Four reasons why they are new and innovative

The term “fast mover” is typically used to describe a runner who is sprinting at a high speed around a track. They possess the determination to remain focused on a single path until they achieve their goal. Here in Asia, startups have been digging and digging to create different AI chips. These fascinating AI chips are changing the world little by little. These four startups prove it.

LLM-focused chips

Have you ever wished ChatGPT could answer your queries faster? The secret lies in the AI chip that powers it. HyperAccel, a semiconductor startup based in South Korea, has recognized that the bandwidth of the chip is critical for the performance of AI models. So, they created a chip that is optimized for high bandwidth, enabling it to fetch a lot of information quickly for learning and computation. This chip maximizes memory bandwidth, which is critical for very large generative AI computations, and incorporates logic that is specialized for LLM inference.

HyperAccel is developing a new semiconductor processor called the “Latency Processing Unit” (LPU). LPUs are equipped with a computer operation engine that optimizes memory bandwidth and performs required computations for inference at high speed. The company has developed model parallelization technology to efficiently distribute LPUs across multiple devices and its own networking technology for data synchronization between LPUs. Recently, HyperAccel launched Orion, an AI server accelerator with its own LPUs. Orion Server is a new generative AI server based on LLM using FPGA technology. With LPUs and Orion, HyperAccel claims that it is about 2.4x more cost-effective than GPUs. They tested Meta’s generative AI, Lama 3, on their platform, and the results showed that the generative AI’s answer rate was 30-50% faster than the existing service. Performance is significantly improved, especially for intricate reasoning processes like coding. Whenever you ask a question, the super-large language model will respond promptly, taking less than a second to provide you with an answer in near real-time.

Domain-specific processor

This is related to what we call “narrow AI”. It refers to the application of AI to a specific task, such as autonomous driving or facial recognition. Domain-specific architectures are processors designed for specific purposes that are trained on data relevant to a particular domain. For the past 70 years, computer processors have been based on the Von Neumann Architecture, which is rule-based. CPUs, GPUs, DSPs, and even CPU bus-connected accelerators are all variants of the Von Neumann Architecture. However, Hailo, an Israeli startup, has developed a new architecture using a unique approach that does not depend on traditional CPU or GPU building blocks.

Hailo has recently unveiled Hailo-8, a specialized deep learning processor that accelerates inference on edge devices. This new architecture boasts a domain-specific processor that offers higher performance, lower power consumption, and minimal latency. It delivers up to 26 teraflops per second (TOPS), enabling the execution of advanced deep learning applications that were only possible in the cloud before. Hailo provides AI processors to various manufacturers across industries, including robotics, advanced driver assistance systems (ADAS), smart cities, and smart homes. As edge technologies continue to be implemented in automotive, smart retail, smart cities, and more, Hailo is well-positioned to serve in this space.

Sound-tracing chips

Have you ever noticed that when you wear Apple’s VisionPro, the sound in your head seems to echo loudly? Well, the reason behind this could be the chip used in the device. A South Korea’s startup called EXARION has developed semiconductors specifically designed for 3D audio and video AI systems. They have focused on creating sound-tracing chips, which work similar to the ray tracing* technology used by GPUs to generate photorealistic images. Sound-tracing, on the other hand, tracks sound paths in real-time, including reflections, diffractions, and other factors that make sound realistic. This technology enables sounds to be clearly distinguished between left and right, forward and backward, up and down, and even changes depending on the terrain and reflective materials in the virtual space.

* Ray tracing: It is a technique used to generate pixel colors and forms by tracing the path of light from a theoretical light source as it reflects off a surface. This approach uses an algorithm that monitors rays to produce a graphical display that is highly realistic.

EXARION has developed a soundtracing chip that offers ultra-realistic 3D audio technology, which is perfect for applications such as metaverse, XR, MR, and gaming. The company is currently working on a prototype of the chip, which will be manufactured on TSMC’s 12-nanometer process. Despite being only 0.3 centimeters wide and deep, the chip can process vast amounts of sound data at low power consumption. EXARION has already validated the soundtracing design with the FPGA, and the chip can be used in various consumer electronics, including smartphones, TVs, headsets, and laptops, as well as augmented reality devices such as Apple’s VisionPro. Welcome!

AIoT chips

What if all the devices and objects we use in our homes, workplaces, and society were equipped with artificial intelligence and the Internet of Things? Axera Tech, a China’s startup, is working towards a future where all devices and objects we use in our daily lives are equipped with artificial intelligence and the Internet of Things. They have developed a system on a chip (SoC) technology that combines high definition, high energy efficiency, quality AI experience, and low power consumption. This technology is applied in a wide range of applications including AI cameras, smart motion cameras, smart transportation systems, solar cell systems, intelligent automotive vision systems, unmanned aerial vehicles, smart manufacturing, smart homes such as cleaning robots, and wearable devices.

Axera Tech is working on expanding the use of AI technology from the cloud to edge devices. Recently, it has developed a range of smart driving chips based on its own core technology. These chips have been integrated into two car models that have already entered large-scale production. As a result, we can now enjoy the convenience of having AI drivers, super-fast navigation systems, and trusted valet-parking assistants in our cars. This technology has made finding directions or parking a lot easier for us.