NVIDIA Pokes Holes in Intel’s Manycore Story

By Michael Feldman

April 3, 2012

As NVIDIA’s upcoming Kepler-grade Tesla GPU prepares to do battle with Intel’s Knight Corner, the companies are busy formulating their respective HPC accelerator stories. While NVIDIA has enjoyed the advantage of actually having products in the field to talk about, Intel has managed to capture the attention of some fence-sitters with assurances of high programmability, simple recompiles, and transparent scalability for its Many Integrated Core (MIC) coprocessors. But according to NVIDIA’s Steve Scott, such promises ignore certain hard truths about how accelerator-based computing really works.

Over the past couple of years, Intel has been telling would-be MIC users that its upcoming Knights Corner coprocessor will deliver the performance of a GPU without the challenges of a having to adopt a new programming model — CUDA OpenCL, or whatever. And since the MIC architecture is x86-based (essentially simple Pentium cores glued to extra wide vector units), developing Knights Corner applications will not be that different than programming a multicore Xeon CPU.

Leveraging that commonality, Intel says their compiler will be able generate MIC executables from legacy HPC source code. And it will do so for applications based on both MPI and OpenMP, the two most popular parallel programming frameworks used in high performance computing. Essentially Intel is promising a free port to MIC.

Not so fast, says Scott, the former Cray alum who joined NVIDIA last year its chief technology officer of the Tesla business. According to him, porting applications for MIC, or even developing new ones, won’t be any easier than programming GPUs, or for that matter, any accelerator. In a blog posted on Tuesday, he described the problems with Intel’s manycore narrative and its claims of superiority over GPU computing.

Scott is not arguing against the MIC as an accelerator, per se. He and most of the community are convinced that HPC needs a hybrid (or heterogeneous) computing to move performance forward without consuming unreasonable amounts of energy. Traditional CPUs, whose cores are optimized for single-threaded performance, are not designed for work requiring lots of throughput. For that type of computing, much better energy efficiency can be delivered using simpler, slower, but more numerous cores. Both GPUs and the MIC adhere to this paradigm; they just come at the problem from different architectural pedigrees.

The problem is that running throughput code on a serial processor sucks up too much energy, which is the situation many users are facing today with conventional CPUs. Conversely, running serial code on a throughput processor is just too slow, and defeats the purpose of having an accelerator in the first place.

Even if low single-threaded performance wasn’t an issue, today’s accelerators live on PCIe cards with limited amounts of memory (usually just a handful a gigabytes) that exists at the end of a PCIe bus. So if the entire application were to run on the accelerator, all its data and instructions would have to be shuttled in from main memory in chunks. Consider that today, with only a portion of the application living on the GPU, the PCIe bottleneck can still hinder performance. Stuffing the whole program on the accelerator would make it that much worse.

So the main thrust of Scott’s critique is that for hybrid computing to work, you have to split the application intelligently between the CPU host and the accelerator. That’s true, he says, whether you’re talking about an x86-based accelerator like MIC or a graphic-based one like Tesla. “The entire game now is how do we deliver performance as power efficiently as possible,” he told HPCwire.

Intel has revealed very little about application performance on the future MIC parts, and has not really addressed how that application split is going to work programmatically, or even that it’s necessary. To date, they and some of the early MIC adopters have mostly talked about recompiling existing codes, based on OpenMP and/or MPI, and running the resulting executable natively on MIC.

Running MPI codes on a manycore architecture is particularly problematic. First there’s the memory capacity problem mentioned above (each MPI process uses quite a bit of data). And then there’s the fact that once the number of MPI processes exceeds the accelerator core count — 50-plus for Knights Corner — the application would have to use the server node’s network card to communicate with MPI processes running on other nodes. As Scott points out in his blog, that’s far too many MPI processes for a typical network interface; all the contention would overwhelm the available bandwidth.

OpenMP has the opposite problem, since most programs using this model don’t scale beyond more than 4 to 8 tasks. As a result, there would no way for most OpenMP applications to utilize the 50-plus cores expected on Knights Corner-equipped nodes. And once again, there’s the memory capacity problem. Like MPI, OpenMP expects to live in the relatively spacious accommodations of the CPU’s main memory.

Scott says if you’re just going to use a compiler to transform your existing application to run on the MIC, you’re not doing hybrid computing at all. More importantly, running the entire code on the accelerator does not take performance into account. After all, the idea is to speed up the application, not just recompile it so that it functionally works. “We don’t think it’s legitimate to talk about ease of programming without talking about performance,” he says.

Scott argues that for applications to take advantage of these new throughput processors, programmers will have delve into some sort of hybrid programming model that splits off the parallel throughput code from the serial code. For NVIDIA GPUs, the parallelism can be exposed with CUDA or with the emerging set of OpenMP-like directives for accelerators, known as OpenACC. There is already an initial CUDA port for x86 developed by PGI, so that’s one option. But the OpenACC framework is likely to reach a larger audience of developers since it offers a higher level of abstraction than CUDA and it looks like it will eventually be folded into the industry-standard OpenMP API.

The idea is that programmers can use OpenACC today to develop GPU-accelerated applications with the anticipation they will be able to use the same code for other accelerator-based hardware platforms, like MIC and AMD’s Fusion or discrete GPU processors. Intel and AMD have not jumped on the OpenACC bandwagon as of yet, but were it to be adopted as a standard and demanded by their customers, they would certainly have to support it.

Even OpenACC is not a magic bullet though. The programmer still has to do dive into the source code  and tell the compiler where and how to carve out parallel code for the accelerator. And as Scott admits, that can be a significant effort, especially for large legacy HPC applications that were written for homogeneous CPU-only machines.

But, he maintains, if you’re interested in taking advantage of the performance offered by throughput processor like GPUs and MIC, the work has to be done. Processor clocks are not likely get any faster than they are today. So the only way to increase performance is via parallelism. As Scott says, “Computers aren’t getting faster, they’re only getting wider.”

Related Articles

The Heterogeneous Programming Jungle

NVIDIA Eyes Post-CUDA Era of GPU Computing

Intel Touts Manycore Coprocessor at Supercomputing Conference

Subscribe to HPCwire's Weekly Update!

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

Data Vortex Users Contemplate the Future of Supercomputing

October 19, 2017

Last month (Sept. 11-12), HPC networking company Data Vortex held its inaugural users group at Pacific Northwest National Laboratory (PNNL) bringing together about 30 participants from industry, government and academia t Read more…

By Tiffany Trader

AI Self-Training Goes Forward at Google DeepMind

October 19, 2017

DeepMind, Google’s AI research organization, announced today in a blog that AlphaGo Zero, the latest evolution of AlphaGo (the first computer program to defeat a Go world champion) trained itself within three days to play Go at a superhuman level (i.e., better than any human) – and to beat the old version of AlphaGo – without leveraging human expertise, data or training. Read more…

By Doug Black

Researchers Scale COSMO Climate Code to 4888 GPUs on Piz Daint

October 17, 2017

Effective global climate simulation, sorely needed to anticipate and cope with global warming, has long been computationally challenging. Two of the major obstacles are the needed resolution and prolonged time to compute Read more…

By John Russell

HPE Extreme Performance Solutions

Transforming Genomic Analytics with HPC-Accelerated Insights

Advancements in the field of genomics are revolutionizing our understanding of human biology, rapidly accelerating the discovery and treatment of genetic diseases, and dramatically improving human health. Read more…

Student Cluster Competition Coverage New Home

October 16, 2017

Hello computer sports fans! This is the first of many (many!) articles covering the world-wide phenomenon of Student Cluster Competitions. Finally, the Student Cluster Competition coverage has come to its natural home: H Read more…

By Dan Olds

Data Vortex Users Contemplate the Future of Supercomputing

October 19, 2017

Last month (Sept. 11-12), HPC networking company Data Vortex held its inaugural users group at Pacific Northwest National Laboratory (PNNL) bringing together ab Read more…

By Tiffany Trader

AI Self-Training Goes Forward at Google DeepMind

October 19, 2017

DeepMind, Google’s AI research organization, announced today in a blog that AlphaGo Zero, the latest evolution of AlphaGo (the first computer program to defeat a Go world champion) trained itself within three days to play Go at a superhuman level (i.e., better than any human) – and to beat the old version of AlphaGo – without leveraging human expertise, data or training. Read more…

By Doug Black

Student Cluster Competition Coverage New Home

October 16, 2017

Hello computer sports fans! This is the first of many (many!) articles covering the world-wide phenomenon of Student Cluster Competitions. Finally, the Student Read more…

By Dan Olds

Intel Delivers 17-Qubit Quantum Chip to European Research Partner

October 10, 2017

On Tuesday, Intel delivered a 17-qubit superconducting test chip to research partner QuTech, the quantum research institute of Delft University of Technology (TU Delft) in the Netherlands. The announcement marks a major milestone in the 10-year, $50-million collaborative relationship with TU Delft and TNO, the Dutch Organization for Applied Research, to accelerate advancements in quantum computing. Read more…

By Tiffany Trader

Fujitsu Tapped to Build 37-Petaflops ABCI System for AIST

October 10, 2017

Fujitsu announced today it will build the long-planned AI Bridging Cloud Infrastructure (ABCI) which is set to become the fastest supercomputer system in Japan Read more…

By John Russell

HPC Chips – A Veritable Smorgasbord?

October 10, 2017

For the first time since AMD's ill-fated launch of Bulldozer the answer to the question, 'Which CPU will be in my next HPC system?' doesn't have to be 'Whichever variety of Intel Xeon E5 they are selling when we procure'. Read more…

By Dairsie Latimer

Delays, Smoke, Records & Markets – A Candid Conversation with Cray CEO Peter Ungaro

October 5, 2017

Earlier this month, Tom Tabor, publisher of HPCwire and I had a very personal conversation with Cray CEO Peter Ungaro. Cray has been on something of a Cinderell Read more…

By Tiffany Trader & Tom Tabor

Intel Debuts Programmable Acceleration Card

October 5, 2017

With a view toward supporting complex, data-intensive applications, such as AI inference, video streaming analytics, database acceleration and genomics, Intel i Read more…

By Doug Black

Reinders: “AVX-512 May Be a Hidden Gem” in Intel Xeon Scalable Processors

June 29, 2017

Imagine if we could use vector processing on something other than just floating point problems.  Today, GPUs and CPUs work tirelessly to accelerate algorithms Read more…

By James Reinders

NERSC Scales Scientific Deep Learning to 15 Petaflops

August 28, 2017

A collaborative effort between Intel, NERSC and Stanford has delivered the first 15-petaflops deep learning software running on HPC platforms and is, according Read more…

By Rob Farber

Oracle Layoffs Reportedly Hit SPARC and Solaris Hard

September 7, 2017

Oracle’s latest layoffs have many wondering if this is the end of the line for the SPARC processor and Solaris OS development. As reported by multiple sources Read more…

By John Russell

US Coalesces Plans for First Exascale Supercomputer: Aurora in 2021

September 27, 2017

At the Advanced Scientific Computing Advisory Committee (ASCAC) meeting, in Arlington, Va., yesterday (Sept. 26), it was revealed that the "Aurora" supercompute Read more…

By Tiffany Trader

How ‘Knights Mill’ Gets Its Deep Learning Flops

June 22, 2017

Intel, the subject of much speculation regarding the delayed, rewritten or potentially canceled “Aurora” contract (the Argonne Lab part of the CORAL “ Read more…

By Tiffany Trader

Google Releases Deeplearn.js to Further Democratize Machine Learning

August 17, 2017

Spreading the use of machine learning tools is one of the goals of Google’s PAIR (People + AI Research) initiative, which was introduced in early July. Last w Read more…

By John Russell

Nvidia Responds to Google TPU Benchmarking

April 10, 2017

Nvidia highlights strengths of its newest GPU silicon in response to Google's report on the performance and energy advantages of its custom tensor processor. Read more…

By Tiffany Trader

GlobalFoundries Puts Wind in AMD’s Sails with 12nm FinFET

September 24, 2017

From its annual tech conference last week (Sept. 20), where GlobalFoundries welcomed more than 600 semiconductor professionals (reaching the Santa Clara venue Read more…

By Tiffany Trader

Leading Solution Providers

Graphcore Readies Launch of 16nm Colossus-IPU Chip

July 20, 2017

A second $30 million funding round for U.K. AI chip developer Graphcore sets up the company to go to market with its “intelligent processing unit” (IPU) in Read more…

By Tiffany Trader

Amazon Debuts New AMD-based GPU Instances for Graphics Acceleration

September 12, 2017

Last week Amazon Web Services (AWS) streaming service, AppStream 2.0, introduced a new GPU instance called Graphics Design intended to accelerate graphics. The Read more…

By John Russell

EU Funds 20 Million Euro ARM+FPGA Exascale Project

September 7, 2017

At the Barcelona Supercomputer Centre on Wednesday (Sept. 6), 16 partners gathered to launch the EuroEXA project, which invests €20 million over three-and-a-half years into exascale-focused research and development. Led by the Horizon 2020 program, EuroEXA picks up the banner of a triad of partner projects — ExaNeSt, EcoScale and ExaNoDe — building on their work... Read more…

By Tiffany Trader

Delays, Smoke, Records & Markets – A Candid Conversation with Cray CEO Peter Ungaro

October 5, 2017

Earlier this month, Tom Tabor, publisher of HPCwire and I had a very personal conversation with Cray CEO Peter Ungaro. Cray has been on something of a Cinderell Read more…

By Tiffany Trader & Tom Tabor

Cray Moves to Acquire the Seagate ClusterStor Line

July 28, 2017

This week Cray announced that it is picking up Seagate's ClusterStor HPC storage array business for an undisclosed sum. "In short we're effectively transitioning the bulk of the ClusterStor product line to Cray," said CEO Peter Ungaro. Read more…

By Tiffany Trader

Intel Launches Software Tools to Ease FPGA Programming

September 5, 2017

Field Programmable Gate Arrays (FPGAs) have a reputation for being difficult to program, requiring expertise in specialty languages, like Verilog or VHDL. Easin Read more…

By Tiffany Trader

IBM Advances Web-based Quantum Programming

September 5, 2017

IBM Research is pairing its Jupyter-based Data Science Experience notebook environment with its cloud-based quantum computer, IBM Q, in hopes of encouraging a new class of entrepreneurial user to solve intractable problems that even exceed the capabilities of the best AI systems. Read more…

By Alex Woodie

HPC Chips – A Veritable Smorgasbord?

October 10, 2017

For the first time since AMD's ill-fated launch of Bulldozer the answer to the question, 'Which CPU will be in my next HPC system?' doesn't have to be 'Whichever variety of Intel Xeon E5 they are selling when we procure'. Read more…

By Dairsie Latimer

  • arrow
  • Click Here for More Headlines
  • arrow
Share This