Sign up for our newsletter and get the latest HPC news and analysis.

Hengeveld: An Industry Perspective on SC11 and Disruptive Innovation

In this special guest feature, Intel’s John Hengeveld reflects on the amazing week that was SC11.

There is so much to cover from SC11.  It was a thrilling week of meetings, technical sessions, and new technology.  I learned a lot, and appreciate how much of great and exciting ride we are in for in the years to come.  The key things I was looking for from my preSC11 column are shown here.

  • New CPUs and the Top500: Interlagos — the AMD Opteron 6200 launched on Monday with a focus on core count and power efficiency per core.   Intel made a press announcement of the performance levels of the future Intel Xeon E5 family as shown on the top500, and further announced that Xeon E5 will support PCIe 3.0.  The top500 list had listings from each of these new CPUs including Cray systems with AMD processors, and HP, Bull and Appro systems with Intel Processors.  At the end of the day, with banner products from both vendors, the industry is set up for a fresh push forward.
  • New Big Systems and new systems across the Globe: While the #1 system surged to over 10PF, the top10 remained unchanged.  I heard about some systems in development that will be coming soon (GENCI Curie, LRZ, Titan) but it was the news that there were not any new top10 systems that surprised me.   More interesting is this occurred while the top500 bottom moved up aggressively from 40.187 TF to 50.94 TF.  Hopefully this pause in the Top10 is a breather while we wait for new systems.
  • GPU vs. Intel MIC part 4: Kepler – Was I the only one disappointed by Jen-Hsun Huang’s keynote? We heard nothing further on Kepler. He struggled to avoid saying Intel and ARM (at one point smoothly saying NVidia when he meant Intel).    He made a case that exascale in 2020 at 20MW is a key goal and that lower power solutions would be required to get there.  But the substance of his talk jumped off of Clayton Christensen’s keynote from last year and talked about the “Innovators Dilemma” on the path to Exascale.

I taught corporate strategy at Portland State University for many years and have often taught the key insights in “The innovators Dilemma” and “The Innovators Solution.”   So I was emotionally connected when Huang started out there.  The key principles of how a “low end disruptor” captures the mainstream of the market with a lower cost “good enough” solution are valuable insights to the technology world in general and HPC in specific.

The point to the referencing Christensen – Was trying to illuminate that GPUs would represent a disruptive innovation for the mainstream of HPC. While he very clearly made the case that NVidia graphics accelerators were once a new market disruptor in gaming and a low-end disruptor in workstations. Where he went off target is trying to stretch that into the HPC space.

At 31 minutes in the keynote, Huang says: “If I can just figure out how to program it, if I can describe all my problems as a triangle. I could solve the worlds problems.”

In this comment, Huang admits that adoption of GPU technology is predicated on adopting an isomorphism.  The programming model of GPU is the transformation of a problem into manipulations of triangles… and thus the debate in the industry now.

The issue in HPC is not “does the industry need the density of performance at lower power” (we do), but rather “must we adopt the isomorphism of thinking of the world as triangles to use the system with highly efficient performance levels”.  This is the core of the GPU vs. Intel MIC architecture debate.

The substance of the keynote was demonstrations of the impact of increased compute density on gaming examples (BF3, Assassin’s Creed, etc) and a plug for his Maximus WS product, which while fun for the gamer in all of us left us feeling a bit… hollow.

MIC:  1TF per socket, I got excited when I found out the Knights Corner silicon would be powered on by SC11 and deeply hoped the gnomes working on it would be able to run Linpack or DGEMM on it by that time.  My friend and boss Joseph Curley pressed the team to complete the demo on time.  We made it, you can see in the picture of Joe with Raj Hazra (the GM of Intel’s Technical Computing Group) proudly holding up one of the first Knights Corner parts in the picture below.  Joe is looking stern in the picture here, no doubt from exhaustion.  He’s been a busy man of late.

The more interesting element of Raj’s talk was the discussion that an Intel MIC product appears to applications as a fully functional compute node able to run its own open source operating system.  This means that many applications will port to MIC with a simple recompile.  Robert Harrison stood up and presented results from porting “10’s of millions of lines of code” to the Intel MIC software development vehicle.

So the GPU vs. MIC debate is engaged in full force.  NVidia and Intel are now mostly publically aligned on the goal, 20MW / Exaflop in this decade.  The debate on performance is over; the debate on programming has begun.

  • PCIe 3: Mellanox announced in their quarterly earnings release some news on their solutions for Infiniband. A few of the new systems in the top500 had Mellanox IB solutions.  Intel announced that their future Xeon E5 processor integrates PCIe 3.0 on die.  No word from AMD on this, and no announcements from other graphics cards or interconnect suppliers. The beginning of the PCIe 3.0 transition is here.  Interconnect bandwidth is going to be a key element in delivering performance in some cluster architectures and in some key workloads.  Again, I expected more from other IB manufacturers.  This transition will accelerate in greater force in the first half of 2012.

** Footnote: A sustaining innovation is the opposite of disruption.  It is the normal progress of maturing of a technology to be accessible to a larger portion of an overall market space.  Unmet needs are now met such that customers see higher value in the product they are using and either pay more for it or buy more of it.  Adding a feature to a product like the iphone4S speech recognition feature is an example.

Comments

  1. Jason Riedy says:

    Um, the goal of an exaflop at 20 MW was set *for* the companies and not by the companies.

  2. John Hengeveld says:

    Intel has adopted 20MW, 2018 1Exaflop as its internal goal.
    nvidia stated its goal was 20MW, 2020… in the keynote.

    Various players can set a bar at a certain level.. when companies like intel and nvidia publically sign up… it counts.

  3. While I agree GPUs are programmed differently for compute than Intel’s MIC I disagree with the assertion that you must think about triangles when programming. OpenCL and CUDA bypass the graphics pipeline.

Resource Links: