The characteristics of Jiushao NPU include high computing power, high energy efficiency and high bandwidth, which are the basis for the higherlevel iteration of intelligent driving technology. It supports mixed precision including INTFPFP, integrates hard acceleration for highprecision fine quantization and Transformer, and can simplify developers' work during quantization and deployment.
In addition, Jiushao NPU also has a threelayer memory architecture with low latency and high throughput, including a largecapacity and highbandwidth NPU dedicated cache, a core module onchip shared cache, as well asn afghanistan phone number list symmetrical dual data paths and a dedicated DMA engine. It improves performance and effective bandwidth, reduces dependence on external storage bandwidth, and achieves an ultimate balance between performance, bandwidth, and cost.
Heizhima Intelligence has developed a new generation of general AI tool chain BaRT. BaRT supports conversion of multiple popular frameworks and models, is natively compatible with PyTorch's reasoning API, and supports Python programming deployment. This enables developers to more conveniently use the Jiushao architecture to develop and deploy AI models.
Another advantage of BaRT is that it supports the industry's mainstream Triton custom operator programming, allowing developers to use Python to write Triton custom operators. These operators can be automatically compiled into hardware acceleration code, thereby further accelerating the deployment of developers' AI models.