1 Based on BERT-Large 128 sequence length training, 64 replica count, 1k batch size. 1.4x faster = Nvidia DGX A100 20,086 samples/second throughput, SambaNova Systems DataScale SN10-8R 28,800 samples/second throughput

Based on MLperf results published as of 11/24/2020.  Inference throughput SambaNova Systems RDU 8632 samples per second and Nvidia A100 1251 samples per second, 8632 / 1251 = 6.90x. Inference latency SambaNova Systems RDU 0.116 samples per second and Nvidia A100 0.799 samples per second, 0.799 / 0.116 = 6.98879x. 

3 CosmicTagger batch size 64, image 1,280×2,048. SambaNova Systems Cardinal SN10 RDU accuracy = 0.902399368584156

4 Based on MLPerf results published as of 11/24/2020 comparing the MLPerf accuracy threshold of 80.25%, with Intel Xeon CPU and Nvidia A100 GPU showing 80.27% accuracy.  SambaNova Systems Cardinal SN10-8R shows 80.46% accuracy achievement.

Based on customer testimonial referenced in Argonne National Labrotory video of system installation and running custom customer model in under 45 minutes