CHIPMAKER Intel has launched its Xeon Phi accelerator card boasting over one TeraFLOPS of peak double-precision floating point computing power.
Intel's Xeon Phi, often seen as the descendant of Larrabee, has finally been launched with Intel revealing the specifications of two cards. The firm revealed that the Xeon Phi 5110P will launch on 28 January with 60 cores running at 1GHz supported by 8GB of GDDR5 memory and will offer 1 TeraFLOPS of peak double-precision floating point computational capability.
Intel might have entered the accelerator market years after Nvidia but the firm has put together an impressive product, at least on paper. While the firm's Xeon Phi 5110P card doesn't quite beat Nvidia's Tesla K20 series in terms of peak computing power, it has considerably higher peak memory bandwidth at 320GB/s and 30 percent more onboard memory, which is crucial when it comes to working with larger datasets.
While Nvidia has been pushing its OpenACC standard for parallel programming, something which Intel termed as simply targeting GPU computing rather than CPU and GPU computing in the case of Intel's Language Extension, Intel said it is committed to move towards the OpenMP 4.0 standard. The firm committed to produce a draft specification compliant compiler by January 2013.
Intel also announced that a cut-down Xeon Phi 3100 series card will make an appearance in the first half of 2013 in actively and passively cooled SKUs. The firm didn't give specific performance figures or board configurations, merely saying that peak double-precision floating point performance will be less than one TeraFLOPS, with the boards having 6GB of GDDR5 memory and bandwidth of 240GB/s.
Intel has gone down the route of trying to make the Xeon Phi parts easier to program than rival products from Nvidia. Unlike Nvidia, which has scored considerable success in the HPC market in a relatively short period of time, Intel has been forced to get into the accelerator market as a growing number of high-end clusters are relying on accelerators to provide most of their computing power.
Intel's Xeon Phi might not deliver high enough peak performance figures to oust Nvidia's Kepler based Tesla boards yet, but given that Intel's first product has its own selling points, such as higher onboard memory capacity and bandwidth, Nvidia can't afford to be complacent with its future GPGPU accelerator product generations. µ