Performance

Here, we are presenting a performance summary of the RBLN ATOM and comparing it to the GPUs and NPUs widely used for inference. All of the numbers described in this table are taken from the official MLPerf^TM Inference v3.0 results.

• Vision / ResNet50

  Vendor	Accelerator	Single Stream Latency	Relative Latency
  Rebellions	RBLN ATOM	0.239 ms	x1.0
  Qualcomm	Cloud AI100	0.336 ms	x1.4
  Nvidia	A2 (Ampere)	0.713 ms	x3.0
  Nvidia	T4 (Turing)	0.818 ms	x3.4

• Language / BERT-Large

  Vendor	Accelerator	Single Stream Latency	Relative Latency
  Rebellions	RBLN ATOM	4.297 ms	x1.0
  Qualcomm	Cloud AI100	7.547 ms	x1.8
  Nvidia	A2 (Ampere)	8.506 ms	x2.0
  Nvidia	T4 (Turing)	6.093 ms	x1.4

For more details, please visit the official MLPerf^TM Inference v3.0 website.