Part of one of Google’s Cloud TPU v4 Pods
Digital transformation is responsible for creating artificial intelligence workloads on an unprecedented scale. These workloads require companies to collect and store mountains of data. Even though business intelligence is extracted from current machine learning models, new data streams are used to create new models and update existing models.
Advertising
Building AI models is complex and expensive. It is also very different from traditional software development. AI models need specialized hardware for accelerated computing and high-performance storage, as well as purpose-built infrastructure to handle the technical nuances of AI.
In today’s world, many critical business decisions and customer services rely on accurate machine learning insights. To train, run, and scale models as quickly and accurately as possible, a company has the knowledge to choose the best hardware and software for its machine learning applications.
Performance Calibration
ML Commons
MLCommons is an open engineering consortium that has made it easier for companies to make machine learning decisions with its standardized benchmarking. Its mission is to make machine learning better for everyone. Testing is done and unbiased comparisons help companies determine which vendor best meets their AI application requirements. The MLCommons Foundation began its first MLPerf benchmarking in 2018.
MLcommons recently conducted a benchmarking program called MLPerf v2.0 Training to measure the performance of hardware and software used to train machine learning models. 250 performance results were reported by 21 different bidders including Azure, Baidu
BIDU
BIDU
GOOG
GOOG
the
the
VIDI
The
IAD
VIDI
The
IAD
This series of tests focused on determining how long it takes to train various neural networks. Faster model training allows for faster model deployment, which impacts total cost of ownership and model ROI.
A new object detection benchmark has been added to MLPerf Training 2.0, which trains the new RetinaNet reference model on a larger and more diverse dataset called Open Images. This new test reflects state-of-the-art ml training for applications such as collision avoidance for vehicles and robotics, retail analytics and many more.
Results
Machine learning has seen a lot of innovation since 2021, both in hardware and software. For the first time since the debut of MLPerf, Google’s cloud-based TPU v4 ML supercomputer outperformed NVIDIA A100 in four of eight training tests covering language (2), computer vision (4), learning by reinforcement (1) and recommendation systems (1).
The higher, the better. TPUs showed significant acceleration in all five benchmarks released during the … [+]
According to the graph comparing the performance of Google and nvidia, Google had the fastest training times for BERT (language), ResNet (image recognition), RetinaNet (object detection), and MaskRCNN (image recognition). On DLRM (recommendation), Google narrowly edged out NVIDIA, but it was a research project and unavailable for public use.
Overall, Google submitted scores for five of the eight benchmarks, the best training times are shown below:
Data: MLCommons
In a chat with Vikram Kasivajhula, Google’s director of product management for ML infrastructure, I asked what approach Google was using to make such dramatic improvements to TPU v4.
“We focused on the problems of heavy model users innovating at the frontiers of machine learning,” he said. “Our cloud product is actually an instantiation of that goal. We also focused on performance per dollar. As you can imagine, these models get incredibly large and expensive to train. One of our priorities is to make sure it is affordable. »
A one-of-a-kind submission
A unique submission was made to MLPerf Training 2.0 by Stanford graduate student Tri Dao. Dao submitted an 8-A100 system for BERT training.
NVIDIA also had a submission using the same setup as Dao. I suspect this was a courtesy submission from NVIDIA to provide Dao with a documented point of comparison.
NVIDIA completed training the BERT model with its 8-A100 in 18.442 minutes while Dao’s submission took only 17.402 minutes. He got faster training time using a method called FlashAttention. Attention is a technique that mimics cognitive attention. The effect enhances some parts of the input data while decreasing other parts – the motivation is that the network should focus more on the small but important parts of the data.
Wrap
Over the past three years, Google has made a lot of progress with its TPU. Likewise, NVIDIA has been using its A100 successfully for four years. Much of the software improvement has been brought to the A100, as evidenced by its long history of achievement.
We’re likely to see NVIDIA submissions in 2023 using both its A100 and the new H100, a beast by any current standard. Everyone was hoping to see the performance of the H100 this year, but NVIDIA didn’t submit it because it wasn’t publicly available.
Software improvements in general were evident in the latest results. Kasivajhula said hardware is only half the story of Google’s improved benchmarks. The other half was software optimizations.
“Many of the optimizations were learned from our own industry-leading YouTube and search benchmark use cases,” he said. “We are now making them available to users. »
Google has also made several performance improvements to the virtualization stack to fully utilize the computing power of CPU hosts and TPU chips. The results of Google’s software improvements have been demonstrated by its peak performance on the image and recommendation models.
Overall, Google’s Cloud TPUs deliver significant performance and cost savings at scale. It will take time to see if the benefits are enough to entice more customers to switch to Google Cloud TPUs.
In the longer term, Google’s better results in the main categories could presage that NVIDIA will achieve less MLPerf results in the future. It is in the interest of the ecosystem to see strong controversies between multiple vendors for the best MLPerf performance results.
One thing is for sure, MLPerf Training 2.0 was much more interesting than previous rounds where NVIDIA picked up performance wins in almost every category.
Full results of MLPerf Training 2.0 are available here.
Paul Smith-Goodson is Vice President and Principal Analyst for Quantum Computing, Artificial Intelligence and Space at Moor Insights and Strategy. You can follow him on To babble for current information on quantum, AI and space.
Note: Moor Insights & Strategy writers and editors may have contributed to this article.
Moor Insights & Strategy, like all research and technology industry analyst firms, provides or has provided paid services to technology companies. These services include research, analysis, consulting, consulting, benchmarking, acquisition matching, and speaking sponsorships. Company has had or currently has paid business relationships with 8×8, Accenture
ACN
ACN
ATEN
ATEN
AMD
AMD
AMZN
AMZN
T
T
AVGO
AVGO
CALX
CALX
CSCO
CSCO
CLDR
CLDR
VALLEY
VALLEY
EXT
EXT
vmw
vmw
IBM
IBM
JBL
JBL
the
the
MRVL
MRVL
MU
MU
MSFT
MSFT
ITENA
ITENA
the
the
QCOM
QCOM
NAKED
NAKED
ORCL
ORCL
PANW
PANW
PXLX
PXLX
PLT
PLT
the
the
RMBS
RMBS
RHT
RHT
S
S
NLOK
NLOK
SYNA
SYNA
the
the
CDT
CDT
VZ
VZ
XLNX
XLNX
ZEN
ZEN
SZ
SZ