Fact Finder
Top 20
So after all that is the presenter being untruthful in saying AKIDA1000 is the most efficient of the semiconductors displayed for comparison purposes in the graphic?Not wanting to dampen the brilliant message of Akida being more efficient (as it is) but it gets up my nose to see wording such as "111x less power". More so because of the "111x less" than the fact that it is actually comparing Energy and not Power.
From a purely mathematical point of view 111 x less power is 1 - 111 = -110 meaning it generates 110 times the amount of power. How fantastic is that—when spotting keywords it actually generates power.
I assume they mean it uses 1/111th of the power or 0.9% of the power, which is 99.1% less and not 111x.
Or maybe even it is 99% more efficient!
I won't get into the inaccuracy of calling it a reduction in power though, Joules measure energy, power is measured in Watts or Joules per second. When converting to instantaneous power, the comparison is very different. But it is overall energy used that IS important anyway. And in this chart Akida uses less energy at 5MHz, and more energy at 300Mhz.
So that it is 99% more efficient is correct.
The bar charts appear to give a much better representation—but these are equally confusing in that they are NOT comparing like against like.
But what would be a better thing to report is total bang for buck - as in number of inferences per Watt. i.e. standard scientific normalization to a constant denominator (the Watt).
Let's round for ease of visualizing
ARM Cortex M4 (120MHz) = 150 inf/W
Jetson Nano - (921 MHz) = 1080 inf/W
Jetson Nano (5W - 614 MHz) = 1520 inf/W
Coral Dev Board = 2850 inf/W
AKD 1000 (300MHz) = 17000 inf/W
AKD 1000 (5MHz) = 27000 inf/W
Now we can say roughly 10x the number of inferences for the same power usage, comparing best against the best. And the comparison becomes 180x the number of inferences for the same power usage when comparing best Akida to worst (ARM).
And Also the ARM Cortex looks far inferior now. And I don't think that was the intention of this presentation.
This leads to yet another of my pet hates. Statistics, damned lies, and statistics. "Facts" can be presented any way to show the message you want to convey.
Charts SHOULD ALWAYS be normalized so as to show same versus same.
Regards
FF
AKIDA BALLISTA
Last edited: