HP's HPC Server Refresh Hints at Exascale Future

Will Enterprise 2020 be hosting a substantial portion of its workloads on GPUs or on other non-traditional architectures?

High-performance servers that use graphics processing units (GPUs) to crunch computationally-intensive workloads are far from a mainstream proposition. That may be about to change soon.

Several vendors ship GPU-based systems, although most (if not all) of these offerings are slated for niche markets, such as high-performance computing (HPC) or for computationally-intensive applications in specific vertical industries.

Hewlett-Packard Co. (HP) has unveiled its densest GPU-based system to date, the SL390s G7. The new G7 updates an existing product (the SL390s) that leverages the Compute Unified Device Architecture (CUDA), a framework developed by Nvidia to host non-graphics workloads on its GPUs.

Although the first-generation SL390s integrates three Nvidia 260 or 270 GPUs onto a single system board, the revamped SL390s G7 integrates eight GPUs. It retains the three discrete x16 PCI-Express (PCI-E) lanes that HP first implemented in its predecessor. This implementation -- which has three times the standard number of PCI-E lanes -- is constrained by chipset "physics," HP officials say.

The new SL390s ships as part of a refresh of HP's ProLiant SL6500 system line. HP also announced the ProLiant SL160S G6 (a dual-socket Intel-based system that supports up to 192 GB of RAM); the ProLiant 165S G7 (a dual-socket AMD-based system that supports up to 288 GB of RAM); and the SL170S G6: similar to the ProLiant 160S, albeit in a half-width configuration.

HP's SL6500-Series deliverables are designed primarily for HPC and for specific applications in niche verticals. Potential applications include not just the obvious (seismic analysis in oil and gas; derivatives crunching in financial services), but the not-so-obvious (e.g., Facebook-sized Web server farms), too. HP's existing SL390s has sold almost exclusively into these verticals.

Officials expect that to remain the case with the G7, but Glenn Keels, director of marketing with HP's worldwide service provider and HPC business unit, thinks the "mainstreamification" of GPU-based processing is only a matter of time.

This will probably become a fait accompli once the industry transitions from petascale to exascale computing. "We see the GPU architectural transformation as being a key architectural transformation. One of those transformations that creates a tipping point in the industry," he explains.

Scientists have suggested that the first exascale-like offerings won't appear before 2018. Keels doesn't dispute this projection and concedes that offerings such as the eight-GPU SL390s G7 are part of the "bleeding edge." At the same time, he says, HP has already GPU-enabled more than a dozen of its system offerings, in addition to several of its key software offerings.

"It's really very early on, but we … already [have] 14 GPU-enabled platforms. Our Cluster Management utility is GPU-aware. We're … the number one OEM supplier of Nvidia technology in servers," says Keels.

While the usual suspects (HPC, oil and gas, financial services) will be among te first to invest in GPU-based technologies, Keels expects other verticals (such as Web server farms) to follow suit. From there, GPU- and other (still emerging) non-traditional architectures will be exploited across all verticals.

"The tipping point is first going to be happening in high-performance computing, but we're also seeing the same kind of approach, [with an] architectural transition happening in the Web space. It isn't just with GPUs, and it's not just CUDA," he indicates.

"I'm not in a position at this point to share what our architectural strategy is for these alternative processors, but realize that GPUs are on the bleeding edge of this architectural tipping point, this transition. It's not crossing the [exascale] chasm yet, but as the price/performance per-watt, per-square-foot improves, it's showing the way."