Back in 1997, Intel’s ASCI Red system stole the show for the 9th iteration of the Top500 list with 7,264 Pentium Pros clocked at 200 MHz (and later outfitted with Pentium II OverDrives at 333 MHz) and peak Linpack performance of 1.068 teraflops.
It took the chipmaker another 32 Top 500 rounds to score the primo spot with an all-Intel system–and just like ASCI Red, the new top super Tianhe-2 toppled records, and carved a rut to pave over for next-generation systems. The ASCI Red maintained its position at the top for three years until the IBM ASCI White slithered by it, just as the Tianhe-2 will likely dominate the list for at least two years, barring any major surprises.
Despite the diminished supercomputer focus organizationally at the time with the closure of their Supercomputing Division just before the ASCI Red win, the company has made significant strides to reshape that landscape–efforts that culminated many years later in a system that shattered performance records with 48,000 Xeon Phi coprocessors, 32,000 Xeons, and Linpack performance hovering at the 33.8 mark.
While it’s only the second time the company has fully owned a top system, some argue that this is the beginning of a new trend–one that Intel is spearheading.
IDC claims that the server business for supercomputing will expand by 36 percent from $11 billion to $15 billion over the next 4 years. And as we noted earlier this morning, the Top 500 founders, analyst community and end user groups are seeing massive uptake in accelerator and coprocessor adoption–with strong emphasis on Xeon Phi.
As the Top 500 organizers noted today, 403 systems in the Top 500 are Intel-based. And as Intel reminded, “Of those systems making their first appearance on the list, Intel-powered installations account for 98 percent. The June edition of the list had recorded 11 systems based on the Intel Xeon Phi coprocessor, including the Petaflops class systems like Milky Way 2 (Tianhe-2) at 54.9 PFlops and Stampede at 8.5 PFlops of peak performance.”
On the heels of this, IDC stated today in its brief on HPC coprocessor growth some surprising figures on the proportion of sites employing co-processors or accelerators in their HPC systems. They claim the number jumped from 28.2% in the 2011 version of the study to 76.9% in 2013. They added that coprocessors and accelerators advanced from slightly more than 1% of all processor parts in 2011 to 3.4% in 2013, with Intel Xeon Phi co-processors and NVIDIA GPUs running neck and neck for leadership, and FPGAs in a respectable third-place position.
Neck and neck for a product that’s been available for a relatively short time–at least compared to GPU? This is indeed a development–one that IDC analyst Steve Conway said today was a shocker, even for his group that keeps firm tabs on this market. While a solid number of these are being implemented for exploration only at this point, Intel is poised to take its rivals to task in the years ahead, which will light a fire under NVIDIA, AMD and others to keep pace.
Instead of letting the Tianhe-2 news ride by itself today, Intel kicked out some developments around two aspects of its business for this growing market. In addition to some new families of Xeon Phi that have been bred for various roles, Intel formally announced some news about its second-gen Knights Landing that can do double duty as a coprocessor or CPU while tapping the 14 nm process and their most recent 3D TriGate transistors. This gives users some choice–allowing them to use it as a PCIe card-based coprocessor for offloading Xeon math, or to tap it as a host processor that’s planted in the motherboard socket, doing all the same things one might expect of a primary processor–but this time with a coprocessor kick.
Intel says that when used as the CPU, Knights Landing will chuck some of the data transfer roadblocks users have over PCIe using other accelerators. Further, to target HPC performance, they will significantly increase the memory bandwidth across the Knights Landing kingdom via integrated on-package memory.
Outside of these Knights Landing revelations, they also let fly on their new generation of Phis that offer some alternatives in terms of performance, price, power, form factors and memory with the ultra performance-tuned 7100 family, and the 3100 family, which sacrifices some performance and options in favor of a lower price. They also added a new member to the 5100 family that spawned last year with the new 5120D, which is focused on higher density environments and hooks the sockets to a mini-board to please the blade crowd.
At this point, it looks like they’ve managed to snare some traction at the highest end of the HPC segment via systems like Stampede and Tianhe-2. But with the additional Phi offerings focused on more middle of the road systems, Intel is making a bet that the “missing middle” of users at the lower end of the Top 500 and just off its cliff who have been holding out on adopting coprocessors or accelerators will now have further incentive given more options in terms of price and performance.