At the supercomputing conference SC2011 today, Intel offered up performance details of its upcoming Xeon E5 processors and demoed their Knights Corner many integrated core (MIC) solution. The new Xeons won't be broadly available until the first half of 2012, but Santa Clara has been shipping the new chips to "a small number of cloud and HPC customers" since September. The new E5 family is based on the same core as the 3960X Intel launched
yesterday, but the company has been surprisingly slow to ramp the CPUs for mass production.
Rajeeb Hazra, general manager of the Intel Datacenter and Connected Systems Group, stated that demand for the new chips was stronger than anything Intel had seen before. "Customer acceptance of the Intel Xeon E5 processor has exceeded our expectations and is driving the fastest debut on the TOP500 list of any processor in Intel's history," said Hazra. "Collecting, analyzing, and sharing large amounts of information is critical to today's science activities and requires new levels of processor performance and technologies designed precisely for this purpose."
Intel may have run into problems with ramping the E5 earlier this year, but benchmarks indicate that the end result has been a remarkably solid CPU. The E5, while important to Intel's overall server lineup, isn't as interesting as the public debut of Knights Corner. Some of you may recall that Intel's canceled GPU (codenamed Larrabee) found new life as the prototype device for future HPC accelerators and complementary products.
According to Intel, Knights Corner packs 50 x86 processor cores into a single die built on 22nm technology. The chip is capable of delivering up to 1TFlop of sustained performance in double-precision floating point code. Intel hasn't released many details at this point, but the company's presentation indicates that the cores operate at 1-1.2GHz. Nvidia's current high-end M2090 Tesla GPU, in contrast, is capable of just 665 DP GFlops. Intel took the time to highlight just how far computing performance has advanced in the past 15 years by comparing the first 1 TF system, ASCI RED, with the performance of a single Knights Corner co-processor. We don't know how much power the new cards will draw, but given that the 1997-era supercomputer could reportedly suck down 850kW at maximum load, the power consumption comparison should be pretty good.
Hazra states that Knights Corner is uniquely easy to develop for. Unlike other co-processors, the MIC is fully accessible and programmable as though it were a fully functional HPC node. Existing x86 applications will run on the card without any need to recompile them in other languages.
Knights Corner isn't shipping yet, but Intel has put the writing on the wall as far as Nvidia is concerned. Once the card starts shipping, we'll soon find out if Nvidia's first-mover advantage and years of work with high-end vendors withstands Intel's superior manufacturing technology and longstanding OEM relationships.