Hi NXP Team,
We are evaluating the i.MX 95 for an edge-AI application and would like to clarify the actual NPU performance and benchmarking information.
According to the latest i.MX 95 Industrial Data Sheet the Neural Processing Unit (NPU) is specified as: 8 eTOPS at 1 GHz, NPU can operate at 800 MHz for reduced power consumption,1 MB embedded SRAM within the NPU.
However, the i.MX 95 Reference Manual appear to indicate an NPU performance of up to 2 TOPS.
Could NXP please clarify the following:
1) What is the reason for the difference between the 8 eTOPS specified in the latest data sheet and the 2 TOPS mentioned in the Reference Manual documentation? (is it per core performance since it has 4 neutron )
https://www.nxp.com/docs/en/data-sheet/IMX95XEC.pdf
https://www.nxp.com/docs/en/reference-manual/IMX95RM.pdf
2)Does the 8 eTOPS figure represent the theoretical peak performance of the NPU at 1 GHz?
3)What exactly does "eTOPS" represent in this specification?
4) Is the 8 eTOPS capability available on all i.MX 95 variants, or does it depend on the specific SoC ordering/part number?
5) Is there any difference in NPU configuration between the i.MX 95 variants that could explain the 8 TOPS vs 2 TOPS numbers?
6) Could NXP provide real-world NPU benchmark results rather than only the theoretical TOPS/eTOPS number?
7) Could you also provide a benchmark comparison between the i.MX 95 NPU and the i.MX 8M Plus NPU using the same model, input resolution and precision?
For an application where the input is coming from a camera, could NXP provide an end-to-end example showing:
Camera → ISP → preprocessing → NPU inference → post-processing
and the achievable FPS/latency for a representative object-detection model?
9) Are there any publicly available tools or benchmark suites that NXP recommends for customers to measure NPU performance on the i.MX 95 EVK?
Thanks