prespin
prespin
EDA for Inference Optimisation

Work with us

Trade-Off Precision for PPA

We use approximate arithmetic to unlock PPA improvements while maintaining model performance.

Early results

Google's open Coral NPU, BF16 multiply-accumulate datapath, 400 MHz. Added error stays within 1/3 of BF16 rounding error.

Area per product (µm²)
Coral BF16
356
Prespin
139
−61%, post place-and-route
Energy per product (pJ)
Coral BF16
1.00
Prespin
0.51
−49%, post place-and-route
Model perplexity
exact BF16
baseline
Prespin
+0.004%
95% CI −0.011% to +0.019%, Qwen2.5-0.5B WikiText-2