TABLE II.

Comparison with the State-of-the-Art Neural Interface SoCs with On-Chip ML

Parameter	JSSC’15 [5]	JSSC’18 [15]	JETCAS’18 [16]	ISSCC’18 [17]	ISSCC’20 [18]	JSSC’20 [19]	TBioCAS’21 [20]	VLSI’21 [21]		JSSC’22 [22]	This Work
Process (nm)	180	180	65	130	65	40	180	28		40	65
Supply Voltage (V)	1.8/1	1.8	0.8	1.2	1.2	0.58	1.5	0.5		1.1/0.7	1.2
# Recording Ch.	16	16	32^	32	8	14^	8	8^		16	256
# Stimulation Ch.	1	16	-	32	1	-	2	-		1	16
Area/ch.* (mm2)	1.563	0.95	0.031^	0.237	0.244	0.182^	0.729	0.013^		0.13	0.014
Energy Efficiency (μJ/class)	2.73	64§	0.0412^	178.7§	0.036^	170.9^	174.4§	0.0015^		0.97	0.227
Memory (kB)	64†	-	1	96†	-	35.8†	64	0.2		134†	3.17‡
Classifier	D2A-LSVM	Ridge Regression	XGB-DT	EDM-SVM	BrainForest	NL-SVM	Coarse (Th.) Fine (LS-SVM)	LR + SGD		GTCA-SVM	NeuralTree
Feature Set	SE	FFT, ApEn	LL, POW, VAR, SE	PLV, CFC, SE	RAF Neurons	SE	MODWT-KDE	SE, LL		SE	PLV, PAC, SE, HFOR, LL, LMP, Hjorth
Dataset (# Patients)	MIT (14)	Local (5)	iEEG (20)	EU-iEEG (-)	EU-iEEG (-)	MIT (24)	MIT (23)	UoM (3)	MIT (24)	MIT (24)	MIT (24)	iEEG(6)
Sensitivity (%)	95.7	96	83.7	100	96.7	96.6	97.8	97.9	97.5	100	95.6	94
Specificity (%)	98	100	88.1	0.81FP/hr	0.8FP/hr	0.28FP/hr	99.7	98.2	98.2	99.5	96.8	96.9
Latency (s)	1	0.76	1.1	<0.1	-	0.71	<0.3	2.6	1.6	0.74	<1

^*Estimated based on SoC area and # of recording channels

^†Including data storage

^{^}Digital back-end only

^‡NeuralTree (2.93kB), FIR (0.2kB), and SPI (0.04kB)

^§Estimated based on AFE and DBE power consumption