"Cray CX1 Experience Seminar" in Tokyo

There was a Cray CX1 Experience Seminar was held by Cray Japan in the afternoon of January 27 in Akihabara Convention Hall in Tokyo. Cray CX1 is the system which I was interested in since I found it which was put casually on the floor in the SC08’s Cray booth last November.

It was a good opportunity for me to attend this seminar since I left IBM and am not in any vendors now. (In this industry, we usually do not accept a competitor's entry to seminars mutually.) Many thanks to the courtesy of the Nakano-san, president of Cray Japan with decades-knowing.

As shown next, the seminar was organized as all-in-one about Cray CX-1 probably similar to CX-1 itself. Therefore it was a galloping seminar. But it was an intelligible seminar and I enjoyed it.

The conference room with 130 seats was almost full with attendees. That means that many people had various kinds of interests about the Cray's personal supercomputer with Accelerators.

(Program)
- Trend of HPC (Cray Japan)
- CX1 overview and its advantage (Cray Japan)
- Parallel Computing by CUDA: Prologue (NVIDIA Japan)
- CX1 Demonstration
(1) Employment of SCRYU/Tetra and LS-DYNA with Cray CX1 Job Launcher (Cray Japan)
(2) SCRYU/Tetra Demonstration on Windows HPC Server 2008 (Software Cradle)
(3) LS-DYNA Demonstration (JSOL)
- CPU information and the latest road map (Intel)
- Efforts of Microsoft for HPC (Microsoft)

Cray CX1 looks one of typical small blade servers which can work by 100V power, for example, IBM Blade Center S. Cray provides a computing blade (two models), a storage blade (two models), and a graphics blade which includes NVIDIA QUADRO FX series or TESLA inside. Since it costs several million yen for four blade-configuration models in a campaign price according to a Cray brochure given in the seminar), the price range is in the Workgroup Server. It looks unusual that Cray which has a track record in supercomputer market of hundreds of million yen offers products for workgroup server, the minimum price range of HPC servers. They are going to enhance Cray CX-1 and follow-on systems in cooperation with Intel, NVIDIA and Microsoft.

According to a talk by NVIDIA, because TESLA is just an expensive and professional QUADRO without graphics function, they provide it around half the price of QUADRO. (In addition, it is supported by huge sales volume and revenue by GEFORCE in consumers market.) They target twice in 18 months for performance improvement. The NVIDIA speaker said that double precision feature had been added last November.

They reported that the number of downloads of CUDA, application development environment for GPU, already exceeded 120,000. CUDA ZONE web site may show such prosperous situation in development of GPU applications.

The CX1 demonstration of SCRYU/Tetra by the Software Cradle, well-known CAE software vendor in Japan, was a CFD calculation applied to an intake manifold and an engine cylinder. It is almost scalable up to eight cores for 180k element-model and up to 16 cores for 1 to 8 million element models (Approx. 10 times faster in 16 cores than 1 core. The performance is not sensitive to a number of elements for 1 million or more element case. InfiniBand is used. )

JSOL (Former Japan Research Institute Solutions) demonstrated the LS-DYNA standard problems (Neon refined and 3 Vehicle Collision) for automotive crash analysis. Their results showed a good scalability up to 32 cores. (26.8 times (IB interconnect) or 23.7 times (GbE interconnect) faster in 32 cores that single core)

I wonder why performance differences are so small between a high bandwidth and low latency IB case and a low speed GbE case. Maybe, both codes are already well tuned in order to minimize communication for moderately parallel computing (I do not confirm it yet).

If such ISV's CAE software could show scalable performance and an acceptable licensing fee structure up to 1,024 cores or more, Industrial sector's usage in HPC would be much more accelerated to larger systems, considering trends of HPC platforms. If so, many things will change. Do you agree?