IT Home On-site news, October 29: XHEART Xinxin Hangtu today officially released its large model chip X7 Ultra, with "large models defining the chip" as its core concept, designing the chip architecture based on the needs of real mass-production models,large model computing power reaching 700 TOPS。
Li Zonglin, CEO of XHEART Xinxin Hangtu, believes thatas intelligent driving moves from multiple small models to end-to-end large models, the way chips are designed also needs to change accordingly. Therefore, when XHEART was founded in 2023, it decided to start from the model requirements of real applications, work backward to the chip architecture, reorganize computing, storage, and data flow, and reserve space for future model evolution.
According to reports, the XHEART Xinxin Hangtu X7 Ultra large model chip has 700 TOPS of large model computing power,with measured 99.99% alignment on mass-production models。
This chip is designed specifically for large models, featuring an extra-large-core NPU architecture; compared with inter-core data interaction, intra-core data interaction delivers aninterconnect speed improvement on the order of TB/s. In addition,the chip's multi-level cache architecture supports large capacity and high concurrency, which can reduce inference latency and power consumption。
The XHEART Xinxin Hangtu X7 Ultra has a built-in high-precision operator acceleration engine dedicated to large models, applying extreme hardware optimization to a series of key Transformer large model operators such as Layernorm, Softmax, and Transpose,natively supporting FP16/FP32 high-precision vector computation, to guarantee model accuracy.
The chip's unlocked NPU hardware parallelism can fully unleash the potential of large model inference,achieving a computing utilization rate of 98.7%. Compared with industry flagship products, its memory bandwidth consumption drops by 80%,and energy efficiency improves by 100%。
The XHEART Xinxin Hangtu X7 Ultra also features China's first mass-produced Armv9 automotive-grade CPU,with performance improved by 20% and energy efficiency improved by 10%. In addition, the large cluster architecture improves multi-core collaboration and L3 memory access performance,integrating a brand-new SVE2 vector processing engine, offering stronger DSP performance than the previous generation's NEON.
IT Home presents a one-image overview of the core highlights of the XHEART autonomous driving large model chip X7 Ultra below: