Purdue Researchers Peer into the ‘Fog of the Machine Learning Accelerator War’

By John Russell

September 27, 2021

Making sense of ML performance and benchmark data is an ongoing challenge. In light of last week’s release of the most recent MLPerf (v1.1) inference results, now is perhaps a good time to review how valuable (or not) such ML benchmarks are and the challenges they face. Two researchers from Purdue University recently tackled this issue in a fascinating blog on ACM SIGARCH – An Academic’s Attempt to Clear the Fog of the Machine Learning Accelerator War.

Tim Rogers and Mahmoud Khairy wrote, “At its core, all engineering is science optimized (or perverted) by economics. As academics in computer science and engineering, we have a symbiotic relationship with industry. Still, it is often necessary for us to peel back the marketing noise and understand precisely what companies are building, why their offering is better than their competitors, and how we can innovate beyond what industrial state-of-the-art provides.

“The more immediately relevant a problem is, the more funding and innovation the solutions will receive, at the cost of increased noise for both customers deciding where to place investments and researchers attempting to innovate without bias. In 2021, the raging machine learning accelerator war is now a well-established battleground of processors, ideas, and noise.”

With that the researchers plunge into describing the two dominant approaches (data parallelism and model parallelism) to solving ML problems and how they match to different chip/system architectures and different problem types and sizes.

Rogers and Khairy credit MLPerf for providing “a SPEC-like benchmark suite that we hope will eventually provide clarity on the best-performing machine learning systems,” but then recognize many companies don’t participate for a variety of economic and other reasons. The larger question they pose is, “In the coming age of customized accelerators, do standardized benchmark suites make sense at all anymore? Broadly, this is likely to be a challenge, but in the context of GEMM-heavy machine learning workloads, we think standardization is both possible and necessary.”

Shown below is a chart they prepared comparing several accelerators including Cerebras and Graphcore, which have not competed in MLPerf. Khairy dug through a variety of public documentation to pull together this “apples-to-apples comparison of 16-bit mixed-precision training:”

The meat of the post is about data parallelism and model parallelism and Rogers and Khairy dig in and provide some comparison metrics. Here’s a summary description of the approaches excerpted from the blog:

  • Data Parallelism. “Generally both NVIDIA and Google rely on data parallelism to scale performance. In this context, data parallelism refers to replicating the network model on different accelerators and assigning a fraction of each training batch to each node (Figure 2). The relatively large DRAM capacity of the TPU and GPU makes network replication feasible, and the dedicated hardware for dense GEMM operations make large batch training efficient. Both vendors have also invested significant effort in making the reduction operation across nodes as efficient as possible. However, the key to exploiting data parallelism is achieving efficient training with a large batch size, which requires a non-trivial amount of hyper-parameter tuning.
  • Model Parallelism. “Cerebras and Graphcore take a different approach, focusing heavily on model parallelism. In model parallelism, each layer of a single model is mapped to hardware, and data flows between each layer, ideally to processing units on the same chip. To exploit this philosophy, both Graphcore and Cerebras devote a large portion of their silicon area to SRAM cells such that large models are placed entirely in SRAM. Cerebras achieves this with its wafer-scale chip, while Graphcore’s IPU servers consist of smaller chips linked via a ring interconnect.”

They emphasize it’s not, strictly-peaking, an either-or question. GPUs and TPUs can exploit model parallelism, and Cerebras/Graphcore servers can exploit data parallelism.

“However, each offering has a particular type of parallelism where its design has an advantage. It seems that one open question academics can help answer is what the right balance between model and data parallelism should be. Data parallelism is difficult to scale, and excessive hyperparameter tuning gives a clear advantage to more prominent players with the time and resources to tune them. In contrast, model parallelism requires less tuning, but as soon as the model no longer fits on-chip or on-node, most of the advantages are lost,” they wrote.

Their blog is a quick read and worth the effort (apologies for excerpting so much material). They emphasize their intent is not to predict the war’s winner but rather predict what weapons will turn the tide: “Academically, the data versus model parallelism tradeoff is interesting. If models continue to grow, exploiting model parallelism on the same chip, wafer-scale or not, will be infeasible. Conversely, increasing the batch size to exploit data parallelism is difficult, and larger models won’t make this easier. It seems that a hybrid approach will be the solution, and hardware that can effectively adapt to different model/data parallelism levels will likely have the most impact.”

Link to ACM SIGARCH blog: https://www.sigarch.org/an-academics-attempt-to-clear-the-fog-of-the-machine-learning-accelerator-war/

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!

Durham University to Test Rockport Networks on COSMA7 Supercomputer

May 24, 2022

Durham University’s Institute for Computational Cosmology (ICC) is home to the COSMA series of supercomputers (short for “cosmological machine”). COSMA—now in its eighth iteration, COSMA8—has been working to an Read more…

SoftIron Measures Its Carbon Footprint to Make a Point

May 24, 2022

Since its founding in 2012, London-based software-defined storage provider SoftIron has been making its case for what it calls secure provenance: a term that encompasses the company’s rigorous accounting of the supply Read more…

TACC Adds Details to Vision for Leadership-Class Computing Facility

May 23, 2022

The Texas Advanced Computing Center (TACC) at The University of Texas at Austin passed to the next phase of the planning process for the Leadership-Class Computing Facility (LCCF), a process that has many approval stage Read more…

ISC 2022: International Association of Supercomputing Centers to Debut

May 23, 2022

At ISC 2022 in Hamburg, Germany, representatives from four supercomputing centers across three countries plan to debut the International Association of Supercomputing Centers (IASC). The IASC aims to bring together publi Read more…

ANL Special Colloquium on The Future of Computing

May 19, 2022

There are, of course, a myriad of ideas regarding computing’s future. At yesterday’s Argonne National Laboratory’s Director’s Special Colloquium, The Future of Computing, guest speaker Sadasivan Shankar, did his best to convince the audience that the high-energy cost of the current computing paradigm – not (just) economic cost; we’re talking entropy here – is fundamentally undermining computing’s progress such that... Read more…

AWS Solution Channel

shutterstock 1103121086

Encoding workflow dependencies in AWS Batch

Most users of HPC or Batch systems need to analyze data with multiple operations to get meaningful results. That’s really driven by the nature of scientific research or engineering processes – it’s rare that a single task generates the insight you need. Read more…

What’s New in HPC Research: Air Pollution Prediction, nOS-V, cuHARM, Quantum Ray Tracing & More

May 19, 2022

In this regular feature, HPCwire highlights newly published research in the high-performance computing community and related domains. From parallel programming to exascale to quantum computing, the details are here. Read more…

Durham University to Test Rockport Networks on COSMA7 Supercomputer

May 24, 2022

Durham University’s Institute for Computational Cosmology (ICC) is home to the COSMA series of supercomputers (short for “cosmological machine”). COSMA— Read more…

SoftIron Measures Its Carbon Footprint to Make a Point

May 24, 2022

Since its founding in 2012, London-based software-defined storage provider SoftIron has been making its case for what it calls secure provenance: a term that en Read more…

ISC 2022: International Association of Supercomputing Centers to Debut

May 23, 2022

At ISC 2022 in Hamburg, Germany, representatives from four supercomputing centers across three countries plan to debut the International Association of Supercom Read more…

ANL Special Colloquium on The Future of Computing

May 19, 2022

There are, of course, a myriad of ideas regarding computing’s future. At yesterday’s Argonne National Laboratory’s Director’s Special Colloquium, The Future of Computing, guest speaker Sadasivan Shankar, did his best to convince the audience that the high-energy cost of the current computing paradigm – not (just) economic cost; we’re talking entropy here – is fundamentally undermining computing’s progress such that... Read more…

HPE Announces New HPC Factory in Czech Republic

May 18, 2022

A week ahead of ISC High Performance 2022 (set to be held in Hamburg, Germany), supercomputing heavyweight HPE has announced a major investment in sovereign Eur Read more…

ECP Director Doug Kothe Named ORNL Associate Laboratory Director

May 16, 2022

The Department of Energy's Oak Ridge National Laboratory (ORNL) has selected Doug Kothe to be the next Associate Laboratory Director for its Computing and Computational Sciences Directorate (CCSD), HPCwire has learned. Kothe will fill the vacancy created by the retirement of Jeff Nichols, whose last day is July 1. Kothe will transition into the position on June 6. As director of the United States' Exascale Computing Project... Read more…

Google Cloud’s New TPU v4 ML Hub Packs 9 Exaflops of AI

May 16, 2022

Almost exactly a year ago, Google launched its Tensor Processing Unit (TPU) v4 chips at Google I/O 2021, promising twice the performance compared to the TPU v3. At the time, Google CEO Sundar Pichai said that Google’s datacenters would “soon have dozens of TPU v4 Pods, many of which will be... Read more…

Q&A with Candace Culhane, SC22 General Chair and an HPCwire Person to Watch in 2022

May 14, 2022

HPCwire is pleased to present our interview with SC22 General Chair Candace Culhane, program/project director at Los Alamos National Lab and an HPCwire 2022 Per Read more…

Nvidia R&D Chief on How AI is Improving Chip Design

April 18, 2022

Getting a glimpse into Nvidia’s R&D has become a regular feature of the spring GTC conference with Bill Dally, chief scientist and senior vice president of research, providing an overview of Nvidia’s R&D organization and a few details on current priorities. This year, Dally focused mostly on AI tools that Nvidia is both developing and using in-house to improve... Read more…

Royalty-free stock illustration ID: 1919750255

Intel Says UCIe to Outpace PCIe in Speed Race

May 11, 2022

Intel has shared more details on a new interconnect that is the foundation of the company’s long-term plan for x86, Arm and RISC-V architectures to co-exist in a single chip package. The semiconductor company is taking a modular approach to chip design with the option for customers to cram computing blocks such as CPUs, GPUs and AI accelerators inside a single chip package. Read more…

AMD/Xilinx Takes Aim at Nvidia with Improved VCK5000 Inferencing Card

March 8, 2022

AMD/Xilinx has released an improved version of its VCK5000 AI inferencing card along with a series of competitive benchmarks aimed directly at Nvidia’s GPU line. AMD says the new VCK5000 has 3x better performance than earlier versions and delivers 2x TCO over Nvidia T4. AMD also showed favorable benchmarks against several Nvidia GPUs, claiming its VCK5000 achieved... Read more…

In Partnership with IBM, Canada to Get Its First Universal Quantum Computer

February 3, 2022

IBM today announced it will deploy its first quantum computer in Canada, putting Canada on a short list of countries that will have access to an IBM Quantum Sys Read more…

Supercomputer Simulations Show How Paxlovid, Pfizer’s Covid Antiviral, Works

February 3, 2022

Just about a month ago, Pfizer scored its second huge win of the pandemic when the U.S. Food and Drug Administration issued another emergency use authorization Read more…

Nvidia Launches Hopper H100 GPU, New DGXs and Grace Superchips

March 22, 2022

The battle for datacenter dominance keeps getting hotter. Today, Nvidia kicked off its spring GTC event with new silicon, new software and a new supercomputer. Speaking from a virtual environment in the Nvidia Omniverse 3D collaboration and simulation platform, CEO Jensen Huang introduced the new Hopper GPU architecture and the H100 GPU... Read more…

Facebook Parent Meta’s New AI Supercomputer Will Be ‘World’s Fastest’

January 24, 2022

Fresh off its rebrand last October, Meta (née Facebook) is putting muscle behind its vision of a metaversal future with a massive new AI supercomputer called the AI Research SuperCluster (RSC). Meta says that RSC will be used to help build new AI models, develop augmented reality tools, seamlessly analyze multimedia data and more. The supercomputer’s... Read more…

PsiQuantum’s Path to 1 Million Qubits

April 21, 2022

PsiQuantum, founded in 2016 by four researchers with roots at Bristol University, Stanford University, and York University, is one of a few quantum computing startups that’s kept a moderately low PR profile. (That’s if you disregard the roughly $700 million in funding it has attracted.) The main reason is PsiQuantum has eschewed the clamorous public chase for... Read more…

Leading Solution Providers

Contributors

Nvidia Dominates MLPerf Inference, Qualcomm also Shines, Where’s Everybody Else?

April 6, 2022

MLCommons today released its latest MLPerf inferencing results, with another strong showing by Nvidia accelerators inside a diverse array of systems. Roughly fo Read more…

D-Wave to Go Public with SPAC Deal; Expects ~$1.6B Market Valuation

February 8, 2022

Quantum computing pioneer D-Wave today announced plans to go public via a SPAC (special purpose acquisition company) mechanism. D-Wave will merge with DPCM Capital in a transaction expected to produce $340 million in cash and result in a roughly $1.6 billion initial market valuation. The deal is expected to be completed in the second quarter of 2022 and the new company will be traded on the New York Stock... Read more…

Intel Announces Falcon Shores CPU-GPU Combo Architecture for 2024

February 18, 2022

Intel held its 2022 investor meeting yesterday, covering everything from the imminent Sapphire Rapids CPUs to the hotly anticipated (and delayed) Ponte Vecchio GPUs. But somewhat buried in its summary of the meeting was a new namedrop: “Falcon Shores,” described as “a new architecture that will bring x86 and Xe GPU together into a single socket.” The reveal was... Read more…

Industry Consortium Forms to Drive UCIe Chiplet Interconnect Standard

March 2, 2022

A new industry consortium aims to establish a die-to-die interconnect standard – Universal Chiplet Interconnect Express (UCIe) – in support of an open chipl Read more…

India Launches Petascale ‘PARAM Ganga’ Supercomputer

March 8, 2022

Just a couple of weeks ago, the Indian government promised that it had five HPC systems in the final stages of installation and would launch nine new supercomputers this year. Now, it appears to be making good on that promise: the country’s National Supercomputing Mission (NSM) has announced the deployment of “PARAM Ganga” petascale supercomputer at Indian Institute of Technology (IIT)... Read more…

Julia Update: Adoption Keeps Climbing; Is It a Python Challenger?

January 13, 2021

The rapid adoption of Julia, the open source, high level programing language with roots at MIT, shows no sign of slowing according to data from Julialang.org. I Read more…

Nvidia Acquires Software-Defined Storage Provider Excelero

March 7, 2022

Nvidia has announced that it has acquired Excelero. The high-performance block storage provider, founded in 2014, will have its technology integrated into Nvidia’s enterprise software stack. Nvidia is not disclosing the value of the deal. Excelero’s core product, Excelero NVMesh, offers software-defined block storage via networked NVMe SSDs. NVMesh operates through... Read more…

Google Launches TPU v4 AI Chips

May 20, 2021

Google CEO Sundar Pichai spoke for only one minute and 42 seconds about the company’s latest TPU v4 Tensor Processing Units during his keynote at the Google I Read more…

  • arrow
  • Click Here for More Headlines
  • arrow
HPCwire