by Jacqueline Emigh
Unless you’re involved in high-end scientific research or algorithmic trading, you probably won’t need a MIC (many Intel core) PC server any time soon. Yet that hasn’t stopped Intel from starting to wow crowds with a new x86-based supercomputing technology codenamed Knights Ferry (KF).
Like Larrabee, KF seeks to bring together massive parallelism with the flexibility of Intel’s x86 architecture. While Intel has put its previous development of Larrabee graphics cards for consumers on hold, the chip maker has kept working on Larrabee technology for HPC users.
Meanwhile, Intel is planning a commercial multicore chip called Knights Corner (KC), also for acceleration of HPC apps. Although KF will offer 32 Intel processing cores on a single chip, KC will reportedly possibly scale to up to 50 of these cores using Intel’s P1270 22nm process.
However, Intel has stated that both KF and KC will complement rather than replace traditional Xeon processors, with “the vast majority of workloads will still [running] best on Intel Xeon processors.”
After releasing specs for KF is August, Intel demo’d use of the new MIC development platform in a series of HPC apps at the SC10 trade show earlier this month.
In a demo jointly produced with the Mayo Clinic, for example, Intel showed how KC can used compressed signals to quickly create high quality medical images, cutting down on the time patients need to spend on getting MRIs done.
To illustrate how apps for standard Intel CPUs can scale to future MIC architectures, Intel generated the Monte Carlo apps using standard C++ code with an Intel MIC-enabled version of Intel Parallel Studio XE 2011 development tools.
Intel, though, did not specify the raw computing performance achieved by KF during the demos at SC10 in GigaFLOPS or TeraFLOPs.
KF is reportedly drop-in compatible with an Intel Xeon chip, without any need to reprogram app code in another development language.
Intel released the specs on KF in conjunction with shipment of the development platform to some HPC customers. Knights Ferry can’t really be described as a GPU, since it is based on x86 cores. However, the card features a standard PCI-Express 2.0 slot, a dual-slot heatskink, and up to 2GB of GDDR memory
Each of the 32 cores on the KF chip features quad threads, a 32KB L1 instruction cache, and a 32KB L1 data cache. Each core also contains a 256KB LB cache, for a total LB cache of 8MB on the chip. According to some experts, this is much more cache than many highly parallel apps even need.
In addition, each processor has a very wide 512-bit vendor unit permitting performance of 16 single-precision floating point operations n a single instruction, with double-precision floating point operations producing half throughput value.