Since may, the British weida launched a new generation of TeslaM2090GPU products, claims to be the world’s fastest efficiency operation, and application in parallel processor HP new ProLiantSL390G7 server. HP SL390 series server could be embedded into eight most M2090, at the same time SL390G7 products provide combining with the GPU CPU mixed operation environment.
It is reported, in HPDISCOVER on congress, HP facing users and partners launched GPUStarterKit (GPU starter kit), this technology does not need in servers use M2090GPU coprocessor to realize HPC on the performance of the 10 trillion times. Starter kit has two ProLiantSL6500 server box, eight ProLiantSL390sG72U computing node, case in the space can hold three GPU, and the product has a low profile in April in the market.
In addition, data on the understanding, each from SL390sG7 server node has two main frequency of 3.06 GHz processor, the X5675 resistance to a total of eight node working together, double the ability of floating point calculations peak performance can reach 1.18 trillion times. Each node configuration for three M2070 no fan GPU coprocessor, make double floating-point calculations of the peak performance ability, add up to 12.36 trillion times. A frame of the GPU and CPU floating-point calculations combined for 13.54 trillion times ability.
HP said GPU starter kit, will the red hat Linux version and enterprise operating system installed in the node together, and cluster manager program and Linux value suit provide high performance computing users expanded version. CUDA r&d environment and normal operation time also apply to this product, rack configuration a DL380 as control node and a 36 of the port InfiniBand switch and a 24 port Ethernet switches, HP user basically put the node after open, and the network and storage, can run application software.
If users need to be more strong performance, HP each node of the server can be configured M2090GPU, because of this product has higher GPU clock frequency, the more big memory bandwidth can be integrated, 512 core, double properties can be floating point calculations per second, make 665 billion 24 GPU coprocessor floating-point calculations of the combined for 16 trillion times.
But, according to HP revealed through 4 U version in ProLiantSL390 series server application, make each server configuration GPU, provide more GPU and CPU floating-point calculations ability total more than 30 trillion times performance.