Hello Forum,
In some of open-source pages have seen that earlier version of G725 has 2 FMA units and now G725 supports 2 CVT units ("Doubled shift-convert unit throughput" in Mali-G725 | 5th Gen Architecture for Premium Mobile GPU – Arm®). But Arm GPU data sheet shows FP32 operations/cycle, FP16 operations/cycle is same in G715, G720 and G725. So, I am confused how to interpret the 2 FMA units as throughput is same.
My basic understanding is each shader core has 4 Arithmetic units or Processing Units (PU) since G710 and each PU has 1 FMA, 1 CVT, 1 SFU pipelines. Now G725 supports 2 CVT's, does that mean it has 1 FMA, 2 CVT, 1 SFU pipelines for each PU?
Thanks
Venkatesh
The data sheet only counts FMAs (as two ops each).
FMA and CVT share issue hardware, so you can't issue more than 2 ops per clock per PU. The extra CVT pipe buys flexibility for issuing different sequences of instructions, but not more ops/clock.
HTH,
Pete