NPU Simulator for GAI LLM DLA (perf/latency) + Genio 420 workflow confirmation

NeuroPilot NeuroPilot - Analytical AI
, ,
1

Hi,

I’ve converted our own Llama-3.2-1B model to DLA using the GAI Deployment Toolkit (NeuroPilot 8, targeting mdla5.3 / MT8371). Two topics I’d like to confirm:

1. NPU Simulator — pre-hardware performance estimation

I came across the MediaTek NPU Simulator (ai-neuropilot.mediatek.com / NeuroPilot) and would like to evaluate on-NPU performance before our Genio 520-EVK arrives.

  • Can the NPU Simulator run/estimate a compiled .dla file on simulated MDLA compute to give performance and latency figures without physical hardware?
  • Does it support GAI/LLM DLAs specifically (the prompt + generative DLA pair from the GAI Deployment Toolkit), or is it intended only for vision/analytical models?
  • If supported, what metrics does it provide (e.g. estimated tokens/sec for prompt + generative, latency), and how close are these estimates to real on-device numbers?
  • What access is needed to use the simulation feature for the Genio 520/420 (MT8371) — is it included with our existing NeuroPilot/GAI access, or a separate request? Any setup guide for loading a GAI DLA?

2. Genio 420 (production target) workflow confirmation

Our production target is the Genio 420 (MT8371LV); we’re using the 520-EVK for development. Given they’re pin-compatible:

  • Do the same GAI Toolkit workflow, NeuroPilot 8, and Yocto v26.0+ requirements apply identically to the Genio 420?
  • Are our DLAs compiled for the 520 (--arch mdla5.3) directly usable on the 420, or does the 420 need a separate compile/target?
  • Any 420-specific differences we should be aware of for LLM-on-NPU deployment?

For context: our conversion is a consistent NP8 stack (NeuroPilot 8.0.11, Neuron 8.2.31, ncc-tflite 8.2.31, mtk_llm_sdk 2.5.3), both DLAs compiled with --arch mdla5.3 --num-mdla 1.

Thanks in advance!

Best regards,
Chris

3

Hi Chris,

Thanks for reaching out!

Here are the updates regarding your questions:

1. NPU Simulator for GAI / LLM DLA
The MediaTek NPU Simulator currently supports analytical models. We’re actively checking internally on the support status for GAI/LLM models (prompt + generative DLA), along with the associated metrics and access requirements. We’ll keep you posted once confirmed.

2. Genio 420 workflow and DLA compatibility

  • Workflow: The software development workflow (GAI Deployment Toolkit, NeuroPilot 8, and Yocto v26.0+) is identical across Genio 420, 520, and 720. There are no software-level deployment differences; any differences are purely hardware-specific.
  • DLA compatibility: Yes, the DLA compiled for the Genio 520 is directly usable on the Genio 420 without re-targeting. To ensure proper execution, please compile using the following configuration:
BACKEND="mdla5.3,edma3.6"
L1_SIZE_KB="256"
NUM_MDLA="1"

(Note: These settings apply identically to Genio 420 / 520 / 720.)

Best,
Jun