Déjà Vu All Over Again

By Nicole Hemsoth

November 16, 2009

Steve Wallach, a supercomputing legend and recipient of the 2008 IEEE Seymour Cray Award, has participated in all 22 supercomputing shows. He is known for his contributions to high performance computing through the design of innovative vector and parallel computing systems. He is co-founder and chief science officer for Convey Computer Corp., a new company with a hybrid-core computer that marries the low cost and simple programming model of a commodity system with the performance of customized hardware architecture.

Never short on opinions, especially when it comes to high performance computing, Steve Wallach talked to HPCwire about the future of HPC and how lessons from the past can point the way for the future.

HPCwire: There’s been a lot of talk about how recent architecture advancements will bring GPU computing into the mainstream for high performance computing with significant speedups and energy savings. You disagree. Why?

Steve Wallach: GPUs are an interesting technology and some applications will probably see significant speed-up, but I don’t see them in the mainstream. Here’s why: programmers will have to put in a lot of effort to get the speed-up. Real-world applications consist of millions of lines of code, and organizations have invested too much money in those programs. If you tell them they have to modify those programs to use your technology, you lose. And it’s not just the software that has to be changed; it is the entire programming eco-structure: debuggers, profilers, and programming memory. Anything that disturbs those underlying realities is destined to become a niche player. This is the biggest difference between an accelerator and a coprocessor. A coprocessor is an extension to the instruction set and is part of the same environment. GPUs are not. A GPU consists of two different programming environments and you have to move the data back and forth between them to get the benefits. The host cannot see the memory of the GPU; there are two separate address spaces.

It’s similar to what we saw with attached array processors in the 80s. What we saw back then that you had to explicitly move and manage the data — which reduced programmer productivity, raised actual cost of ownership and ultimately reduced performance. Like back then, the GPU programming model is different from its host.

GPUs initially did not have ECC correctable memory, now they do. This, however, demonstrates their lack of general purpose computing requirements. You have to work hard to make it work and not every application is amenable. The memory structure of a GPU is meant to be optimal for sequential access, but many programs require non-unity stride which will reduce performance for those applications. Classical supercomputers from Cray, Convex, NEC, and Fujitsu had very high bandwidth, highly interleaved main memory. A GPU is not going to be a general-purpose or a wide-spread solution for technical and software reasons. You can only execute the “hot spot” on the GPU, for example, and still need a classical host like the x86. It is not an integrated system. And, as of now, GPUs do not support virtual memory.

The GPU is really just a contemporary version of an attached array processor. If you look at the last 30 years, the architectures that have succeeded in the long term have been the ones that are easiest to program and that fit into the current environment. New languages take time to be learned and adopted. Organizations can’t hire the right people to program the machines. Each new full-time equivalent programmer who has to be hired can easily add $200,000 to $300,000 to the costs of the new system per year. This is not a new phenomenon; it has been true for the past few decades. The time to reconfigure is really expensive.

HPCwire: You’ve said that “software is the ‘Trojan Horse’ of high-performance computing.” What do you mean by that?

Wallach: As an organization, you accept the hardware — the horse — and then the next day the software warriors pour out and devour your IT department. As technology enthusiasts, we get excited by new technologies based on peak performance micro-architecture and the software questions come later as well as questions about “how do I fit it into my environment?” and “will I be able to achieve this level of performance with my applications?”

This has been true for the last 30 and will be true for the next 30 years. If you go back to the 80s — you had all kinds of interesting technologies like array processors and others, but the ones that had the best software succeeded such as Convex, Cray and Alliant. They succeeded because the programmer could leverage the technology for their FORTRAN and C environments. Integrated solutions like these succeeded and companies like CDC failed because their software was part of an anemic development environment. As another example from the past, the Japanese (Fujitsu and NEC) had exceptional software environments.
 
Fast forward to today. It’s like déjà vu all over again. A lot of new technologies are evolving but are not dealing with the software environment. Previous FPGA vendors had this problem. They were not integrated with the host environment. Vector processors, such as ClearSpeed, have this problem and this is true of all accelerators and GPUs.

The GPUs have some great technologies for visualization, for example, but are not integrated. You have to learn how to program in new languages like CUDA and there aren’t a lot of major applications written in CUDA. Programmers have to re-code or set up source translators that facilitate FORTRAN to CUDA. There are no translators for FORTRAN to assembly code and from a technical perspective it is much more efficient to go from FORTRAN to assembly code. Source to source translators are NOT as efficient as compilation to assembly code.

HPCwire: You talk about Convey’s hybrid-core computer as being an application specific, low power node. What is the significance of this description to the market?

Wallach: In the past decade, every generation has added new, specific instructions to general purpose computers to speed performance. For example, the current x86 system enhances image processing and new instructions have been developed to enhance vector processing. Since clock cycles are basically flat, you will see the trend toward specific instructions built into the microprocessors increasing. If one instruction can replace 10 instructions, you will have reduced power for that application. Our view is that it is now time to step up and increase the functionality of this approach. We advocate having one instruction to replace 100 instructions. Now you don’t have to rely on clock cycles to increase performance. You are relying instead on data and control paths. This approach is extremely useful for Convey and allows us to significantly increase performance while reducing power requirements, footprint and overall facility costs for a data center.

HPCwire: In order to be successful, do you think new computing paradigms need to leverage existing eco-structures like Linux and Windows?

Wallach: Absolutely. As I said before, new languages mean higher costs and lower productivity. In VC deals, whenever I hear that you have to program in a new language to make it work, I turn it down.

With new computing paradigms, you get several benefits when they leverage existing eco-structures like Linux and Windows. First off, they are more easily acceptable in the marketplace. If I’m the data center manager, I don’t have to hire anyone new or have training for a new eco-structure. No need to program in OCCAM, for example. I call programs that don’t take into consideration legacy systems and that are obscenely difficult to integrate, “pornographic” programs — you can’t always describe them exactly, but you know them when you see them. In 1984, I converted a FORTRAN program from CDC to ANSI FORTRAN to see what they were doing and it was awful. In the contemporary world, CUDA is the new pornographic programming language.

In addition, Windows and Linux allow for adoption of related technologies from other industries without changing the programming environment. Industry innovators such as the researchers at Lawrence Berkeley National Laboratory believe, for example, that future supercomputers will use the processors found in cell phones and other hand-held devices. Why? Because they use so little energy and have proven that they can handle sophisticated tasks (October 2009, IEEE Spectrum: “Low-Power Supercomputers” ). It is easy for the manufacturers to build chips designed for specific HPC applications just like they build different chips for each Smartphone brand. Chip manufacturers will also provide the software — compilers, debuggers, profiling tools, even complete Linux operating systems — tailored to each specific chip they sell which will make the new systems easy to integrate into a current environment.

HPCwire: Last year in HPCwire you said the future of HPC involves improved software, in particular more widespread use of PGAS languages and optical interconnects. Is this still the case?

Wallach: Yes. I believe the need for optical interconnects increases as we build large systems. The efficiency of scaling in parallel processing has to do with bandwidth and latency. Optical interconnects are much more efficient in terms of speed and power as compared to copper. PGAS (partitioned global address space) languages allow programmers a global view of their dataset and are much more efficient. PGAS languages also make it much easier to program highly parallel systems — they are much better than MPI.

HPCwire: Speaking of software, where is Convey on its development of different software personalities?

Wallach: We are on track with our development of personalities. Convey’s personalities are application architectures and instruction sets that support a wide array of application-specific solutions. Rather than develop hundreds of unique applications, we a creating a manageable number of personalities that can be leveraged in hundreds of different ways. We’ve shipped a range of different personalities for different customers, and we’ve got several others in development.

In the end, we anticipate developing around a dozen different core personalities. This is consistent with what leading researchers have determined, also. For example, in the study published by the University of California at Berkeley, “The Landscape of Parallel Computing Research: A View from Berkeley,” researchers define what they call MOTIFs or computer application structures for HPC. They describe 13 computer application structures on the Y access with the X access representing a particular application and how it uses that structure. Berkeley’s view is consistent with ours that there are approximately a dozen different personalities that cover the full spectrum of computing. In our development, we add a third element to the equation — the memory system — and see this as a three-dimensional grid. For this case, a unity stride (access sequential elements — dense data); or a highly interleaved (access non-sequential elements — multiple independently accessible memory banks — sparse data); or a “smart” memory (PIM – perform specific operations in the memory system — thread based) system is required for optimal performance.

We are on track to have personalities with memory structure and instruction sets with these MOTIFs, which is where we believe computing is going. For the HC-1, we ultimately anticipate 13 MOTIFs — but some will use the same personality.

HPCwire: Convey has just started shipping production units, can you tell us about the company’s early customers and how they’re using the HC-1?

Wallach: Early applications for the HC-1 follow the classic profile of HPC applications: signal-image processing, computer simulations, bioinformatics, and other applications we can’t discuss at this time. We have HC-1s going into the world’s leading research labs, all of which we will talk about during SC09 at our booth.

You can catch up with Steve Wallach during SC09, where he is participating in a talk on “HPC Architectures: Future Technologies and Systems” from 1:30-2:00 p.m. on Thursday (Rm. E143-144); or at Convey’s booth (#2589).

Subscribe to HPCwire's Weekly Update!

Be the most informed person in the room! Stay ahead of the tech trends with industry updates delivered to you every week!

MLPerf Inference 4.0 Results Showcase GenAI; Nvidia Still Dominates

March 28, 2024

There were no startling surprises in the latest MLPerf Inference benchmark (4.0) results released yesterday. Two new workloads — Llama 2 and Stable Diffusion XL — were added to the benchmark suite as MLPerf continues Read more…

Q&A with Nvidia’s Chief of DGX Systems on the DGX-GB200 Rack-scale System

March 27, 2024

Pictures of Nvidia's new flagship mega-server, the DGX GB200, on the GTC show floor got favorable reactions on social media for the sheer amount of computing power it brings to artificial intelligence.  Nvidia's DGX Read more…

Call for Participation in Workshop on Potential NSF CISE Quantum Initiative

March 26, 2024

Editor’s Note: Next month there will be a workshop to discuss what a quantum initiative led by NSF’s Computer, Information Science and Engineering (CISE) directorate could entail. The details are posted below in a Ca Read more…

Waseda U. Researchers Reports New Quantum Algorithm for Speeding Optimization

March 25, 2024

Optimization problems cover a wide range of applications and are often cited as good candidates for quantum computing. However, the execution time for constrained combinatorial optimization applications on quantum device Read more…

NVLink: Faster Interconnects and Switches to Help Relieve Data Bottlenecks

March 25, 2024

Nvidia’s new Blackwell architecture may have stolen the show this week at the GPU Technology Conference in San Jose, California. But an emerging bottleneck at the network layer threatens to make bigger and brawnier pro Read more…

Who is David Blackwell?

March 22, 2024

During GTC24, co-founder and president of NVIDIA Jensen Huang unveiled the Blackwell GPU. This GPU itself is heavily optimized for AI work, boasting 192GB of HBM3E memory as well as the the ability to train 1 trillion pa Read more…

MLPerf Inference 4.0 Results Showcase GenAI; Nvidia Still Dominates

March 28, 2024

There were no startling surprises in the latest MLPerf Inference benchmark (4.0) results released yesterday. Two new workloads — Llama 2 and Stable Diffusion Read more…

Q&A with Nvidia’s Chief of DGX Systems on the DGX-GB200 Rack-scale System

March 27, 2024

Pictures of Nvidia's new flagship mega-server, the DGX GB200, on the GTC show floor got favorable reactions on social media for the sheer amount of computing po Read more…

NVLink: Faster Interconnects and Switches to Help Relieve Data Bottlenecks

March 25, 2024

Nvidia’s new Blackwell architecture may have stolen the show this week at the GPU Technology Conference in San Jose, California. But an emerging bottleneck at Read more…

Who is David Blackwell?

March 22, 2024

During GTC24, co-founder and president of NVIDIA Jensen Huang unveiled the Blackwell GPU. This GPU itself is heavily optimized for AI work, boasting 192GB of HB Read more…

Nvidia Looks to Accelerate GenAI Adoption with NIM

March 19, 2024

Today at the GPU Technology Conference, Nvidia launched a new offering aimed at helping customers quickly deploy their generative AI applications in a secure, s Read more…

The Generative AI Future Is Now, Nvidia’s Huang Says

March 19, 2024

We are in the early days of a transformative shift in how business gets done thanks to the advent of generative AI, according to Nvidia CEO and cofounder Jensen Read more…

Nvidia’s New Blackwell GPU Can Train AI Models with Trillions of Parameters

March 18, 2024

Nvidia's latest and fastest GPU, codenamed Blackwell, is here and will underpin the company's AI plans this year. The chip offers performance improvements from Read more…

Nvidia Showcases Quantum Cloud, Expanding Quantum Portfolio at GTC24

March 18, 2024

Nvidia’s barrage of quantum news at GTC24 this week includes new products, signature collaborations, and a new Nvidia Quantum Cloud for quantum developers. Wh Read more…

Alibaba Shuts Down its Quantum Computing Effort

November 30, 2023

In case you missed it, China’s e-commerce giant Alibaba has shut down its quantum computing research effort. It’s not entirely clear what drove the change. Read more…

Nvidia H100: Are 550,000 GPUs Enough for This Year?

August 17, 2023

The GPU Squeeze continues to place a premium on Nvidia H100 GPUs. In a recent Financial Times article, Nvidia reports that it expects to ship 550,000 of its lat Read more…

Shutterstock 1285747942

AMD’s Horsepower-packed MI300X GPU Beats Nvidia’s Upcoming H200

December 7, 2023

AMD and Nvidia are locked in an AI performance battle – much like the gaming GPU performance clash the companies have waged for decades. AMD has claimed it Read more…

DoD Takes a Long View of Quantum Computing

December 19, 2023

Given the large sums tied to expensive weapon systems – think $100-million-plus per F-35 fighter – it’s easy to forget the U.S. Department of Defense is a Read more…

Synopsys Eats Ansys: Does HPC Get Indigestion?

February 8, 2024

Recently, it was announced that Synopsys is buying HPC tool developer Ansys. Started in Pittsburgh, Pa., in 1970 as Swanson Analysis Systems, Inc. (SASI) by John Swanson (and eventually renamed), Ansys serves the CAE (Computer Aided Engineering)/multiphysics engineering simulation market. Read more…

Choosing the Right GPU for LLM Inference and Training

December 11, 2023

Accelerating the training and inference processes of deep learning models is crucial for unleashing their true potential and NVIDIA GPUs have emerged as a game- Read more…

Intel’s Server and PC Chip Development Will Blur After 2025

January 15, 2024

Intel's dealing with much more than chip rivals breathing down its neck; it is simultaneously integrating a bevy of new technologies such as chiplets, artificia Read more…

Baidu Exits Quantum, Closely Following Alibaba’s Earlier Move

January 5, 2024

Reuters reported this week that Baidu, China’s giant e-commerce and services provider, is exiting the quantum computing development arena. Reuters reported � Read more…

Leading Solution Providers

Contributors

Comparing NVIDIA A100 and NVIDIA L40S: Which GPU is Ideal for AI and Graphics-Intensive Workloads?

October 30, 2023

With long lead times for the NVIDIA H100 and A100 GPUs, many organizations are looking at the new NVIDIA L40S GPU, which it’s a new GPU optimized for AI and g Read more…

Shutterstock 1179408610

Google Addresses the Mysteries of Its Hypercomputer 

December 28, 2023

When Google launched its Hypercomputer earlier this month (December 2023), the first reaction was, "Say what?" It turns out that the Hypercomputer is Google's t Read more…

AMD MI3000A

How AMD May Get Across the CUDA Moat

October 5, 2023

When discussing GenAI, the term "GPU" almost always enters the conversation and the topic often moves toward performance and access. Interestingly, the word "GPU" is assumed to mean "Nvidia" products. (As an aside, the popular Nvidia hardware used in GenAI are not technically... Read more…

Shutterstock 1606064203

Meta’s Zuckerberg Puts Its AI Future in the Hands of 600,000 GPUs

January 25, 2024

In under two minutes, Meta's CEO, Mark Zuckerberg, laid out the company's AI plans, which included a plan to build an artificial intelligence system with the eq Read more…

Google Introduces ‘Hypercomputer’ to Its AI Infrastructure

December 11, 2023

Google ran out of monikers to describe its new AI system released on December 7. Supercomputer perhaps wasn't an apt description, so it settled on Hypercomputer Read more…

China Is All In on a RISC-V Future

January 8, 2024

The state of RISC-V in China was discussed in a recent report released by the Jamestown Foundation, a Washington, D.C.-based think tank. The report, entitled "E Read more…

Intel Won’t Have a Xeon Max Chip with New Emerald Rapids CPU

December 14, 2023

As expected, Intel officially announced its 5th generation Xeon server chips codenamed Emerald Rapids at an event in New York City, where the focus was really o Read more…

IBM Quantum Summit: Two New QPUs, Upgraded Qiskit, 10-year Roadmap and More

December 4, 2023

IBM kicks off its annual Quantum Summit today and will announce a broad range of advances including its much-anticipated 1121-qubit Condor QPU, a smaller 133-qu Read more…

  • arrow
  • Click Here for More Headlines
  • arrow
HPCwire