Artificial intelligence (AI) semiconductor company Graphcore has once again demonstrated the competitiveness of its Intelligence Processing Unit (IPU) system by achieving record-breaking performance in the latest MLPerf 1.1 benchmark tests.

Graphcore held an online press conference on the 9th, announcing the latest IPU-POD system and MLPerf benchmark test results, and introducing Graphcore's plans for accelerating domestic AI innovation, as well as its domestic business performance and direction.

At this press conference, VP of Global Sales, Febriz Moizan, expressed confidence that the Graphcore IPU system outperformed the NVIDIA DGX A100.

In a recent MLPerf benchmark, Graphcore IPU-POD16 outperformed NVIDIA's DGX A100 in training the computer vision model ResNet-50.

Training ResNet-50 The NVIDIA DGX A100 took 29.1 minutes, while Graphcore's IPU-POD16 took 28.3 minutes.

This represents a 24% performance improvement over the first MLPerf test results using software alone, which is especially noteworthy given that GPUs are typically used to train ResNet-50 models.

Graphcore also recently released benchmark results for its newly released IPU-POD128 and IPU-POD256.

Graphcore submitted the system to the MLPerf 'Commercially Available' category, demonstrating its commitment to continuous system scaling and performance improvement.

For the IPU-POD128 and IPU-POD256, which boast the highest performance ever achieved on Graphcore IPU systems, the training time for the ResNet-50 model was only 5.67 minutes and 3.79 minutes, respectively.

For the natural language processing (NLP) model BERT, Graphcore submitted IPU-POD16, IPU-POD64, and IPU-POD128 training data to both the Open and Closed categories.

In particular, the latest IPU-POD128 showed superior performance in the open sector with a training time of 5.78 minutes.

Overall, the BERT model training performance improved by 5% for IPU-POD16 and 12% for IPU-POD64 compared to the previous MLPerf benchmark.

Additionally, Graphcore's flagship product, the IPU-POD256, demonstrated the potential for tangible performance benefits by recording just 1.8 hours for training the EfficientNet B4 model.

Looking at the raw data from the MLPerf tests, we can see that each vendor's system is connected to a host processor.The sheer number of processors is striking. Some participating companies even designate one CPU for every two AI processors.

In contrast, Graphcore consistently maintains the lowest host processor-to-IPU ratio. The IPU uses the host server only for data movement, eliminating the need for the host server to dispatch code at runtime.

Therefore, the fewer host servers required for an IPU system, the more flexible and efficient expansion is possible.

For natural language processing models such as BERT-Large, IPU-POD64 requires only one dual-CPU host server.

Because ResNet-50 requires more host processor support for image preprocessing, four dual-core servers are assigned per IPU-POD64. This results in a system-to-host processor ratio of 1:8, which is lower than all other systems participating in MLPerf. It's also noteworthy that in this MLPerf 1.1 benchmark test, Graphcore achieved the fastest single-server training time for BERT training, taking just 10.6 minutes.

Graphcore has been committed to improving AI training performance through ongoing optimizations of its Poplar software development kit (SDK) and the release of new IPU-POD products. Notably, following the first MLPerf test in June, the company demonstrated significant improvements this time, demonstrating Graphcore's unwavering commitment to continuous innovation.

Graphcore has taken a fundamentally different 'innovative approach' from other companies in the industry from the system design stage, such as separating the host server and AI computing. Graphcore is conducting software updates to improve performance at least every three months and is working on implementing and optimizing new models and workloads for IPUs.