The MT04 AI accelerator delivers exceptional compute acceleration with up to 32x sparsity support, combining ultra-high performance, low power consumption, and industry-leading energy efficiency. Optimized for computer vision, natural language processing, multimodal inference, and other data center AI workloads, it serves large-scale inference applications across internet, telecommunications, smart cities, life sciences, and autonomous driving. Its proprietary dual-sparsity algorithm and hardware-software co-design achieve orders of magnitude performance gains while significantly reducing total cost of ownership (TCO).

1、Deeply Integrated Architecture

Leverages hardware-software co-design to tightly couple compute cores with high-bandwidth on-chip memory, 4th-generation tensor cores optimized for sparse computations, and a unified memory architecture with zero-copy data access.

2、Intelligent Sparsity Engine

Purpose-built for AI inference with high-rate sparse tensor cores supporting up to 32x model sparsity, adaptive precision encoding for lossless model compression, and a dynamic pruning engine achieving near-ideal hardware utilization.

3、Flexible Scalable Design

Integrates specialized accelerators for diverse workloads including custom sparse processing units, an 8K60 video codec engine supporting AV1/HEVC/VP9, 200GB/s JPEG decoding throughput with zero CPU overhead, and a vector search engine optimized for 1B+ embeddings.

4、Professional Multimedia Platform

Dedicated hardware acceleration for intelligent video analytics supporting simultaneous processing of 32x 4K30 video streams per card, real-time object detection with sub-2ms latency, and CPU offload for JPEG/PNG decoding up to 1000fps.