphics card to play games.