Supercomputing slice of history, Cray, has finally boldly gone where other supercomputer makers have gone before and released details of its GPU equipped machine.
The XK6, like the name suggests, is a cousin to the XE6 which was the AMD Opteron-based machine that the company announced a year ago. Cray said that XK6 systems will be available in the second half of the year.
Cray has come late to the GPU but that did not stop its vice president Barry Bolding, vice president, of Cray’s products division claiming it was still unique in the GPU computing space.
This uniqueness comes because of Crays long-term commitment to “heterogeneous computing” and is track record for “building productive petascale systems”, apparently.
The XE6 was born out of Cray’s background with vector-based supercomputing and compilers which recently was shown in its full glory with the X2 super.
According to HPC the XK6 is not that different from Cray’s CPU.based offerings. It is a blade based bit of kit with four of the eight AMD Opteron sockets and the other four Nvidia Tesla GPU modules. Each of the four-node blades is made up of two Gemini interconnect chips, four Opteron CPUs, and four Nvidia Tesla 20-series GPUs.
Cray used the X2090 Tesla which is a more compact form of the new M2090 module. It has a 665 gigaflop (double precision) GPU, 6 GB of GDDR5, and 178 GB/second of memory bandwidth.
A single 24 blades cabinet will deliver 70 teraflops.
So it is a fairly conservative processor and GPU mix because Cray thinks that not all supercomputing workloads need a GPU.
The XK6 will use the XE6’s software with GPU-specific libraries and tools like Nvidia’s CUDA SDK. Cray is developing its own OpenMP-based compiler that will be targeted for GPU acceleration which is a higher-level programming model than CUDA.