New Player
AI semiconductors developed by Korean fabless startup Rebellions outperformed those from global semiconductor giants such as Nvidia and Qualcomm in performance tests.
MLPerf AI semiconductor test results released by ML Commons, a non-profit organization on April 6 (local time) showed that the AI semiconductor Atom developed by Rebellions posted a latency of 4297 ms in a language model test. It was 1.4 times faster than that of Nvidia (T4 with a latency of 6093 ms) and 1.8 times faster than that of Qualcomm (Cloud AI100 with a latency of 7547 ms). This indicator is a single-stream test that measures the time it takes for one chip to process one input. It is used most importantly because it is not affected by the sizes and processes of semiconductors.
The Atom also recorded 1.4 times faster performances than its competitors in the field of vision which processes graphics. The Atom’s single-stream latency in the vision field was 0.239 ms, 1.4 times faster than that of Qualcomm (Cloud AI100 with a latency of 0.336 ms) and 3 times faster than that of Nvidia (A2 with a latency of 0.713 ms).
This Rebellions achievement is the first record in the Korean system semiconductor industry. In the meantime, Korean fabless companies have participated in MLPerf vision tests, but Rebellions is the first to beat Nvidia and Qualcomm in the language and vision fields at the same time. MLPerf is evaluated as the most reliable test in the AI semiconductor industry.