Intel x86 processors continue to dominate HPC servers while the number of cores per processor also keeps rising, perhaps no surprises there. Also somewhat anticipated, the amount of memory per core, per processor, and per node is rising. These are the top line results of Intersect360 Research’s latest HPC sites survey on processor use.
A bit more intriguing were findings that ARM is generating some early interest – “for the first time in our surveys we note the appearance of an HPC system based on the ARM architecture” – and that accelerators share of systems market slipped very slightly after four years of steady growth. The latter may just be a pause. Nvidia (NASDAQ: NVDA) still owns the market but Intel (NASDAQ: INTC) is making inroads. Intersect360 does suggest the high tide of standardized processors may be starting to ebb.
“Rather than just look at microprocessors, you have to look at all of the different multi-core and manycore options. The majority of that has been Nvidia GPUS so far. We’re now seeing close to half of the HPC market having at least some accelerator or manycore component attached. When we get to [Intel] Knights Landing that’s not strictly just an accelerator. There’s going to be a great deal of diversity in the market, potentially a move back toward specialization away from what you might have called standardization or commodity,” said Intersect360 Research CEO Addison Snell.
- Brand disloyalty – Despite Intel’s growing share of the x86 processor market in HPC (and elsewhere), we believe customers still treat these components as commodities, and will switch vendors when price/performance for their particular applications warrants a switch.
- Multi-core processors – The single-core processor is nearly extinct. The most common processor in HPC is the eight-core CPU, but it is unclear what will be the most popular configuration in the future.
- Memory requirements – Memory per core continues to rise. This could be a response to alleviating the more significant bottleneck in external storage or as a result of running applications with larger data sets.
- Accelerators – No longer an afterthought for HPC users, accelerators are being used for production work across an array of applications. We expect that within a year or two the majority of new systems will be equipped with accelerators of some sort.
This latest Intersect360 Research report, HPC User Site Census: Processors, released late in October, examines a wide range of processor deployment by system type (HPC clusters, SMP, MPP, servers) as well as trends in memory and accelerator use. The most recent surveys were conducted in the second quarter of 2015. In addition, Site Census data obtained in 2013 and 2014 was used to aggregate additional information across all sites. The report presents results from a total of 781 HPC systems at 453 sites.
Despite the slight dip in accelerators’ share of systems, they are now entrenched in HPC according the report: Current offerings “from Nvidia, Intel AMD can deliver well over a peak teraflop of double precision floating point performance, delivering more than five times the peak performance of a high-end Power or x86 CPU. FPGA performance is more difficult to measure, although in recent years, much more powerful implementations (some with conventional ARM processors embedded into the chip) have become available. According to our surveys, approximately a third of HPC systems operating today contain some sort of accelerator,” reports Intersect360.
About half of all accelerated systems today have 20 or more accelerators, notes Interset360; two years ago the number was ten. Intersect360 contends this overall growth indicates more codes are able to take advantage of accelerator technology and likely being run in production environments, rather than for just test and evaluation.
The other main accelerator findings reported include:
- Nvidia remains the dominant supplier, with a footprint in 77.8% of accelerated systems. That represents a slight reduction of market share from two years ago, when it supplied 85.3% of those systems.
- Intel is the second most popular supplier with 11.3% of accelerated systems, a three-fold increase from our 2013 data. Given that its Xeon Phi coprocessor line has only been available since 2012, Intel has managed to capture a significant share of the accelerator market in a relatively short period of time.
- AMD (NASDAQ: AMD) has captured 1.8% of the market, compared to 0.5% in 2013. A renewed focus on its FirePro server GPU line over the last two years may account for AMD’s larger representation in our latest surveys.
- The share of accelerated systems in those most recently deployed/modified (2014+) declined slightly for the first time in four years.
Intersect360 had predicted Intel’s entrance into market in 2012 would validate accelerator model of computing for HPC and helping diversify the market. That certainly seems to be the case. According the report, early users of Intel’s Xeon Phi platform had interest in accelerators, but resisted adopting GPU technology. Looking ahead, Intersect360 expects Intel to capture additional market share from Nvidia over the next two years.
Increasing use of memory at the node, processor, and core levels remains trend. Part of the reason, suggests Intersect360, is the desire to relieve IO bottlenecks. Likewise, the sheer growth in the number of cores per processor is driving more memory use – a trend that seems likely to continue.
For more about the processor report, visit Intersect360 Research at: http://www.intersect360.com/industry/reports.php .
Intersect360 also released its latest interconnect report at roughly the same time. Here’s a link to it: http://www.intersect360.com/industry/reports.php?id=130