Siggraph Asia starts rolling

AFTER 34 years of US domicile, the first Far Eastern version of the famous Siggraph Singapore. Next year it'll be in Yokohama, and 2010, Seoul. Events follow the money and that is moving in huge buckets from West to East in these days of crisis.

The four day inaugural event, gathered a bunch of luminaries including Rob Cook, a VP of Pixar animation - the first Oscar winner for software;  along with some 6,000 graphics geeks that are estimated to be attending the fest.

Singapore also happens, by sheer coincidence, to host a number of top animation studios. Lucasfilm Animation Singapore, the Forced One's main base outside the US of A, was in charge of Star Wars: The Clone Wars - therefore the visitors were able to see some tricks from the Indy Jones and Ironman movies here too. Lucasfilm was one of the main Siggraph Asia job fair participants, offering over a hundred CGI job positions from some fifteen employers here.

Let's start with some familiar faces here at the exhibition - how about front-door booths by Nvidia cum Leadtek, and ATI/AMD and Sapphire? Both Green and Red GPU camps chose their "preferred" Asia Pacific professional graphics card partners to co-host the booths right at the show entrances. No, there was no GeForce GTX295 here - but Quadro FX5800, FX4800 and CX (the Photoshop flavour) led the play at the Nvidian booth.

The Nvidia Quadro team coveres some quite interesting points, including SLI scaling benefits on more polygon and plain texture throughput-oriented engineering apps compared to the a-bit-more-effects-oriented gaming. Also on show was what 4GB RAM on  the uber expensive Quadro FX5800 can do. What do you do with 4GB on a GPU, besides Tesla-like CUDA or OpenCL computation? Well, voxel-based volumetric 3D needs all the memory it can get. Shelling out 3,500 American bucks per card may just make sense for those guys... or seven grand if doing it in SLI.

According to Nvidia, the channel and OEM prices of Quadro cards don't differ much - now at least. In the days of the Quadro 980 generation, a Tier 1 PC workstation vendor could sometimes sell you a spare Quadro at half it's usual official price - not any more.

Nvidia's Quadro team still has its humongous market share lead in the profitable workstation graphics sector, but AMD's workstation graphics group has made tremendous leaps forward in the past year - partly thanks to our own Charlie, who helped the open source community get hold of ATI driver code as well. Vastly improved Linux drivers and more optimised OpenGL followed naturally then. The FirePro 9270 and upcoming 'X2' OpenGL profi cards, with 2GB and 4GB GDDR5 memory respectively, present formidable competition to the Quadros when it comes to raw FP power: 2.5x more double precision peak flops - 240 GFLOPs on the FirePro 9270 compared to the Quadro FX5800 is a given, but we have to see how much of it will be reflected in the actual OpenCL application code.

Sapphire seemed to contribute a bit more to the DAAMIT booth than Leadtek to the Nvidia one. Sapphire's 56-inch 3840x2160 QuadHD LCD displayed a crossfired FirePro 3-D hi-res output. And 600W-class workstation PSUs were shown there as well. So, a bit more than just passing off ATI's own profi cards. Most engineers prefer the 16:10 golden ratio display format for 3840x2400 resolution like the old IBM T221 - PC. A workstation, is not a TV, and minute LCD panel cost savings on overpriced large displays to be the same format as TV are not our problem.

The Red Gang also managed to solve the problem of running one computational task across two or more GPUs in their Brook+ and OpenCL implementations. Previously, this was a major problem due to the lack of cache and memory coherency between the GPUs in a single system. Nvidia CUDA needs to follow on this to widen the benefits of its multi GPGPU Tesla and Quadro configurations too.

Talking about multi GPGPU stuff, we all know about the QuadroPlex dual GPU workstation add-ons, and Tesla 1U-high quad 4 TFLOP SP GPGPU setups for supercomputing. After all, quite a few hundred of these were purchased recently by TITECH in Tokyo to replace the Clearspeed cards in their Tsubame Opteron supercomputer cluster.

Now - or, to be more precise, a quarter from now - AMD has something just as impressive. Its partner, Aprius Inc, has a lovely 4U rackmount enclosure with two modules in there, each holding up to four full-length PCIe X16 v2 graphics cards plus a full speed PCIe optical link to the workstation. If you use them with FirePro 9270, you get ten teraflops of raw single precision power in that rack, and two teraflops of double precision peak - not bad at all.

By the time the final box is ready, you may be able to put together eight dual-GPU cards in there for some 20 TFLOPs SP and 4 TFLOPS DP, for not much more than US$ 30 grand... very interesting. Let's just make sure the application code can come anywhere near that performance: since the GPGPU performance is limited to the stuff that fits to the on-card memory. Nvidia still has twice the per-GPU advantage there due to the 512-bit per GPU memory path.

To finish off the hardware side, IBM also showed some graphics stuff, plus the iDataplex X86 tightly integrated cluster solution. Unfortunately, their scalable OpenGL cluster graphics wasn't there - too bad since IBM pioneered the approach with their SGE (Scalable Graphics Engine, not Sun Grid Engine) at the start of this century.

In summary, nothing really new hardware wise here, but a pretty nice competitive show-off for the two main GPU purveyors in the professional field. Tomorrow, for the final show day, we talk a bit about those animation studios headed by the Lucasfilm 'Jedi masters' themselves, other such gurus and of course the software. µ