As businesses assessed the damage and began digging out, the picture wasn't as gloomy as they might have feared - WSJ, on the tsunami that killed thousands
|
|
|
Fujitsu offers 16-core Sparc chip to HPC customers
SUPERCOMPUTER VENDOR Fujitsu is offering its customers access to the same kit that powers the world record holding K-Computer.
Fujitsu is offering customers the PRIMEHPC FX10, which is built from the same components that power the K-Computer. The cluster can be scaled to 1,024 racks with a total of 98,304 nodes delivering up to 23.2 petaflops of compute power.
Fujitsu's 1.848GHz Sparc IXfx chip - the firm is one of Oracle's few Sparc licensees - can access 32GB or 64GB RAM, for a cluster maximum of 6,291TB connected to a 85GB/sec memory bus. The inter-node transfer rate, which is of vital importance for massively parallel high performance computing (HPC) workloads, is 5GB/s each way.
Not surprisingly, Fujitsu runs a Linux-based operating system on the PRIMEHPC FX10 and offers optimised compilers for C, C++ and Fortran.
Fujitsu's K-Computer, which at the time was powered by 8-core Sparc chips, took the top spot in the prestigious Top500 list earlier this year. Since then the K-Computer has been upgraded and is expected to be in first place once again when the latest Top500 is published next week.
While Fujitsu shows off its 16-core Sparc chip, next week's Top500 list will also see AMD's 16-core Bulldozer Opteron chips appear, with the Jaguar tipped for an upgrade. Interestingly, even with GPGPU acceleration in vogue, Fujitsu's and AMD's HPC partners are still heavily reliant on traditional CPU power for many workloads.
Fujitsu said that it expects to ship 50 PRIMEHPC FX10 clusters within the next three years. µ
Tags: Hardware
Share this:
Share
a hell of a lot cheaper per core than full blown chips ever will be.
500 shaders are a lot cheaper than 500 intel or amd cores. a factor 10,20 or even 30 is nothing special
and they do math / pattern manipulation and so on, so extremely well and very fast.
and for that reason, there will be more of them. lots and lots of them.
With the evolution on the CPU designs heading more and more towards throughput computing -- more cores and more wider vector ISA extensions, the dedicated parallel number crunchers like GPUs are founding them selfs in more and more narrowing market niche.
The heterogeneous computing paradise, that AMD so much dreamed of, is evaporating in the mist of the harsh reality. OpenCL is still nothing more than an experimental field for a few academia researches to play with, and the API development itself has become quite stagnant in the last year. Intel, with its next architecture - Haswell, have every chance to hit the 1 TFLOPS mark for a single CPU, without the driver and API overhead of the traditional GPGPU implementation. GPUs will, after all, get back to what they do and have done best - drawing pretty pixels and polygons on the screen... preferably in an integrated form as part of their master, the CPU.