NXP samples first i.MX8 with an NPU
Jan 7, 2020 — by Eric Brown 3,222 viewsNXP unveiled a 1.8GHz, quad -A53 “i.MX8M Plus” SoC with a 3D GPU, a Cortex-M7 MCU, a HiFi4 DSP, dual ISPs, and a 2.3 TOPS NPU. A Linux BSP and EVK are also in the works.
NXP has begun sampling its first processor with an AI chip. Applications for the i.MX8 Plus include “people and object recognition for public safety, industrial machine vision, robotics, hand gesture, and emotion detection with natural language processing.”
The i.MX8M Plus is like a faster version of the i.MX8M Nano, but with a 2.3-TOPs Neural Processing Unit (NPU) and more coprocessors. The NPU can process AI algorithms on MobileNet v1 at over 500 images per second, claims NXP.


NXP i.MX8M Plus block diagram (left) and EVK with camera option
(click images to enlarge)
The i.MX8M Plus joins many other high-end SoCs with edge AI chips designed for local edge inference processing of sensor data to reduce the latency and cost of sending data to the cloud. Many of the newcomers to our recent New Years catalog of 136 open-spec Linux SBCs are built on NPU-enabled SoCs such as the Rockchip RK3399Pro or Amlogic AM311D, or use separate NPUs such as Google’s Edge TPU or Intel’s Movidus Myriad X.
— ADVERTISEMENT —
Like the i.MX8M Nano, the i.MX8M Plus has 4x Cortex-A53 cores fabricated with a 14nm LPC FinFET process. The i.MX8M Plus is faster than the Nano, with -A53 clock rates listed on the product page at up to 1.8GHz (or 2GHz according to the announcement), compared to 1.5GHz on the Nano.
Like the Nano, the Plus provides GC7000UL 3D and GC520L 2D GPUs with OpenGL ES 2.0/3.0/3.1, Vulkan, and OpenCL 1.2 support. Unlike the Nano, there’s a VPU with H.265 or H.264, but like the similarly 14nm, quad -A53 i.MX8M Mini, it’s limited to HD instead of 4K on the i.MX8M and up to hexa-core i.MX8 Quad/QuadPlus/QuadMax.
Like the Nano, there’s a Cortex-M7 real-time chip as opposed to the lower-power Cortex-M4 cores on the other i.MX8 SoCs. Cortex-M7 speed is increased to 800MHz compared to 650MHz for the Nano.
Like the i.MX8M and i.MX8 Quad, the i.MX8M Plus adds a Tensilica HiFi4 DSP clocked at 800MHz for natural language processing. On most models, there are also dual ISPs for stereo vision or a single 12-megapixel camera. The ISPs have an input rate of up to 375 MP/s and support HDR and fisheye lens correction.
The mix of coprocessors is much like that of TI’s AI-enhanced Sitara AM5729, which powers the BeagleBone AI. However, the AM5729 is limited to dual Cortex-A15 CPU cores.


i.MX8M Plus comparison with other i.MX8 SoCs (left) and different Plus models
(click images to enlarge)
The i.MX8M Plus is the first i.MX8 SoC to support ECC RAM, PCIe Gen3, LVDS, or dual CAN FD. Unlike its siblings, it supports dual Gigabit Ethernet ports one of which supports Time-Sensitive Networking (TSN), a feature found on NXP’s QorIQ LS1028A. Unlike the Nano and Mini there’s support for HDMI 2.0a and USB 3.0, including Type-C.
The 15 x 15mm Plus shares the extensive digital audio features found on the other i.MX8 SoCs. There’s a security block and up to 15-year lifecycle support. Some models will support up -40 to 105°C temperatures at 1.6GHz.
An i.MX8M Plus EVK is in the works that will ship with a Linux BSP. An image was supplied as part of a preliminary product page for a 4K MIPI camera option to the EVK (see farther above). This is an 8-megapixel MIPI-CSI2 camera optimized for the Plus’ ISPs and named for its 4-Lane D-Phy output. No other details on the DVK appear to be available.
Further information
NXP has begun sampling the i.MX8M Plus for customers. More information may be found in the announcement and product page.
Please comment here...