April 15, 2013

Future Challenges of Large-Scale Computing

Nicole Hemsoth

Now in its 28th year, the International Supercomputing Conference, ISC’13, is fast approaching. On Monday, June 17, Bill Dally, chief scientist at NVIDIA and senior vice president of NVIDIA Research, will deliver the opening keynote, titled “Future Challenges of Large-Scale Computing.”

Dally will address the multiple advances that will be necessary in order for the community to achieve the potential of HPC and data analytics going forward. The thrust of his talk will be on the challenges around power, programmability, and scalability, and most notably the role that energy-efficiency will play in determining system performance.

ISC’13 will be held from June 16-20, 2013, at the Congress Center Leipzig (CCL) in Leipzig, Germany.

In this Q&A Mr. Dally shares his views on where HPC is headed in the context of such important topics as heterogenous computing, the memory wall, government belt-tightening, and more…

HPCwire: Are different types of workloads, such as big data, HPC and Web 2.0, beginning to demand different types of processors? Will server processors diversify over the next five to ten years, or will they converge?

Bill Dally: HPC, Web servers, and big data all have similar requirement for processors. Within these applications there are program segments that are limited by single-thread performance and other segments that are limited by throughput. To meet this need, there will be a convergence on heterogeneous multicore processors where each “socket” will contain a small number of cores optimized for latency (like today’s CPU cores) and many more cores optimized for throughput (like today’s GPU cores).

HPCwire: The increase in processor performance seems to be outpacing memory technology. What can be done about the memory wall?

Dally: There are three aspects of memory relevant here: bandwidth, latency and capacity. To address the slow scaling of memory bandwidth we plan to move to memory technologies that involve placing memory dice on the same package as the processor chip and connecting them with very high-bandwidth, low-energy links. This on-package memory technology will enable us to scale memory bandwidth with processor performance holding the Byte/FLOP ratio roughly constant for the next few generations.

Memory latency is remaining roughly constant as processor performance increases. We deal with this by increasing parallelism to hide the latency. With adequate parallelism we can keep the memory pipeline full – using all of the available bandwidth.

Memory capacity is largely a matter of cost. The challenge here is that high-bandwidth memories, like on-package or stacked DRAM, cost significantly more than commodity memory. Thus, for cost-sensitive applications we are likely to see a two-tiered memory system with a moderate capacity, high-bandwidth on-package memory and a high-capacity commodity memory. A non-volatile memory technology like flash or phase-change memory could have a place in such a hierarchy as well.

HPCwire: How important will 3D stacked chip technology be to processors and memory? When do you think we’ll see the first commercial products?

Dally: Placing memory on-package will be critical to scale bandwidth. Stacking technology is important to extend the capacity of this high-bandwidth memory.

Stacked memory is shipping today. However, most of this today uses wire bonds, not through-silicon vias. At the 2013 GPU Technology Conference (GTC) this past March, we announced that we expect to introduce stacked memories with our Volta architecture-based generation of GPUs – in about 2016.

HPCwire: Government austerity restrictions look as though there could be pressure to reduce investments in exascale computing, especially in the US. How capable is industry of driving these initiatives by itself?

Dally: Industry will continue to move forward on its own on exascale projects, however progress will be much slower than with government assistance.

It is disappointing that government priorities are such that investment in computing innovation is being scaled back. At the same time, other nations like China are investing heavily in computing. Even the EU with all of its economic problems is moving forward with their exascale program. With reduced investment, the US runs a grave risk of giving up its leadership in computing.

HPCwire: With current technology, it seems as though exascale computing would require so much energy as to render it impractical. Will we see new breakthrough technologies to sufficiently reduce power consumption to make exascale practical and affordable?

Dally: Improving energy efficiency to reach the goal of a sustained exaflops on a real application in 20MW is a significant challenge. However, I am optimistic that we can meet this challenge. There are many emerging circuit, architecture and software technologies that have the potential to dramatically improve the energy efficiency of one or more parts of the system. For example, at NVIDIA we have recently developed a new signaling technology that reduces the energy required by communication by more than an order of magnitude, and we have developed an SRAM technology that permits operation at dramatically lower voltages – and hence lower power. It won’t be a single breakthrough technology that will get us to the exascale energy goal, it will be multiple breakthroughs – at least one in each of the multiple areas that require improvement – processor, communication, memory, etc. We have a number of research projects that are targeted at these different areas. If a sufficient number of these projects have successful outcomes, we will meet the goal.

These improvements, however, depend on research, which in turn will be slowed considerably without government funding.

HPCwire: What do you see as the biggest challenges to reaching exascale?

Dally: Energy efficiency and programmability are the two biggest challenges.

For energy, we will need to improve from where we are with the NVIDIA-Kepler-based Titan machine at Oak Ridge National Laboratory in Tennessee, which is about 2GFLOPS/Watt (500pJ/FLOP) to 50GFLOPS/Watt (20pJ/FLOP), a 25x improvement in efficiency while at the same time increasing scale – which tends to reduce efficiency. Of this 25x improvement we expect to get only a factor of 2x to 4x from improved semiconductor process technology.

As I described before, we are optimistic that we can meet this challenge through a number of research advances in circuits, architecture and software.

Making it easy to program a machine that requires 10 billion threads to use at full capacity is also a challenge. While a backward compatible path will be provided to allow existing MPI codes to run, MPI plus C++ or Fortran is not a productive programming environment for a machine of this scale. We need to move toward higher-level programming models where the programmer describes the algorithm with all available parallelism and locality exposed, and tools automate much of the process of efficiently mapping and tuning the program to a particular target machine.

A number of research projects are underway to develop more productive programming systems – and most importantly the tools that will permit automated mapping and tuning.

Changing a large code base, however, is a very slow process, so we need to start moving on this now. As with energy efficiency, progress will be slowed without government funding.

About Bill Dally

Bill Dally is chief scientist at NVIDIA and senior vice president of NVIDIA Research, the company’s world-class research organization, which is chartered with developing the strategic technologies that will help drive the company’s future growth and success.

Dally first joined NVIDIA in 2009 after spending 12 years at Stanford University, where he was chairman of the computer science department and the Willard R. and Inez Kerr Bell Professor of Engineering. Dally and his Stanford team developed the system architecture, network architecture, signaling, routing and synchronization technology that is found in most large parallel computers today.

Dally was previously at the Massachusetts Institute of Technology from 1986 to 1997, where he and his team built the J-Machine and M-Machine, experimental parallel computer systems that pioneered the separation of mechanism from programming models and demonstrated very low overhead synchronization and communication mechanisms. From 1983 to 1986, he was at the California Institute of Technology (Caltech), where he designed the MOSSIM Simulation Engine and the Torus Routing chip, which pioneered wormhole routing and virtual-channel flow control.

Dally is a cofounder of Velio Communications and Stream Processors. He is a member of the National Academy of Engineering, a Fellow of the American Academy of Arts & Sciences, a Fellow of the IEEE and the ACM. He received the 2010 Eckert-Mauchly Award, considered the highest prize in computer architecture, as well as the 2004 IEEE Computer Society Seymour Cray Computer Engineering Award and the 2000 ACM Maurice Wilkes Award. He has published more than 200 papers, holds more than 75 issued patents and is the author of two textbooks, “Digital Systems Engineering” and “Principles and Practices of Interconnection Networks.”

Dally received a bachelor’s degree in electrical engineering from Virginia Tech, a master’s degree in electrical engineering from Stanford University and a PhD in computer science from Caltech.

Share This