Nvidia’s Ampere A100 GPU: Up to 2.5X the HPC, 20X the AI

By Tiffany Trader

May 14, 2020

Nvidia’s first Ampere-based graphics card, the A100 GPU, packs a whopping 54 billion transistors on 826mm2 of silicon, making it the world’s largest seven-nanometer chip. Launched today during a pre-recorded “kitchen keynote” from Nvidia chief Jensen Huang, the Ampere architecture follows in predecessor Volta’s footsteps, a mega-GPU that turns up the dial on transistors, AI specialization and overall performance. Both the 12nm Volta and the 7nm Ampere are manufactured by TSMC.

Jensen Huang holds up new Ampere A100 GPU during his “kitchen keynote.” There was no on-site GPU Technology Conference (GTC) this year, due to the COVID-19 outbreak.

HPC workloads that leverage Ampere benefit from a 250 percent increase in peak double-precision floating point performance over older brother Volta owing to new IEEE-compliant tensor core instructions for HPC processing. Benchmarking conducted by Nvidia realized speedups for HPC workloads ranging between 1.5x and 2.1x over Volta (see the “Accelerating HPC” chart further down the page). Peak single precision performance gets a theoretical 10-20X boost with the addition of TensorFloat-32 (TF32) tensor cores.

Other new features include:

• Multi-instance GPU (aka MIG) which enables a single A100 GPU to be partitioned into as many as seven separate GPUs.

• Third-generation Nvidia NVLink fabric, which doubles the high-speed connectivity between GPUs.

• And structural sparsity, which introduces support for sparse matrix operations in Tensor cores, and accelerates them by two times.

Here’s a comparison chart showing key specs for Volta V100 and Ampere A100 GPUs:

Ampere – named after the French physicist and mathematician – doubles down on Nvidia’s single universal GPU strategy introduced with the Volta architecture that supported HPC, AI and graphics. “[With Ampere A100], instead of having a whole bunch of Volta GPU servers, a bunch of T4 GPU servers, and CPU servers, all of that can now be run on one unified Ampere server,” said Huang in a briefing held for media yesterday.

“This is unquestionably the first time that we’ve unified the acceleration workload of the entire datacenter into one single platform,” Huang added. “You know, everything from video analytics to image processing to voice to training to inference to data processing is now on one unified server.”

Here Huang is referring to the new DGX machine, also launched today. The DGX A100 is powered by eight A100 GPUs that together provide 320GB of memory with 12.4TB per second bandwidth. Six NVSwitches with third-generation NVLink fabric connect the GPUs, providing 4.8TB per second of bi-directional bandwidth. Each DGX 100 system offers nine Mellanox ConnectX-6 200Gb/s network interfaces and 15TB Gen4 NVMe storage.

Nvidia DGX A100 with nearly 5 petaflops FP16 peak performance (156 FP64 Tensor Core performance)

With the third-generation “DGX,” Nvidia made another noteworthy change. Instead of dual Broadwell Intel Xeons, the DGX A100 sports two 64-core AMD Epyc Rome CPUs. The move could signal Nvidia’s pushback on Intel’s emerging GPU play or may have been motivated by AMD’s price-performance story.

An Arm CPU option is also in the wings now, and with robust ongoing development activity around GPU-accelerated Arm, we would not be surprised to see another CPU shakeup in a fourth-gen DGX. Nvidia could choose an Arm server chip from Marvell or Ampere (imagine the co-branding opportunity) or it could decide to go full-bore and add an internally-developed Arm CPU to its ever-growing stack. CUDA 11 debuts full Arm64 support.

Speaking of Nvidia’s stack — with the introduction of DGX-1 four years ago on through the announced (and now completed) acquisition of Mellanox, Nvidia shied away from the systems-maker appellation, aware of the potential conflict with OEM partners. That reluctance seems to have have diminished.

A100 GPU HPC application speedups compared to NVIDIA Tesla V100 (Source: Nvidia)

“We develop vertically fully integrated systems to pioneer new form factors of computers,” said Paresh Kharya, director of product management for Nvdia’s datacenter and cloud platforms in the same media briefing. “We open the entire system as well and turn them into elemental building blocks. So our ecosystem, the entire industry can buy pieces or whole,” he continued, referencing the HGX A100 design that will be supported by partnering cloud service providers and server makers.

Huang later added, “Nvidia is really a datacenter scale computing company. We used to be a PC company 30 years ago, then we became a workstation company, then we became a server company, and we ultimately worked our way into the datacenter. But the future of computing: the datacenter is the computing unit. The future of computing is really datacenter scale; applications are going to run in the entire datacenter all at the same time.”

Nvidia also announced the creation of the DGX A100 SuperPod, spanning 140 DGX A100 systems (1,120 GPUs), 170 Mellanox Quantum 200G IB switches, 280 TB/s network fabric (over 15km of optical cable), and 4 petabyes of all-flash networked storage. The entire system delivers nearly 22 petaflops of peak double-precision performance (700 petaflops of “AI computing”) and was built in under three weeks, Nvidia said.

Nvidia has added four SuperPods to its internal supercomputer, called Saturn-V, boosting its total “AI supercomputing” capacity to nearly 5 exaflops, which Nvidia says makes Saturn-V the fastest AI supercomputer in the world. The total double-precision capacity of all those GPUs comes out to about 93 peak petaflops, but Huang clarified that Saturn-V is not really one system. It sits in four different locations under one management interface. Nvidia uses the system for software development applied to computer graphics, robotics, self-driving cars, healthcare and its new recommender system, Merlin.

DGX A100 cluster. Photo courtesy Argonne National Laboratory

But Nvidia has another reason for building Saturn-V. “If we’re a datacenter scale-company developing chips, systems and software at the datacenter scale, it stands to reason that we should build it ourselves,” said Huang.

The DGX A100 lists at $199,000 and is shipping now. The first order went to the U.S. Department of Energy’s Argonne National Lab, which is using the cluster to better understand and fight COVID-19. Other early adopters in the HPC research community include Indiana University, Jülich Supercomputing Center, Karlsruhe Institute of Technology, Max Planck Computing and Data Facility, and NERSC at the DOE’s Berkeley National Laboratory.

There is robust partner support. The list of cloud service providers and systems builders expected to integrate the A100s includes Alibaba Cloud, Amazon Web Services (AWS), Atos, Baidu Cloud, Cisco, Dell Technologies, Fujitsu, GIGABYTE, Google Cloud, H3C, Hewlett Packard Enterprise, Inspur, Lenovo, Microsoft Azure, Oracle, Quanta/QCT, Supermicro and Tencent Cloud.

The HGX A100 reference design, the backbone of the DGX A100, comes in a four- and an eight-GPU configuration. The four-GPU HGX A100 offers full interconnection between GPUs with NVLink, while the eight-GPU version provides full GPU-to-GPU bandwidth through NVSwitch. With the new multi-instance GPU (MIG) architecture, the Ampere servers, as we heard Huang call them, can be configured as 56 small GPUs for inferencing or as eight GPUs working together for training or HPC workloads.

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!

With New Owner and New Roadmap, an Independent Omni-Path Is Staging a Comeback

July 23, 2021

Put on a shelf by Intel in 2019, Omni-Path faced a uncertain future, but under new custodian Cornelis Networks, OmniPath is looking to make a comeback as an independent high-performance interconnect solution. A "significant refresh" – called Omni-Path Express – is coming later this year according to the company. Cornelis Networks formed last September as a spinout of Intel's Omni-Path division. Read more…

PEARC21 Panel Reviews Eight New NSF-Funded HPC Systems Debuting in 2021

July 23, 2021

Over the past few years, the NSF has funded a number of HPC systems to further supply the open research community with computational resources to meet that community’s changing and expanding needs. A review of these systems at the PEARC21 conference (July 19-22) highlighted... Read more…

Chameleon’s HPC Testbed Sharpens Its Edge, Presses ‘Replay’

July 22, 2021

“One way of saying what I do for a living is to say that I develop scientific instruments,” said Kate Keahey, a senior fellow at the University of Chicago and a computer scientist at Argonne National Laboratory, as s Read more…

PEARC21 Plenary Session: AI for Innovative Social Work

July 21, 2021

AI analysis of social media poses a double-edged sword for social work and addressing the needs of at-risk youths, said Desmond Upton Patton, senior associate dean, Innovation and Academic Affairs, Columbia University. S Read more…

Summer Reading: “High-Performance Computing Is at an Inflection Point”

July 21, 2021

At last month’s 11th International Symposium on Highly Efficient Accelerators and Reconfigurable Technologies (HEART), a group of researchers led by Martin Schulz of the Leibniz Supercomputing Center (Munich) presented a “position paper” in which they argue HPC architectural landscape... Read more…

AWS Solution Channel

Accelerate innovation in healthcare and life sciences with AWS HPC

With Amazon Web Services, researchers can access purpose-built HPC tools and services along with scientific and technical expertise to accelerate the pace of discovery. Whether you are sequencing the human genome, using AI/ML for disease detection or running molecular dynamics simulations to develop lifesaving drugs, AWS has the infrastructure you need to run your HPC workloads. Read more…

PEARC21 Panel: Wafer-Scale-Engine Technology Accelerates Machine Learning, HPC

July 21, 2021

Early use of Cerebras’ CS-1 server and wafer-scale engine (WSE) has demonstrated promising acceleration of machine-learning algorithms, according to participants in the Scientific Research Enabled by CS-1 Systems panel Read more…

With New Owner and New Roadmap, an Independent Omni-Path Is Staging a Comeback

July 23, 2021

Put on a shelf by Intel in 2019, Omni-Path faced a uncertain future, but under new custodian Cornelis Networks, OmniPath is looking to make a comeback as an independent high-performance interconnect solution. A "significant refresh" – called Omni-Path Express – is coming later this year according to the company. Cornelis Networks formed last September as a spinout of Intel's Omni-Path division. Read more…

Chameleon’s HPC Testbed Sharpens Its Edge, Presses ‘Replay’

July 22, 2021

“One way of saying what I do for a living is to say that I develop scientific instruments,” said Kate Keahey, a senior fellow at the University of Chicago a Read more…

Summer Reading: “High-Performance Computing Is at an Inflection Point”

July 21, 2021

At last month’s 11th International Symposium on Highly Efficient Accelerators and Reconfigurable Technologies (HEART), a group of researchers led by Martin Schulz of the Leibniz Supercomputing Center (Munich) presented a “position paper” in which they argue HPC architectural landscape... Read more…

PEARC21 Panel: Wafer-Scale-Engine Technology Accelerates Machine Learning, HPC

July 21, 2021

Early use of Cerebras’ CS-1 server and wafer-scale engine (WSE) has demonstrated promising acceleration of machine-learning algorithms, according to participa Read more…

15 Years Later, the Green500 Continues Its Push for Energy Efficiency as a First-Order Concern in HPC

July 15, 2021

The Green500 list, which ranks the most energy-efficient supercomputers in the world, has virtually always faced an uphill battle. As Wu Feng – custodian of the Green500 list and an associate professor at Virginia Tech – tells it, “noone" cared about energy efficiency in the early 2000s, when the seeds... Read more…

Frontier to Meet 20MW Exascale Power Target Set by DARPA in 2008

July 14, 2021

After more than a decade of planning, the United States’ first exascale computer, Frontier, is set to arrive at Oak Ridge National Laboratory (ORNL) later this year. Crossing this “1,000x” horizon required overcoming four major challenges: power demand, reliability, extreme parallelism and data movement. Read more…

Quantum Roundup: IBM, Rigetti, Phasecraft, Oxford QC, China, and More

July 13, 2021

IBM yesterday announced a proof for a quantum ML algorithm. A week ago, it unveiled a new topology for its quantum processors. Last Friday, the Technical Univer Read more…

ExaWind Prepares for New Architectures, Bigger Simulations

July 10, 2021

The ExaWind project describes itself in terms of terms like wake formation, turbine-turbine interaction and blade-boundary-layer dynamics, but the pitch to the Read more…

AMD Chipmaker TSMC to Use AMD Chips for Chipmaking

May 8, 2021

TSMC has tapped AMD to support its major manufacturing and R&D workloads. AMD will provide its Epyc Rome 7702P CPUs – with 64 cores operating at a base cl Read more…

Intel Launches 10nm ‘Ice Lake’ Datacenter CPU with Up to 40 Cores

April 6, 2021

The wait is over. Today Intel officially launched its 10nm datacenter CPU, the third-generation Intel Xeon Scalable processor, codenamed Ice Lake. With up to 40 Read more…

Berkeley Lab Debuts Perlmutter, World’s Fastest AI Supercomputer

May 27, 2021

A ribbon-cutting ceremony held virtually at Berkeley Lab's National Energy Research Scientific Computing Center (NERSC) today marked the official launch of Perlmutter – aka NERSC-9 – the GPU-accelerated supercomputer built by HPE in partnership with Nvidia and AMD. Read more…

Ahead of ‘Dojo,’ Tesla Reveals Its Massive Precursor Supercomputer

June 22, 2021

In spring 2019, Tesla made cryptic reference to a project called Dojo, a “super-powerful training computer” for video data processing. Then, in summer 2020, Tesla CEO Elon Musk tweeted: “Tesla is developing a [neural network] training computer called Dojo to process truly vast amounts of video data. It’s a beast! … A truly useful exaflop at de facto FP32.” 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…

CentOS Replacement Rocky Linux Is Now in GA and Under Independent Control

June 21, 2021

The Rocky Enterprise Software Foundation (RESF) is announcing the general availability of Rocky Linux, release 8.4, designed as a drop-in replacement for the soon-to-be discontinued CentOS. The GA release is launching six-and-a-half months after Red Hat deprecated its support for the widely popular, free CentOS server operating system. The Rocky Linux development effort... Read more…

CERN Is Betting Big on Exascale

April 1, 2021

The European Organization for Nuclear Research (CERN) involves 23 countries, 15,000 researchers, billions of dollars a year, and the biggest machine in the worl Read more…

Iran Gains HPC Capabilities with Launch of ‘Simorgh’ Supercomputer

May 18, 2021

Iran is said to be developing domestic supercomputing technology to advance the processing of scientific, economic, political and military data, and to strengthen the nation’s position in the age of AI and big data. On Sunday, Iran unveiled the Simorgh supercomputer, which will deliver.... Read more…

Leading Solution Providers

Contributors

HPE Launches Storage Line Loaded with IBM’s Spectrum Scale File System

April 6, 2021

HPE today launched a new family of storage solutions bundled with IBM’s Spectrum Scale Erasure Code Edition parallel file system (description below) and featu 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…

10nm, 7nm, 5nm…. Should the Chip Nanometer Metric Be Replaced?

June 1, 2020

The biggest cool factor in server chips is the nanometer. AMD beating Intel to a CPU built on a 7nm process node* – with 5nm and 3nm on the way – has been i Read more…

GTC21: Nvidia Launches cuQuantum; Dips a Toe in Quantum Computing

April 13, 2021

Yesterday Nvidia officially dipped a toe into quantum computing with the launch of cuQuantum SDK, a development platform for simulating quantum circuits on GPU-accelerated systems. As Nvidia CEO Jensen Huang emphasized in his keynote, Nvidia doesn’t plan to build... Read more…

Microsoft to Provide World’s Most Powerful Weather & Climate Supercomputer for UK’s Met Office

April 22, 2021

More than 14 months ago, the UK government announced plans to invest £1.2 billion ($1.56 billion) into weather and climate supercomputing, including procuremen Read more…

Q&A with Jim Keller, CTO of Tenstorrent, and an HPCwire Person to Watch in 2021

April 22, 2021

As part of our HPCwire Person to Watch series, we are happy to present our interview with Jim Keller, president and chief technology officer of Tenstorrent. One of the top chip architects of our time, Keller has had an impactful career. Read more…

Quantum Roundup: IBM, Rigetti, Phasecraft, Oxford QC, China, and More

July 13, 2021

IBM yesterday announced a proof for a quantum ML algorithm. A week ago, it unveiled a new topology for its quantum processors. Last Friday, the Technical Univer Read more…

Senate Debate on Bill to Remake NSF – the Endless Frontier Act – Begins

May 18, 2021

The U.S. Senate today opened floor debate on the Endless Frontier Act which seeks to remake and expand the National Science Foundation by creating a technology Read more…

  • arrow
  • Click Here for More Headlines
  • arrow
HPCwire