While AI products are becoming popular recently, when looking at the "Seeed Studio Grove AI HAT for Edge Computing Artificial Intelligence Board" it mentions a RISC-V, a 230GMULps 16-bit KPU (Neural Network Processor), a FPU, etc.

What is this "230GMULps"? What could be the processor architecture in it if it can be illustrated by an example?

Update: I initially thought of a specific model. As comments indicated that it is the number of multiplications, so I updated the question to ask for an example architecture.

  • 1
    \$\begingroup\$ GMul = billion multiplications, so 230 GMULPs = 230000000000 16-bit multiplications per second. \$\endgroup\$ – nanofarad Sep 18 '20 at 15:45
  • 1
    \$\begingroup\$ A complete guess, but ops might count things other than multiplication (e.g. the addition and accumulation that are also needed on a neural network and are likely cheaper/easier than a multiplication) \$\endgroup\$ – nanofarad Sep 18 '20 at 15:49
  • 1
    \$\begingroup\$ I cannot at the moment. It is merely a guess based on past experience without any specific architecture I can think of offhand to formally justify it, so I'd rather not post a full answer that I can't fully justify. \$\endgroup\$ – nanofarad Sep 18 '20 at 15:56
  • 1
    \$\begingroup\$ "Giga Multiplications Per Second" (GMULPS) measures the system throughput. Sometimes you see Giga Connection Updates Per Second (GCUPS) used as the metric, but it depends on the batch size. So some prefer a general throughput metric that can be applied to any batch size. I think they are claiming several thousand parallel multipliers. But why don't you just ask the seller? \$\endgroup\$ – jonk Sep 18 '20 at 16:13
  • 2
    \$\begingroup\$ There’s a lot of puffery in the AI industry. The most-abused metric might very well be TOPs. It’s not easy to actually achieve that number in practice. The more meaningful number is inferences/s for a given network. So, the network and the throughout you need should guide your decision, not TOPs. \$\endgroup\$ – hacktastical Sep 18 '20 at 17:46

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.