Specifications
AI Processor
2x Huawei Ascend 310P (Da Vinci architecture)
Form Factor
Single-slot full-height full-length PCIe card
Host Interface
PCIe 4.0 x16
AI Performance (card)
280 TOPS INT8 / 140 TFLOPS FP16
AI Performance (per chip)
140 TOPS INT8 / 70 TFLOPS FP16
Memory
96GB LPDDR4X with ECC (48GB per chip)
Memory Bandwidth
408 GB/s (whole card)
Video Decode
H.264/H.265 hardware decode, up to 64-ch 1080p30 per card
Video Encode
H.264/H.265 hardware encode, up to 4-ch 1080p30
Precision Support
INT8, FP16
Power Consumption
Up to 150W (dual-chip, card-level)
Cooling
Passive heatsink, requires server airflow
Software Stack
Huawei CANN, MindSpore, AscendCL, PyTorch Adapter (torch-npu), ATC
Operating System
openEuler, EulerOS, Ubuntu 20.04/22.04 LTS
Deployment
Huawei Atlas servers and partner x86/Arm servers with CANN support
Certifications
CE, FCC, RoHS
Overview
The Huawei Atlas 300I Duo is the high-memory member of Huawei's Ascend-based inference card family. It packages two Ascend 310P AI processors onto one full-height, full-length PCIe 4.0 x16 board and pairs them with 96GB of ECC LPDDR4X memory at a combined 408 GB/s, delivering 280 TOPS INT8 and 140 TFLOPS FP16 across the card.
The large memory pool is the headline: it allows quantised parameter-heavy LLMs, long-context key-value caches and multi-stream video pipelines to remain resident on-device, cutting host-to-device transfers and enabling dense inference nodes without a full datacenter GPU. Each chip is individually a 140 TOPS INT8 / 70 TFLOPS FP16 part, and the card exposes both to the Ascend software stack.
The card runs within the Huawei Ascend ecosystem (CANN, MindSpore, torch-npu and AscendCL) and integrates with Atlas inference servers and CANN-supported partner platforms. QS Compute supplies the Atlas 300I Duo for AI inference buildouts, RAG appliances and intelligent video analysis systems.
Key Benefits
96GB ECC LPDDR4X keeps large models and KV caches resident. 280 TOPS INT8 across two Ascend 310P processors. Single-slot PCIe 4.0 x16 form factor for dense multi-card nodes. Full Ascend software stack with CANN, MindSpore and torch-npu.
Applications
LLM and generative AI inference, retrieval-augmented generation (RAG) and vector search, recommendation and ranking, intelligent video analytics, OCR and speech analysis, and edge-to-cloud AI inference nodes.
Request a Quote — HUAWEI ATLAS 300I DUO — DUAL ASCEND 310P 96GB AI INFERENCE CARD
QS Compute — global B2B supply of AI computing hardware, edge AI systems and accelerators. Volume pricing, 15-day sample lead time.
Get Your Quote →