How Is AI Performance (TOPS) Calculated on Genio Platform?

NeuroPilot
,
1

How does MediaTek determine the AI performance rating (in TOPS) for the Genio platform APU, as shown in datasheets (e.g., Genio 700 NPU: 4 TOPS; Genio 1200 NPU: 4.8 TOPS)? What calculation method, precision format (such as INT8 , FP32 , or fixed point), and operational metrics (like MACS or FLOPS ) are used to obtain these numbers? Please clarify the relation between these metrics and how the performance figures are derived.

2

Genio calculates NPU AI performance in TOPS (Tera Operations Per Second), a metric reflecting the processor’s peak compute capability. The calculation uses the following method:

  • The performance rating is based on INT8 precision, which is the standard for Genio NPUs.

  • Core capability is measured by the number of MACS (Multiply–Accumulate operations per second). Each MAC is equivalent to 2 operations (1 multiply + 1 accumulate).

  • Formula:
    TOPS = Number of MACs × 2 × Operating Frequency (in Hz) ÷ 1e12

    • For example: If an MDLA has 1024 MAC units and operates at 500 MHz:
      TOPS = 1024 × 2 × 500,000,000 ÷ 1e12 ≈ 1 TOPS

  • Actual performance figures (such as 4 TOPS or 4.8 TOPS) are summarized based on the total number of MDLA cores, configured frequency, and INT8 operation mode.

  • Practical interpretation: TOPS depends on specific hardware implementation and will vary with frequency or voltage adjustments.

Conversion reference:

  • TOPS: Tera-Operations per Second (1 TOPS = 1e12 Ops/sec)
  • FLOPS: Floating Point Operations per Second
  • MACS: Multiply–Accumulate Operations per Second (1 MAC = 2 Ops)

Here recommends using the provided datasheet values for typical platform benchmarking and product selection.