Interesting MS is specifying 40 TSOPs on the CPU or the CPU makers have decided that.
MS Windows does seems to need more than than some alternative OS's.
Fp16 seems to be the default for AI/ML.
RTX4080 is max 48TFLOPs
I have no idea what the Pi5 is doing, but with more optimized code it is not bad, nearly usable Image may be NSFW.
Clik here to view.
ARM have their Ethos tech, which might be useful for Pi6, the N78 is 1 to 10 TSOP Image may be NSFW.
Clik here to view.
ARMs Compute library has these types.
Will the RTX5090 have Int8/Binary ops?
Compute costs $$ and power, more compute at lower power/$ is good.
Needed in fact as AGI is unleashed.
MS Windows does seems to need more than than some alternative OS's.
Fp16 seems to be the default for AI/ML.
RTX4080 is max 48TFLOPs
I have no idea what the Pi5 is doing, but with more optimized code it is not bad, nearly usable Image may be NSFW.
Clik here to view.
ARM have their Ethos tech, which might be useful for Pi6, the N78 is 1 to 10 TSOP Image may be NSFW.
Clik here to view.
With the performance increases I see with Binary and Ternary NN perhaps Ethos gen3 can do better?The Arm Ethos-N NPUs improve the inference performance of neural networks. The NPUs target 8-bit integer quantized Convolutional Neural Networks (CNN).
ARMs Compute library has these types.
While some like to think CUDA on Nvidia is a must have for Desktops, software optimization is being improved.Support for multiple data types: FP32, FP16, INT8, UINT8, BFLOAT16
Will the RTX5090 have Int8/Binary ops?
Compute costs $$ and power, more compute at lower power/$ is good.
Needed in fact as AGI is unleashed.
Statistics: Posted by Gavinmc42 — Sun Mar 24, 2024 5:46 am
