AM62A7-Q1: C7x+MMA versions between AM62A and TDA4VEN

Hello,

Yes, you are correct that these C7x's are not identical. AM62A uses C7504 whereas TDA4AEN (architecturally, very similar to AM67A) uses C7524. This is generally not exposed because C7 is a closed architecture such that the low-level features are not directly accessible -- C7xMMA is treated as a black-box AI accelerator.

That said, there are a few differences that impact the user-level

• Increased L2 cache size (1.25 MB in AM62A vs. 2.25 MB in TDA4AEN)
• Histogram / lookup table support in AEN, not present in AM62A. This mainly impacts nonlinear activations
• 2nd C7xMMA on AEN -- each C7xMMA is a 256-bit variant that equates to 2 TOPS

BR,
Reese

Thanks for the clarification! 

Can you clarify what  nonlinear activations are?

Ah, yes of course.

I'll first note that there are many good resources online to describe what these are and why they exist. They may give a more intuitive explanation than my text below

In short form, nonlinear-activations are generally pointwise functions that happen on the output of larger layers (e.g. convolution) that had mixed together many points/values. This makes a sort of 'break' in an otherwise linear algorithm that lets it generalize to more complex functions and patterns in the underlying data. In essence, neural networks MUST have these non-linearities to function on complex data.

Some examples of this are ReLU, SiLU, sigmoid, tanh (I'd recommend checking out an image of the x-y plot for these. Some are really simple like RELU, such that there's almost no extra compute involved. Others like SiLU or tanh are more complex to compute -- for these, a look-up table is a good way to accelerate for quantized data. For 8-bit quantization, you basically need a 256-element table, from which the non-linear result is fetched.

There's a mention of some of these activations here in our documentation: https://github.com/TexasInstruments/edgeai-tidl-tools/blob/master/docs/supported_ops_rts_versions.md#feature-set-comparison-across-devices.

BR,
Reese

Hi Reese 

thanks for the good read. I'll take a deeper dive in this. in the meantime, my customer and I put together this table, can you help me fill out the rest? 

Thanks!