The supercomputer buyer is unwilling to spend several months to put up the colonizing of the super large-scale figure processor – central processing unit. They hope to buy the system that dispose and finish in advance and come into operation in a short time.
As two weeks after Great Britain Hill and Knowlton claims to introduce Tesla M2090 figure processor coprocessor used for server, Hewlett-Packard’s high performance calculation executive vice president Mike. Hamilton just writes in his blog, should meet the Integrated server framework demand in advance, the challenge that the high performance calculation group faces is to realize 10,000 billion floating point calculated performance at least, and the price is in under 100,000 dollars.
Hewlett-Packard will face GPU Starter Kit that high performance calculation user and cooperative partner put out in Las Vegas soon next week figure processor actuator kit Do not need the figure processor coprocessor to realize the expectation of the foot Hamilton while using M2090 fan-free in the server. The actuator tool is coated with butterfly type server chassis of two ProLiant SL6500s, computation node of eight ProLiant SL390s G7 2Us, can hold three figure processors in the space of chassis, Hewlett-Packard has already put this product on market in a low-key attitude in April.
Each the intersection of server and node have two primary frequency 3.06GHz to strong the intersection of X5675 and processor, eight nodes work together altogether, the peak value characteristic of the floating point computational capability of double precision can reach for 1,180 billion times. The primary frequency of coprocessor – these processors is 1.15GHz that each node has disposed three M2070 fan-free figure processors, perhaps only 448 are activated in 512 cores, the sum of the crest value characteristics of the floating point computational capability of double precision is 12,360 billion times. The sum of the floating point computational capability of central processing unit in a framework and figure processor is 13,540 billion times.
The figure processor actuator kit will be mounted in the node in advance with enterprise edition Linux operating system of the red cap, Hewlett-Packard’s own colonial supervisor procedure and expanding edition to high performance calculation users’ Linux value suit is the same. CUDA researches and develops the environment and is suitable for these models in uptime. Framework dispose one DL380 as control node and one 36 carry oral the intersection of InfiniBand and switchboard and one, 24 carry oral the intersection of Ethernet and switchboard. After users basically open the node, even have the network and store, can operate the application software.
If willing in Hewlett-Packarding, they can make density more, the server with stronger function. The first step is M2090 figure processors transferred to Great Britain Hill and Knowlton, the clock rate of this processor is higher, memory bandwidth too old, figure processor integrate 512 core, double precision the intersection of floating point and calculated performance can up to 665 billion per second. The sum of the floating point computational capability of 24 figure processor coprocessors is 16,000 billion times.
But it is more outstanding that Hewlett-Packard passes and carries out the transition to ProLiant SL390 serial 4U editions of the butterfly type server, the the intersection of edition and every two servers dispose eight figure processors ‘ There are a large number of spaces that can hold in the framework . Through carrying out the transition to this kind of more bull butterfly type server and adopting chassis of four SL6500s, figure the intersection of processor and characteristic can up to 31,900 billion per second, eight the intersection of server and nodal each nodal the intersection of floating point and computational capability can reach for 1,180 billion times per second, the crest value characteristic totals for 33,100 billion times per second.
According to this conjecture, the figure processor actuator kit may be a mutation of stemming from a certain above-mentioned type, we have reasons to guess if disposition and expenses that describe according to Hamilton are in under 100,000 dollars, general expenses should among between 150,000 dollars and 175,000 dollars big to reach, offer Tesla the intersection of figure and price of series, M of processor, Great Britain so, very difficult to be accurate judged . 4U have space can hold other eight SL390 series the intersection of G7 and node in the framework perhaps, can turn over one time of growth in characteristic of the framework in this way, the sum of the computational capability of floating point reaches for 66,200 billion times, then the price should be among between 300,000 dollars and 350,000 dollars.
12 makes ...