The February 2023 issue of IEEE Spectrum is here!

Close bar

These Might Be the Fastest (and Most Efficient) AI Systems Around

Almost 2000 entries ranked in MLPerf’s latest inferencing list

5 min read
NVIDIA A30 Tensor Core GPU

The machine learning industry’s efforts to measure itself using a standard yardstick has reached a milestone. Forgive the mixed metaphor, but that’s actually what’s happened with the release of MLPerf Inference v1.0 today. Using a suite of benchmark neural networks measured under a standardized set of conditions, 1,994 AI systems battled it out to show how quickly their neural networks can process new data. Separately, MLPerf tested an energy efficiency benchmark, with some 850 entrants for that.

This contest was the first following a set of trial runs where the AI consortium MLPerf and its parent organization MLCommons worked out the best measurement criteria. But the big winner in this first official version was the same as it had been in those warm-up rounds—Nvidia.

Entries were combinations of software and systems that ranged in scale from Raspberry Pis to supercomputers. They were powered by processors and accelerator chips from AMD, Arm, Centaur Technology, Edgecortix, Intel, Nvidia, Qualcomm, and Xilininx. And entries came from 17 organizations including Alibaba, Centaur, Dell Fujitsu, Gigabyte, HPE, Inspur, Krai, Lenovo, Moblint, Neuchips, and Supermicro.

Despite that diversity most of the systems used Nvidia GPUs to accelerate their AI functions. There were some other AI accelerators on offer, notably Qualcomm’s AI 100 and Edgecortix’s DNA. But Edgecortix was the only one of the many, manyAI accelerator startups to jump in. And Intel chose to show off how well its CPUs did instead of offering up something from its US $2-billion acquisition of AI hardware startup Habana.

Before we get into the details of whose what was how fast, you’re going to need some background on how these benchmarks work. [Click here if you want to skip the background.] MLPerf is nothing like the famously straightforward Top500 list of the supercomputing great and good, where a single value can tell you most of what you need to know. The consortium decided that the demands of machine learning is just too diverse to be boiled down to something like tera-operations per watt, a metric often cited in AI accelerator research.

First, systems were judged on six neural networks. Entrants did not have to compete on all six, however.

  • BERT, for Bi-directional Encoder Representation from Transformers, is a natural language processing AI contributed by Google. Given a question input, BERT predicts a suitable answer.
  • DLRM, for Deep Learning Recommendation Model is a recommender system that is trained to optimize click-through rates. It’s used to recommend items for online shopping and rank search results and social media content. Facebook was the major contributor of the DLRM code.
  • 3D U-Net is used in medical imaging systems to tell which 3D voxel in an MRI scan are parts of a tumor and which are healthy tissue. It’s trained on a dataset of brain tumors.
  • RNN-T, for Recurrent Neural Network Transducer, is a speech recognition model. Given a sequence of speech input, it predicts the corresponding text.
  • ResNet is the granddaddy of image classification algorithms. This round used ResNet-50 version 1.5.
  • SSD, for Single Shot Detector, spots multiple objects within an image. It’s the kind of thing a self-driving car would use to find important things like other cars. This was done using either MobileNet version 1 or ResNet-34 depending on the scale of the system.

Competitors were divided into systems meant to run in a datacenter and those designed for operation at the “edge”—in a store, embedded in a security camera, etc.

Datacenter entrants were tested under two conditions. The first was a situation, called “offline”, where all the data was available in a single database, so the system could just hoover it up as fast as it could handle. The second more closely simulated the real life of a datacenter server, where data arrives in bursts and the system has to be able to complete its work quickly and accurately enough to handle the next burst.

Edge entrants tackled the offline scenario as well. But they also had to handle a test where they are fed a single stream of data, say a single conversation for language processing, and a multistream situation like a self-driving car might have to deal with from its multiple cameras.

Got all that? No? Well, Nvidia summed it up in this handy slide:

MLPERF inference 1.0 slideImage: NVIDIA

And finally, the efficiency benchmarks were done by measuring the power draw at the wall plug and averaged over 10 minutes to smooth out the highs-and-lows caused by processors scaling their voltages and frequencies.

Here, then, are the tops for each category:


Datacenter (commercially available systems, ranked by server condition)

Image ClassificationObject DetectionMedical ImagingSpeech-to-TextNatural Language ProcessingRecommendation
System nameNF5488A5Dell EMC DSS 8440 (10x A100-PCIe-40GB)NVIDIA DGX-A100 (8x A100-SXM-80GB, TensorRT)Dell EMC DSS 8440 (10x A100-PCIe-40GB)Dell EMC DSS 8440 (10x A100-PCIe-40GB)NF5488A5
ProcessorAMD EPYC 7742Intel(R) Xeon(R) Gold 6248 CPU @ 2.50GHzAMD EPYC 7742Intel(R) Xeon(R) Gold 6248 CPU @ 2.50GHzIntel(R) Xeon(R) Gold 6248 CPU @ 2.50GHzAMD EPYC 7742
No. Processors222222
No. Accelerators810810108
Server queries/s271,2468,265479.65107,98726,7492,432,860
Offline samples/s307,2527,612479.65107,26929,2652,455,010

Edge (commercially available, ranked by single-stream latency)

Image ClassificationObject Detection (small)Object Detection (large)Medical ImagingSpeech-to-TextNatural Language Processing
System nameNVIDIA DGX-A100 (1x A100-SXM-80GB, TensorRT, Triton)NVIDIA DGX-A100 (1x A100-SXM-80GB, TensorRT, Triton)NVIDIA DGX-A100 (1x A100-SXM-80GB, TensorRT, Triton)NVIDIA DGX-A100 (1x A100-SXM-80GB, TensorRT)NVIDIA DGX-A100 (1x A100-SXM-80GB, TensorRT)NVIDIA DGX-A100 (1x A100-SXM-80GB, TensorRT)
ProcessorAMD EPYC 7742AMD EPYC 7742AMD EPYC 7742AMD EPYC 7742AMD EPYC 7742AMD EPYC 7742
No. Processors222222
No. Accelerators111111
Single stream latency (milliseconds)0.4313690.255811.68635319.91908222.5852031.708807
Multiple stream (streams)1344192056
Offline samples/s38011.650926.6985.51860.607314007.63601.96

The Most Efficient


Image ClassificationObject DetectionMedical ImagingSpeech-to-TextNatural Language ProcessingRecommendation
System nameGigabyte R282-Z93 5x QAIC100Gigabyte R282-Z93 5x QAIC100Gigabyte G482-Z54 (8x A100-PCIe, MaxQ, TensorRT)NVIDIA DGX Station A100 (4x A100-SXM-80GB, MaxQ, TensorRT)NVIDIA DGX Station A100 (4x A100-SXM-80GB, MaxQ, TensorRT)NVIDIA DGX Station A100 (4x A100-SXM-80GB, MaxQ, TensorRT)
ProcessorAMD EPYC 7282 16-Core ProcessorAMD EPYC 7282 16-Core ProcessorAMD EPYC 7742AMD EPYC 7742AMD EPYC 7742AMD EPYC 7742
No. Processors222111
No. Accelerators558444
Server queries/s78,502155737243,38910,203890,334
System Power (Watts)5345482261131413021342

Edge (commercially available, ranked by single-stream latency)

Image ClassificationObject Detection (small)Object Detection (large)Medical ImagingSpeech-to-TextNatural Language Processing
System nameAI Development KitNVIDIA Jetson Xavier NX (MaxQ, TensorRT)AI Development KitNVIDIA Jetson Xavier NX (MaxQ, TensorRT)NVIDIA Jetson Xavier NX (MaxQ, TensorRT)NVIDIA Jetson Xavier NX (MaxQ, TensorRT)
ProcessorQualcomm Snapdragon 865NVIDIA Carmel (ARMv8.2)Qualcomm Snapdragon 865NVIDIA Carmel (ARMv8.2)NVIDIA Carmel (ARMv8.2)NVIDIA Carmel (ARMv8.2)
No. Processors111111
AcceleratorQUALCOMM Cloud AI 100 DM.2eNVIDIA Xavier NXQUALCOMM Cloud AI 100 DM.2NVIDIA Xavier NXNVIDIA Xavier NXNVIDIA Xavier NX
No. Accelerators111111
Single stream latency0.851.6730.44819.08372.3757.54
System energy/stream (joules)

The continuing lack of entrants from AI hardware startups is glaring at this point, especially considering that many of them are members of MLCommons. When I’ve asked certain startups about it, they usually answer that the best measure of their hardware is how it runs their potential customers’ specific neural networks rather than how well they do on benchmarks.

That seems fair, of course, assuming these startups can get the attention of potential customers in the first place. It also assumes that customers actually know what they need.

“If you’ve never done AI, you don’t know what to expect; you don’t know what performance you want to hit; you don’t know what combinations you want with CPUs, GPUs, and accelerators,” says Armando Acosta, product manager for AI, high-performance computing, and data analytics at Dell Technologies. MLPerf, he says, “really gives customers a good baseline.”

Due to author error a mixed metaphor was labelled as a pun in an earlier version of this post. And on 28 April the post was corrected, because the column labels “Object Detection (large)” and “Object Detection (small)” had been accidentally swapped.

The Conversation (0)

Will AI Steal Submarines’ Stealth?

Better detection will make the oceans transparent—and perhaps doom mutually assured destruction

11 min read
A photo of a submarine in the water under a partly cloudy sky.

The Virginia-class fast attack submarine USS Virginia cruises through the Mediterranean in 2010. Back then, it could effectively disappear just by diving.

U.S. Navy

Submarines are valued primarily for their ability to hide. The assurance that submarines would likely survive the first missile strike in a nuclear war and thus be able to respond by launching missiles in a second strike is key to the strategy of deterrence known as mutually assured destruction. Any new technology that might render the oceans effectively transparent, making it trivial to spot lurking submarines, could thus undermine the peace of the world. For nearly a century, naval engineers have striven to develop ever-faster, ever-quieter submarines. But they have worked just as hard at advancing a wide array of radar, sonar, and other technologies designed to detect, target, and eliminate enemy submarines.

The balance seemed to turn with the emergence of nuclear-powered submarines in the early 1960s. In a 2015 study for the Center for Strategic and Budgetary Assessment, Bryan Clark, a naval specialist now at the Hudson Institute, noted that the ability of these boats to remain submerged for long periods of time made them “nearly impossible to find with radar and active sonar.” But even these stealthy submarines produce subtle, very-low-frequency noises that can be picked up from far away by networks of acoustic hydrophone arrays mounted to the seafloor.

Keep Reading ↓Show less