Is it possible to directly pipe analog signals between analog tiles?

[dingandrew]

From my understanding the column wires of the crossbar tiles contain the summed currents which are then passed to an op-amp current-to-voltage converter then to an ADC. I would like to pass this output voltage signal directly to another crossbar input, completely skipping the ADC converter.

I know the ADC can be disabled, but I believe this has the effect of perfect ADC conversions. I would like to maintain an analog signal, with all of its noise, as inputs to another crossbar tile.

Is this possible?

[maljoras]

Hi @dingandrew,
that is correct, the summed currents are the outputs. To "turn-off" all digital compute, you could set the inp_res and out_res to -1 to avoid any quantization. The inp_bound and out_bound needs to be adjusted, too, those will clip the inputs and outputs into a certain range. Finally, you need to make sure that all peripheral digital compute is turned off, that is forward/backward.noise_management = NoiseManagementType.NONE and forward.bound_management=BoundManagementType.NONE as well as weight_scaling_omega=0 and forward.out_scale=1.0.

[maljoras]

We should add some kind of "AnalogOutputPreset" to the Presets to make this settings simpler.

[kaoutar55]

@maljoras would this preset allow disabling ADCs? then does it make sense also to have also an DigitalOutputPreset?

[maljoras]

@kaoutar55 The default is an assumption that we have digital outputs. Otherwise using activations etc in pytorch does not make much sense

[dingandrew]

Hi @maljoras,

So this would have the effect of skipping the ADC right?

And this also brings up another point. Is it possible to extend the analog layers to implement activation/pooling layers. For example, a RELU could be implemented with a op-amp. The idea behind this is that the calculations would stay in the analog domain for longer, avoiding costly A-D or D-A conversions.

[maljoras]

You could simply add an analog pooling layer (with the expected noise sources etc) to your liking by adding a pytorch module and build a DNN using that custom pooling (while AnalogLayers have ADC switched off). This is the advantage of using pytorch as it is very flexible in adding custom layers.

[dingandrew]

Would this be the correct RPU configuration for turning off all digital circuitry?

mapping = MappingParameter( weight_scaling_omega=0.0) 


rpu_config = SingleRPUConfig(device=ConstantStepDevice(), mapping=mapping)

rpu_config.forward.inp_res = -1
rpu_config.forward.out_res = -1
rpu_config.forward.inp_bound = 100
rpu_config.forward.out_bound = 100
rpu_config.forward.w_noise_type = WeightNoiseType.ADDITIVE_CONSTANT
rpu_config.forward.out_scale = 1.0
rpu_config.forward.bound_management = BoundManagementType.NONE
rpu_config.forward.noise_management = NoiseManagementType.NONE
rpu_config.backward.noise_management = NoiseManagementType.NONE
    

[maljoras]

The inp_bound and out_bound should probably be a smaller number. Even for a purely analog circuit, you will have some output and input bounds as the currents will saturate at some point. There might be also additional non-idealities, you might want to consider. Eg. ir_drop could be enabled, or you could add some output non-linearity / noise with some additional pytorch code.