colbert qps
0
20
40
60
80
throughput, qps
quantization config, colbert rescorer (hybrid prefetch, k=20, pf=100, concurrency=8)
50
no-quant
(f32)
31
no-quant
(f16)
53
no-quant
(turbo4)
50
scalar
int8
47
product
x16
27
binary
(f32)
45
binary
(f16)
49
turbo4
tq1