Special | Report
Feature-Rich IP Core Solutions Target Mainstream Market
In an aim to bring premium applications to mainstream devices, ARM has released its ARM AI Platform, which is comprised of a wide range of scalable and affordable IP core solutions.

RM Holdings has lined up an army of new processor core IPs that are affordable, scalable and feature-rich enough to bring down premium and intuitive user experiences into everyday mainstream devices, like digital TVs, PCs, and mobile phones.

With the latest line of products, consumers can toy with their mid-tier mobile devices to get a game console-like user experiences, like virtual reality (VR) and augmented reality (AR), and jump on the newly emerging and hotly coveted artificial intelligence (AI) to interface with their machines through voice and image.

At a press conference in Seoul, Stephen Barton, Product Manager of Client Line Business with ARM Holdings, said, “As entertainments point at the direction, people like developers and end customers want to bring hi-fidelity games down from game consoles into their mainstream devices. Second is that machine learning is needed to be everywhere. You need to do voice recognition on your digital (TV) and face recognition on your mobile devices. And, you also want to bring down AR, VR, and full stream of videos on demand from premium devices into your mainstream devices.”

Stephen Barton
Stephen Barton, Product Manager of Client Line Business with ARM Holdings
Stephen Barton, Product Manager of Client Line Business with ARM Holdings
Scalability of Catered Devices
To keep with shifts in consumers’ appetites toward new premium features, which previously were never available on their mid-tier and entry-level devices, ARM’s new sets of IP solutions are designed to scale up from entry to middle-tiers, depending on price points and key features.

In particular, ARM’s Ethos N-57 and N-37 neural processing units (NPUs) are designed to perform machine learning, or AI inferences, to enable users to interact with their machines, like mobile phones and digital TVs, through face and speech recognition.

ARM’s Mali G-57 graphics processing unit (GPU) is another piece of the IP sets, and is the first Valhall architecture-based graphics processor. The Mali G-57 GPU is designed to power high-resolution user interface in the 4K and 8K digital TV as well as AR and high-fidelity gaming. The last piece of the new ARM IP rollout is the Mali D-37 display processor unit (DPU) for full-HD and 2K resolution, which is built on the smallest silicon footprint ever, according to ARM.

Tucked in the ARM AI platform, all these new silicon IPs will pair together with ARM Cortex A77 and Cortex A55 mobile CPUs to power a wide range of mainstream and ultra-efficient devices.

For example, mobile SoC chip makers can put together Cortex A77 and A55 with Ethos N-57, Mali GPU G-57, and Mali D-37 in a single silicon platform to power mainstream devices. They can also piece together Cortex A-55, Mali G-57, Ethos N-37, and Mali D-37 to power ultra-efficient devices.

According to ARM, the Mali G-57 performs 30 percent better than its predecessor G-52, consumes 30 percent less power, and performs machine learning workloads up to 60 percent, as the Valhall architecture boasts of great improvements in computing power, gaming performances, and data storage capacity.

For example, the Mali Valhall architecture has a wide execution engine of 16 warps and 32 lanes, boasting of one engine per shader core. That translates to 33 percent more computing power in the same area when compared with its predecessor G52. The Valhall architecture also has the capability to handle 4 texels per cycle, as it is linked to a quad texture mapper. It has a highly efficient cache memory system that can handle twice amounts of data in half the time.

“The Valhall architecture is the culmination in the evolution of execution engines from four engines to 16 engines. So, it can perform more computing power without increasing area density. It has a simplified and user–friendly complier, removing the restriction in using instruction sets. That allows for better scheduling and dynamic scheduling, as well. This is the reason why you can have better performances,” said Barton.

ARM’s Mali D-37 is the first Komeda architecture-based DPU for mainstream devices, delivering 2K and full-HD images on the smallest area of less than 1sq.mm with 16nm circuitry. This DPU is energy efficient enough to save 30 percent of power.

ARM has a very scalable Ethos NPU product lineup that can address a wide range of price points from premium market segment down to entry level. While ARM N-77 is a premium NPU for mission-critical AI applications, ARM N-57 and N-37 NPU are middle-and entry-level NPUs that are designed to play machine learning inferences on mainstream and cost-sensitive devices, respectively.

Especially, ARM N-37 is designed to be used in such applications where low memory bandwidth can be a problem.

“The Ethos NPU family is mainly designed to support the widest variety of requirements, particularly mobile applications. They are general-purpose NPUs that can run on any deep learning neural network. So they can support video, audio, speech use case applications,” Jem Davies, General Manager with ARM Machine Learning Business Unit.

“Digital TVs might include super-resolution cameras, which feature object detection, image classification, and so on. At the low-end, we often see single-use cases, such as surveillance camera just doing one thing like object detection and very restricted resources. Typically, they have very limited DRAM bandwidth, or very low-cost DRAM interfaces.”

The Ethos N-57 is designed to support mainstream smart phones and smart homes where multiple use cases are running in parallel and more computation power is in demand. It ensures that the applications areas must have balanced performances, DRAM bandwidth and power consumption, and silicon area.

Jem Davies
Jem Davies, General Manager with ARM Machine Learning Business Unit
Jem Davies, General Manager with ARM Machine Learning Business Unit
Boasting of 4 teraops per second performance and 5 teraop per watt power efficiency, N-77 can deliver greatest power and area efficiency. Meanwhile, N-57 can deliver 2 teraops per second with a 512KB internal SRAM.

“Many of our silicon partners want to buy IPs to scale across a number of different performance points and a number of different market segments. So our scalable IPs will be extremely attractive,” stressed Davies.