DSP, Micros & Memory


Scaling ML processing with external AI accelerators

23 November 2022 DSP, Micros & Memory AI & ML

At NXP, we craft our i.MX applications processors and our i.MXX RT crossover microcontroller (MCU) portfolios and families of devices to cover a very wide spectrum of market needs. From verticals needing very low-power dissipation to other end products needing a complex heterogeneous compute platform with multiple CPUs, 2D-3D GPUs, DSPs and NPU machine learning accelerators, our product teams must solve a multi-dimensional optimisation problem. The ultimate goal is to deliver all the compute power and connectivity your application and products need, with minimal unused features so that the area footprint, power dissipation, and cost requirements of the devices are also met.

When talking about machine learning, there are some end uses such as autonomous vehicles and natural language processing that push the compute requirements for edge devices to limits of tens and hundreds of tera operations per second (TOPs). With more emphasis on developing efficient ML models specifically for the edge, and using techniques such as quantisation and pruning, many of the edge machine learning applications fit in the

Giga-Ops to low single digit TOPs range of ML compute performance. NXP solutions natively cover this large portion of AI processing needs.

The software investment dominates the hardware choice these days, especially when considering roadmaps and multiple generation of products. There is a clear benefit to choosing the same or similar applications processor when developing a range of end products, so that the development effort translates to higher maturity and quality, with reuse. Different market segments require different performance on some portion of the application, but the fundamentals remain unchanged between versions and tiers of products. Choosing a scalable processor family such as i.MX applications processors gives the developer flexibility in advanced features and performance range, while simultaneously providing a common base compute architecture and feature set to leverage across both the portfolio and common software enablement elements. NXP provides GStreamer and NNStreamer frameworks to simplify deployment of vision applications with ML.X GStreamer is used as a framework for creating streaming media applications, abstracting the hardware layer to allow the use of any i.MX SoC, without having to change the underlying vision pipeline software.

Application requirements and markets evolve even after a product is launched in the market. So, what does one do when more is still needed from the selected applications processor? Going back to the selection process and looking for a higher performance processor is usually not a preferred option. Adding another device to provide additional acceleration when needed is a possible path, especially with high-speed high bandwidth, low latency chip-to-chip connectivity options such as PCIe. This is where NXP’s ecosystem partners with dedicated ML accelerator chips can help.

Kinara is such an NXP ecosystem partner that develops the Ara-1 Edge AI processors for dedicated ML acceleration. The Gstreamer- and NNStreamer-based vision pipeline support offered by NXP, and the set of Kinara-developed Gstreamer compatible plugins, make it seamless to integrate Ara-1X into NXP inference pipelines and to also easily migrate the design to different i.MXX applications processors if the feature requirements change.

Combining the native MLX processing capabilities of NXP’sX i.MXX applications processors with dedicated MLX accelerators from NXP ecosystem partners such as Kinara, creates an even larger scalability range than that offered by NXP alone while still maintaining software reuse.


Credit(s)



Share this article:
Share via emailShare via LinkedInPrint this page

Further reading:

Ryzen-based computer on module
Altron Arrow AI & ML
SolidRun announced the launch of its new Ryzen V3000 CX7 Com module, configurable with the eight-core/16-thread Ryzen Embedded V3C48 processor.

Read more...
QuecPython live demonstration
Quectel Wireless Solutions DSP, Micros & Memory
QuecPython allows designers to adapt Quectel’s modules quickly, with a low-code approach to suit their precise requirements in less time and at reduced cost, while maintaining high security standards.

Read more...
Robust and customisable SBC
Altron Arrow DSP, Micros & Memory
Pairing the powerful i.MX8M Plus System on Module (SoM) from SolidRun, which features the i.MX 8M Plus SoC from NXP, this high-performance SBC is set to transform industrial environments.

Read more...
New family supports future cryptography
Altron Arrow DSP, Micros & Memory
NXP has introduced its new i.MX 94 family, which contains an i.MX MPU with an integrated time-sensitive networking (TSN) switch, enabling configurable, secure communications with rich protocol support in industrial and automotive environments.

Read more...
Fast and reliable 4G connectivity worldwide
TRX Electronics DSP, Micros & Memory
Powered by a powerful Quectel LTE Cat 4 modem, the Arduino Pro 4G module’s fast data throughput and high bandwidths ensure reliable and quick data download and upload, even in remote locations.

Read more...
NXP’s all-purpose microcontroller series
Altron Arrow DSP, Micros & Memory
NXP has released its MCX A14x and A15x series of all-purpose microcontrollers which are part of the larger MCX portfolio that shares a common Arm Cortex-M33 core platform.

Read more...
ESP32-P4 SoC
iCorp Technologies DSP, Micros & Memory
Espressif Systems announced its latest SoC, the ESP32-P4 which is powered by a RISC-V CPU, with an AI instructions extension, an advanced memory subsystem, and integrated high-speed peripherals.

Read more...
Non-terrestrial network module
Altron Arrow Telecoms, Datacoms, Wireless, IoT
Fibocom unveiled its MA510-GL (NTN), a non-terrestrial networks module which is compliant with 3GPP Release 17 standard.

Read more...
Microchip SoC FPGA
ASIC Design Services DSP, Micros & Memory
Microchip Technology introduced the RT PolarFire SoC FPGA, the first real-time Linux capable, RISC-V-based microprocessor subsystem on a proven RT PolarFire FPGA platform.

Read more...
QLC Flash memory using BiCS tech
EBV Electrolink DSP, Micros & Memory
KIOXIA announced it had started shipping its 2 Tb Quad-Level-Cell memory devices with its 8th-generation BiCS FLASH 3D flash memory technology.

Read more...