Huawei Atlas 300I A2 — Ascend 910B4 64GB HBM Inference Card

Specifications

Overview

The Huawei Atlas 300I A2 upgrades the Atlas inference card line to the Ascend 910 series silicon. Based on the Ascend 910B4 processor, it integrates AI cores, a general-purpose processor and codec hardware to serve inference, video analysis and feature-retrieval workloads from a single card.

It is offered with 24GB, 32GB or 64GB of HBM2e memory and publishes 560 TOPS INT8 / 280 TFLOPS FP16, making it far better suited than LPDDR4X-based cards to large language models and long-context inference. Feature retrieval, clustering, OCR, speech analysis and video structuring are first-class use cases, and secure boot is supported.

As with the rest of the Atlas family, the 300I A2 runs on Huawei's CANN stack with MindSpore and the PyTorch Ascend adapter. QS Compute supplies the Atlas 300I A2 for LLM inference nodes, RAG and vector search appliances and intelligent video platforms.

Key Benefits

560 TOPS INT8 / 280 TFLOPS FP16 from a single Ascend 910B4. Up to 64GB HBM2e for large-model inference. Feature retrieval and secure boot built in. CANN + MindSpore software ecosystem.

Applications

Large language model inference, retrieval-augmented generation and vector search, video structuring and analytics, OCR and speech analysis, and datacenter AI inference.

Related Products