Startup Rebellions’ AI chip outperforms competitors

Rebellions Inc.‘s AI chip Atom [Photo provided by Rebellions]

Rebellions Inc. received top performance rating for its artificial intelligence chip from MLPerf benchmark, one of the most reliable tests to assess AI chip technologies, according to the company on Thursday.

The Korean startup submitted its AI chip Atom to have it tested for the MLPerf’s Bidirectional Encoder Representations from Transformers language model. The results showed that Atom was 1.5 times to twice as better than Cloud AI 100, the latest AI chip from Qualcomm Inc., or Nvidia Corp.’s A2 and T4 graphics processing units with equivalent specifications.

MLPerf is a system performance benchmark that measures the performance of individual machine learning workloads. AI technologies from global leading tech companies, like Qualcomm, Nvidia and Google LLC, have been tested under the benchmark, which mostly assesses the performance of the transformers, such as the GPT series.

“Our chip stands out because we have demonstrated global competitiveness in both linguistic and non-linguistic models,” said Rebellions Chief Executive Officer Park Sung-hyun. “Architectures that can accelerate various AI algorithms, including both linguistic and non-linguistic models, are the essence of AI chip design.”

In addition to language models, Rebellions’ Atom chip also surpassed performance standards in terms of the single-stream processing speed, performing more than 1.4 times better than Qualcomm’s AI chip and more than three times better than Nvidia‘s equivalent GPUs.

“It is significant that our chip outperformed those from market-dominant global players, particularly when AI chips are gaining attention around the world thanks to Chat GPT,” Park said.

Indeed, Rebellions has been receiving more requests for sample Atom chips from Big Tech data centers like Google and Microsoft. Rebellions released its AI chip for use at data centers earlier in February.

The company plans to commercialize its AI chip next year, “with the price at about one-tenth of Nvidia’s GPU A100,” the company said.