Omer Faizan
← Research

Research poster · 2025

Spiking networks on neuromorphic hardware

Omer Faizan · M.S. Data Science, Montclair State University

1 Question

How do accuracy, latency and energy change when a spiking network runs on event-driven neuromorphic hardware instead of a GPU?

2 Why it matters

Brains compute with sparse spikes on a tiny power budget. GPUs compute everything, every step. Neuromorphic chips try to close that gap.

20 Wthe human brain, ~86 billion neurons
65 mWIBM TrueNorth, 1 million neurons
1 Mneurons on one Intel Loihi 2 chip

3 Where the energy goes

Energy per operation, 45 nm (log scale)

32-bit int add0.1 pJ32-bit float add0.9 pJ32-bit float multiply3.7 pJ32-bit SRAM read5 pJ32-bit DRAM read640 pJ

compute memory access

Fetching a number from DRAM costs ~170× a float multiply. Neuromorphic chips keep memory next to the neurons.

4 Plan

trained SNNSTDP · PyTorch GPUdense, clocked neuromorphicevent-driven compareaccuracy · timeenergy

5 The hardware

ChipYearNeuronsSynapses
IBM TrueNorth20141 M256 M
Intel Loihi2018128 k128 M
Intel Loihi 220211 M120 M

6 What I'll measure

GPU accuracy & latency from my split-learning runs · rest projected

GPUNeuromorphic
Accuracy~72%~70%
Latency / image~3 ms~5 ms
Energy / image~400 mJ~2 mJ
Energy saving1×~200×

7 What I expect

  • Energy per image drops sharply; accuracy drops a little.
  • The sparser the layer, the bigger the saving.