Fast Rewind: 2016 Was a Wild Ride for HPC
Some years quietly sneak by. 2016, not so much. It’s safe to say there are always forces reshaping the HPC landscape but this year’s bunch seemed like a noisy lot. Among the noisemakers: TaihuLight, DGX-1/Pascal, Dell EMC & HPE-SGI et al., KNL to market, OPA-IB chest thumping, Fujitsu-ARM, new U.S. President-elect, BREXIT, JR’s Intel Exit, Exascale (whatever that means now), [email protected], whither NSCI, Deep Learning mania, HPC identity crisis…You get the picture.
Far from comprehensive and in no particular order – except perhaps starting with China’s remarkable double helping atop the Top500 List – here’s a brief review of ten 2016 trends and a few associated stories along with occasional 2017 SWAGs.
1. Make Way for Sunway
China’s standing up of the 90-plus Petaflops Sunway (TaihuLight) on the June Top500 List caused a lot of consternation outside of China including too many dismissive comments: It was this. It wasn’t that. It’s a trick. Never happened. I like Thomas Sterling’s characterization in his ISC16 closing keynote. To paraphrase (and with apologies to Thomas) – Get over it and enjoy the engineering triumph for a moment.
Are there legitimate questions? Sure. Nevertheless, built with homegrown processors as well as other components, TaihuLight’s marks China’s breakout in terms of going it alone. The recent U.S. election seems poised to heap more fuel on China’s inner fire for self-control over its technology destiny. China’s Tihan-2 (34 Pflops) is still solidly in the number two spot. China and the U.S each had 171 computers on the most recent Top500, the first time for such parity. HPCwire managing editor Tiffany Trader’s piece on China’s standing up Sunway looks at the achievement and her second piece from SC16 looks at the new rivalry (links below).
2. GPU Fever
In April, NVIDIA announced its latest and greatest GPU, the impressive P100 (Pascal architecture) and its DGX-1 development box. It promptly stood up a clustered version of the latter – the DGX SATURNV (124 DGX-1 servers, each with eight P100s) – that placed 28th on the November 2016 Top500. Clearly the rise of GPU computing is continuing and gaining strength. Indeed the rise of accelerator-assisted heterogeneous computing is broadly gaining momentum. The SATURNV was also remarkably energy-efficient at 8.17 gigaflops/watt.
Given the rapid emergence of deep learning in all of its forms – and yes training is generally quite different than inferencing – it seems very likely that the GPU craze will continue, including various mixed-precision offerings for better DL performance. Intel, IBM, ARM, pick-your-favorite company, are all in the accelerated computing hunt. FPGAs. SoCs. Soon neuromorphic options. Right now, NVIDIA is the clear leader. Here are two looks at both the Pascal and DGX-1
3. Who’s Who in HPC?
This is a question, not the usual introduction to a list of prominent players. Mergers and acquisitions are reshaping the technology supplier world (again) and at the largest scales. Does that mean we’re in a period of consolidation? HPE buys SGI. Dell "buys" EMC. SoftBank acquires ARM. Mobile chip king Qualcomm’s $39B purchase of automotive chip supplier NXP is being billed as the largest semiconductor deal ever. (One wonders how Cray will fare in a land of Giants. Cray has cutting edge supercomputing technology – is that enough?) There were many more deals but here’s a look at three majors.
Let’s start with HPE-SGI. HPE, of course, is itself the result of venerable Hewlett-Packard’s split-up a year ago. SGI was kept alive by a Rackable deal some years ago. Pairing the two creates a giant with strength at all levels of HPC with HPE seeking, among other things, to leverage SGI’s shared memory technology, enterprise Hanna platforms, and high-end supercomputer offerings. We’ll see. Here are two looks at this most recent of deals (closed in November).
Dell EMC. Has a certain ring to it. Dell, now private, is looking to leverage scale – always a Dell goal – as well as add technologies needed to offer complete compute-storage solutions. Both Dell and EMC were already juggernauts and they now declare that Dell EMC’s combined revenues make it the top provider (revenue) in HPC, displacing HPE (servers.) Like most system suppliers, Dell EMC is increasingly focused on “packaged HPC solutions”, to the extent such a thing is possible, to help drive advanced scale computing adoption in the enterprise. Combined revenue is expected to be in the $70-plus billion range.
Here are two brief articles examining Dell EMC plans.
Perhaps more surprising was the SoftBank acquisition of ARM. The worry was SoftBank might meddle in ARM’s openness but that doesn’t seem to be the case. Indeed, ARM, like others, seems to be gaining some momentum. Fujitsu, of course, is switching from Sparc to ARM for its post K computer and around SC16 ARM announced it is being supported in the latest OpenHPC stack. Then just a week or so ago, ARM announced plans to purchase toolmaker Allinea. The latter suggests SoftBank is making good on its promise to infuse new resources.
On the technology front, ARM introduced its Scalable Vector Extensions (SVE) for HPC in August which provides the flexibility to implement vector units with a broad choice of widths (128 to 2048 bits at 128-bit increments); applications can be compiled once for the ARMv8-A SVE architecture, and executed correctly without modification on any SVE-capable implementation, irrespective of that implementation’s chosen SVE vector width. This is another arrow in ARM’s HPC quiver and integral to Fujitsu’s plans to use the processor.
Market signals from ARM chip suppliers have been a bit more mixed and it will be interesting to watch ARM traction in 2017, not least in China. Here are three articles looking at ARM’s progress and that SoftBank purchase.
4. Containers Start Penetrating HPC
This is perhaps not an earth-shaking item, but nevertheless important. Virtualization technologies are hardly new and containers, mostly from Docker, have been sweeping through the enterprise. HPC has taken the hint with a pair offerings introduced recently – Singularity and Shifter – which may help individual researchers and small research teams gain easier access to NSF resources.
Shifter was developed first (2015) at NERSC for use with Edison supercomputer. Singularity (2016) came out of Lawrence Berkley National Laboratory this year and has aspirations for making the use of containers possible across many leadership class computers, such as TACC and SDSC for example. Gregory Kurtzer (LBNL) is Singularity’s leader. The idea is simple but not so easy to execute on leadership class computers: create ‘containers’ in which researchers can put their complete application environment and reproducibly run it anywhere Singularity is installed. Adoption has been surprisingly fast says Kurtzer.
Here’s an interview by Trader with Kurtzer from this past fall on Singularity progress as well as a backgrounder on Shifter.
5. Deep Learning – The New Everything
Earlier this year, Google’s DeepMind’s AlphaGo platform defeated one of the world’s top GO players. Just training the system took a couple of years. No matter. Expectations are crazy high for DL and machine learning re: autonomous driving, precision medicine, recommender systems of all variety, authentication and recognition (voice and image), etc. We are clearly just at the start of the DL/ML era. Let’s leave aside true AI for the moment and all the scary and hopeful voices surrounding it.
Deep learning and machine learning are doing productive things now. Sometimes it’s as simple as complicated pattern recognition (I enjoyed writing that). Other times it is blended with traditional simulation & modeling applications to speed and refine the simulation computation. In many cases it’s the only way to realistically make sense of huge datasets.
The DL language is all over the place. Vendors and users alike seem to have nuanced preferences for AI over DL over ML over cognitive computing. There’s also a growing froth of technologies competing for sway: GPUs, FPGAs, DSPs, brain-inspired architectures. If SC15 last year seemed like a loud pivot towards driving HPC into the enterprise, SC16 was nearly totally focused on DL and data analytics of one or another sort – chip makers, systems builders, job schedulers and provisioning tools all seemed to be chanting the same (or at least similar) verse.
Here are six articles from the 2016 archive on various aspects of DL (use, technology, etc.).
6. Murky Public Policy
So much of HPC depends on public policy and government funding. The change in administration muddies what was already pretty murky water. One example is the tentative new Secretary of Energy Rick Perry’s various slips of the tongue over what the agency is and expressing a desire to abolish DoE. Hmm. Not sure he had the agency’s full compass in mind when he spoke. Anyway, conventional wisdom is Washington’s emphasis will shift from R&D to decidedly D and defense-related.
One initiative that’s been hanging in the wind through 2016 is NSCI, the National Strategic Computing Initiative. Greeted with enthusiastic fanfare at its inception by Presidential Executive Order at the end of July 2015, later regarded somewhat skeptically because inaction, and now mostly reborn as an umbrella label for pr-existing programs – it is easy to wonder if NSCI will survive. No doubt many of its pre-existing components will and one hopes several of the newer goals such as HPC workforce expansion also will have legs (and funding).
The DoE Exascale project – which now and perhaps rightfully takes pain to set useful exascale level computing rather than a numeric LINPACK score as its goal – has speeded its schedule somewhat. The first machine is now expected in 2021. There’ve been suggestions more changes are coming, so it is perhaps premature to say much. Maybe more funding or urgency will follow given the international landscape seems increasingly dominated by nationalistic agendas versus regional cooperation. China’s ascent may also trigger more HPC R&D spending…or not.
The extent to which NSCI frames broad goals for HPC advancement make it an interesting, though perhaps fading, indicator and wish list. Here are five articles looking at NSCI and the U.S. Exascale Project including a recent interview with Paul Messina, director of U.S. Exascale efforts.
7. How’s ‘HPC’ Business
While we’re talking policy, it’s worth looking at the business (and broad technology) climate. IDC is one of the key keepers of the notes here – remember the saying that he/she who keeps the notes has power – and IDC reports 2016 was a good year and 2017 looks better. 2016 clocked in at roughly $11.6B (server revenue) and 2017 projected at $12.5B, a nice growth.
Speaking at the SC16 IDC breakfast update, Bob Sorensen, VP Research, noted “HPC universe is expanding in ways that are not directly observed…because we haven’t quite decided what the definition of HPC should be.” He identified work being done with new hardware and software for deep learning. “From the training phase, the computationally intensive part where you go and train a deep neural network to understand a tough problem – exaflops regression kind of training,” involving GPUs, (Xeon) Phis and FPGAs.
The times they are a changin’. Here are IDC’s two market updates from SC16 and ISC16, along with a ten-year retrospective from Addison Snell, CEO, Intersect360 Research, which notched its ten-year anniversary right around the SC16 timeframe. (Congrats Addison!)
8. Surveying the Tronscape
Turning from business to the technology landscape, it's worth noting there are few speakers better able to capture and prioritize the breadth of HPC technology at any given moment than Thomas Sterling, director of CREST; he would demur from any such notion. His annual closing keynote at ISC is substantive and entertaining and perhaps casts a somewhat wider and more technology-in-the-trenches net than the items called out here earlier.
Here’s an account of ISC16 Sterling’s talk. It’s a fast read and well worth the effort. Typically Thomas, he touches on a lot of ground and with clarity.
9. OpenPOWER Pushes Toward Liftoff
2017 has to be IBM's and OpenPOWER’s lift off. Last year HPCwire put a rather harsh lens over the effort describing its many challenges:
“2016 promises to be pivotal in the IBM/OpenPOWER effort to claim a non-trivial chunk of the Intel-dominated high-end server landscape. Big Blue’s stated goal of 20-to-30 percent market share is huge. Intel currently enjoys 90-plus percent share and has seemed virtually unassailable. In an ironic twist of the old mantra ‘no one ever got fired buying IBM’ it could be that careers at Big Blue rise or fall based upon progress,” I wrote.
There have been several Power8 and Power8+ (NVLink) systems available. IBM has optimized one system (PowerAI) for DL/ML, everyone’s darling target, and worked with HPC cloud provider Nimbix to put the latest Power technology in its cloud including tools to make using it easier. The Power9 roadmap has been more fully described and first Power9 chips are expected this year, including support of the CORAL project. There are quite a few more items checked off on the IBM/OpenPower done list.
HPCwire will again review IBM/OpenPOWER’s progress early in 2017, but based on a conversation with Ken King and Brad McCredie at SC16, the pieces of the puzzle – technology, product, channel, growing interest from hyperscales, lower price point product from partners, continuing investment in advances such as OpenCAPI – are in place. Sales are what’s needed next. Below is a link to last year’s broad article and one describing the PowerAI offering.
10. Marvelous Marvin Remembered
Before ending it is good to recall that last January (2016), artificial intelligence pioneer Marvin Minsky died at age 88 – a sad way to start a year so thoroughly dominated by discussion around AI precursors deep learning, machine learning, and cognitive computing (whatever your preferred definition).
The New York Times obituary by Glenn Rifkin is well worth reading. Here’s a brief excerpt: “Well before the advent of the microprocessor and the supercomputer, Professor Minsky, a revered computer science educator at M.I.T., laid the foundation for the field of artificial intelligence by demonstrating the possibilities of imparting common-sense reasoning to computers.”
What would an end-of-year article be without a couple of plaudits, re: Bill Gropp, was elevated to Acting Director, NCSA, and also won the 2016 Ken Kennedy award. Well deserved. NCSA celebrated turning thirty. James Reinders left Intel making us all wonder where he will reappear and in what capacity. Here’s his brief farewell published in HPCwire. There are many more but we’ll stop here.
Quantum and neuromorphic computing efforts kept gaining momentum. Two big neuromorphic computing systems were stood up in the spring, and IBM sold a TrueNorth-based system to LBNL for collaboration. The quantum picture still seems a bit fuzzy to me, but Bo Ewald shows no slowdown in evangelizing (article) and Los Alamos National Lab is racing to develop a broader range of applications for its D-Wave machine. The Center for Evaluation of Advanced Technology (CENATE) at PNNL ramped up fast (update here) and will hold its first workshop in 2017.
Less of a bonus trend and more under the expected label, Knights Landing (KNL) systems began cropping up everywhere in the second half of the year (Intel’s SC16 recap). The OmniPath-InfiniBand competition continued in force. Seagate continued its drive into HPC; DataDirect Networks maintained its strength at the high-end, continued its push into the enterprise, and adopted software defined strategy with vigor. OpenHPC delivered version 2.0 of its open source HPC stack and tools and now supports ARM as well as x86. One wonders if IBM will take the plunge.
There are many more important trends warranting notice here – the growth of NSF cyberinfratructure for example – but sadly my deadline is here too. See you in 2017.