How many gigaflops GPU MALI T624 MP6 reaches?

Hi kevin_23,

The architectural graphics arithmetic performance is 17 fp32 flops per clock, per arithmetic pipeline, per core. Mali-T620 series has 2 arithemetic pipelines, and based on other Mali devices I would expect the clock frequency to be somewhere around 450Mhz (not sure on the exact performance of the part in the Note 3 Neo - you may get a better answer on the Samsung forums). So with 6 cores I would expect ...

17 x 2 x 6 x 450MHz = 91.8 GFLOPS.

However, two important things to note.

1. Most graphics content heavily uses fp16 rather than fp32 - for Mali this means we can get (approximately) double the performance in terms of peak FP16 flops throughput.
2. More importantly, peak FLOPS numbers are a terrible way to measure graphics performance. Very little content is limited by maths processing capability - memory accesses for per-vertex data, texturing are normally the limiting factor for 99% of content on the market - either in the GPU or downstream of the GPU in the memory system. In our experience to date very little gaming content other than artificial benchmarks even gets remotely close to using the available "flops" before they start slowing down for other reasons.

HTH,
Pete

Hi kevin_23,

The architectural graphics arithmetic performance is 17 fp32 flops per clock, per arithmetic pipeline, per core. Mali-T620 series has 2 arithemetic pipelines, and based on other Mali devices I would expect the clock frequency to be somewhere around 450Mhz (not sure on the exact performance of the part in the Note 3 Neo - you may get a better answer on the Samsung forums). So with 6 cores I would expect ...

17 x 2 x 6 x 450MHz = 91.8 GFLOPS.

However, two important things to note.

1. Most graphics content heavily uses fp16 rather than fp32 - for Mali this means we can get (approximately) double the performance in terms of peak FP16 flops throughput.
2. More importantly, peak FLOPS numbers are a terrible way to measure graphics performance. Very little content is limited by maths processing capability - memory accesses for per-vertex data, texturing are normally the limiting factor for 99% of content on the market - either in the GPU or downstream of the GPU in the memory system. In our experience to date very little gaming content other than artificial benchmarks even gets remotely close to using the available "flops" before they start slowing down for other reasons.

HTH,
Pete