3 ms·
To summarize, it appears that their new chips can do 96x96 dot product in a single cycle (multiplying neural network weights by values), and have hardware for a
by andr 7y ago
To summarize, it appears that their new chips can do 96x96 dot product in a single cycle (multiplying neural network weights by values), and have hardware for activations (ReLU, softmax, tanh) and pooling (as in convolutional neural networks). This results in a crazy 144 trillions ops/second.
How does this compare to TPUs and the Neural Engine in iPhone CPUs?
- LoSboccacc 7y agosomething I don't understand, is this chip brand new? is it essential to the autonomous drive?
- DeonPenny 7y agoChip is new and is needed for autonomous driving
- Traster 7y agoCan they actually do a 96x96 dot product in a single cycle? My understanding was that they can do 1 96x96 dot product per cycle - that is to say that their throughput is a 96x96 dot product but it'll be heavily pipelined.