CB

Cambricon 玄思1000 (Xuansi 1000)

rack
4× MLU290-M5
MLU-Link dual-ring topology across the four cards internally, with 8x 100GB/s bidirectional MLU-Link exposed externally; hosts attach over two groups of PCIe 4.0 x16 at 64 GB/s bidirectional through Mini SAS HD connectors, two cards sharing each; scale-out is 2x 200Gbps HDR InfiniBand single-port adapters
2021
Power
2.3
kW Total
Memory
131
GB Total

Cambricon's four-card accelerator appliance, and the only whole-machine product Cambricon lists. A 2U chassis 447 by 87 by 735 mm and 29 kg holds four MLU290-M5 cards with ECC memory, four 1.92 TB NVMe drives in U.2 slots, and a typical power draw of 2,300 W from two 3,000 W supplies in 1+1 redundancy. Compute is quoted as 2.05 PetaOPS INT8, 1 PetaOPS INT16 and 256 TOPS CINT32 across the chassis, with FP32 and FP16 listed among supported precisions but given no throughput figure. The INT8 total is exactly four times the 512 TOPS Cambricon publishes for a single MLU290-M5. Memory is 128 GB across the four cards; Cambricon states no per-card capacity, so none is shown on the card entry. The four cards link internally in an MLU-Link dual ring and expose MLU-Link externally, which is what lets several chassis scale up rather than only out. Cambricon notes that two of these units combined with a CPU server form an eight-card system, a configuration it publishes no separate specification for. Video decode covers H.264, HEVC, VP8 and VP9.

We will point you at suppliers who have it. Free, and no signup.

System Details

GPU Configuration

GPU Count: 4 GPUs
Architecture: MLUv02
Interconnect: MLU-Link dual-ring topology across the four cards internally, with 8x 100GB/s bidirectional MLU-Link exposed externally; hosts attach over two groups of PCIe 4.0 x16 at 64 GB/s bidirectional through Mini SAS HD connectors, two cards sharing each; scale-out is 2x 200Gbps HDR InfiniBand single-port adapters

System Specifications

Form Factor: rack
Total Power: 2.3 kW
Total Memory: 131 GB
Memory Bandwidth: 4920 GB/s

Powered by Cambricon MLU290-M5

This system utilizes 4 × Cambricon MLU290-M5 GPUs, each delivering exceptional performance for AI and HPC workloads.

Per GPU TDP

350W

Per GPU Memory

Process Node

Architecture

MLUv02

Documentation & Resources

Official Datasheet

Cambricon 玄思1000 (Xuansi 1000) technical specifications

Download PDF

Cambricon MLU370-X8 product page

Cambricon • 2022-03-21

View Document ↗

Typical Use Cases

AI/ML Training
High-Performance Computing
Data Analytics

The Cambricon 玄思1000 (Xuansi 1000) runs 4× MLU290-M5 GPUs over MLU-Link dual-ring topology across the four cards internally, with 8x 100GB/s bidirectional MLU-Link exposed externally; hosts attach over two groups of PCIe 4.0 x16 at 64 GB/s bidirectional through Mini SAS HD connectors, two cards sharing each; scale-out is 2x 200Gbps HDR InfiniBand single-port adapters.