GTI, short for Gyrfalcon Technology Inc., is a team comprised of innovative-minded AI scientists from Silicon Valley, along with seasoned engineering experts and entrepreneurs from the semiconductor chip industry.

GTI's AI processor chip, the "Lightspeeur(R) 2801S Light Spear," successfully completed tape-out in September 2017. Boasting an exceptional energy efficiency ratio of up to 9.3 TOPS/Watt—several orders of magnitude higher than other solutions currently available in the AI edge computing and data center machine learning markets—the chip features two key characteristics: ultra-low power consumption combined with minimal memory usage, alongside high levels of parallelism and in-situ computing. These attributes make it a powerful enabler for scaling diverse AI applications across both edge devices and cloud environments.

Light Spear (R) Built on the APiM architecture, which leverages memory as the AI processing unit, this design eliminates the extensive data movement found in other architectures, significantly reducing power consumption. The architecture supports true on-chip parallelism and in-situ computing, effectively overcoming the performance bottleneck caused by memory bandwidth limitations. It features 28,000 parallel computing cores and does not require external storage units for AI inference tasks.

Light Spear (R) supports Convolutional Neural Networks with a distributed memory block architecture that is highly optimized for CNN computations and accommodates multi-layer structures, where each layer can have varying dimensions. Additionally, its built-in model compression algorithms enable fast, low-power CNN processing. Furthermore, Lightspeeur(R) seamlessly integrates with popular open-source deep learning frameworks such as Caffe and TensorFlow.

Light Spear (R) 2801S is manufactured using a 28nm process and has already entered mass production. Measured performance results are as follows:

1. Efficiency power ratio is 9.3 TOPS/W

2. The single-chip boasts a peak computing performance of 5.6 TOPS, enabling VGG networks to achieve up to 130 FPS.

3. Running the VGG network achieves up to 30 FPS at a power consumption of 180 milliwatts, while running the AlexNet network delivers up to 48 FPS.

4. ImageNet image classification network accuracy: VGG achieved 66%, while AlexNet scored 58%.

This chip requires no external components. DDR, featuring multi-chip seamless connectivity, supports AI applications with more demanding computational requirements—such as AI servers, autonomous driving, smart security, and industrial applications operating under complex conditions—in a variety of scenarios.

Lightspeeur™ The Guangmao series is the world's first dedicated neural network processor chip solution capable of simultaneously supporting image and video processing, as well as voice and natural language processing.

Lightspeeur™ The chip delivers outstanding performance in terms of its exceptional energy efficiency ratio, Artificial Intelligence Compared to other solutions in the market, edge computing stands several orders of magnitude ahead in the field of data center machine learning.  

Whether in training mode or inference mode, Lightspeeur™ These chips deliver ultra-high-density computing performance, successfully overcoming the performance bottleneck caused by memory bandwidth and enabling true on-chip parallelism and in-situ computation.

Lightspeeur™ Chip support CNN ， RNN And LSTM Network models that also support popular open-source frameworks, such as Caffe, TensorFlow And MXNet 。 Software Development The package offers a one-stop development kit. 。

Contact: Mr. Song Contact number: 13509685126, email: sales 1@silone.cn