北京数渡

Member of Vantone Group (600246.SH)

English

8-Card Scale Up Solution

Solution Overview

For high-speed interconnection in 8-card AI servers, Sudotech innovatively adds PCIe Gen5 x16 Fabric Links between Switches. This eliminates GPU communication bottlenecks caused by inter-CPU communication constraints. In the dual-Switch architecture, each Switch connects upstream to a CPU, fully unleashing the computing and memory performance of dual CPUs. It avoids performance bottlenecks from cross-CPU computation and memory access, significantly improving training and inference performance.

Supports Fabric Link for PCIe 5.0 Switch cascading

Reduces performance loss in computation and memory access across NUMA domains Achieves dual improvements in cluster communication efficiency and overall computing utilization Delivers equivalent performance for intra-Switch communication, and significantly outperforms industry solutions for inter-Switch communication

Application Fields

• Edge Computing

• Education and Scientific Research

• Large Model Training

• Scientific Computing

• High‑Performance Computing (HPC)

16-Card Scale Up Solution

Solution Overview

16-Card GPU Interconnection Solution adopts 4 PCIe Switches to high-speed interconnect 16 GPUs, which greatly enhances the All-to-All communication performance between GPUs and improves the performance of MoE models. It reserves x8 network card expansion to build large-scale clusters. A single machine can provide an ultra-large video memory of 1.5+ TB, enabling fast operation of the full-size Qwen model and the lossless-precision full-performance version of the DeepSeekR1 671B model with just one device.

Application Fields

• Visual Effects

• Scientific Computing and Simulation

• High-Concurrency Inference

• Large Model Training

• High-Performance Computing (HPC)

High-Density All-Flash System Solution

Solution Overview

Extreme structural design enables PB-level capacity in a 1U chassis with high space utilization. Multi-PCIe chip interconnection supports up to 24 NVMe drives to build high-performance storage arrays. Dual-controller redundant design, supporting RAID protection, snapshot, compression, deduplication, encryption, etc. Security features including automatic alarm, automatic reconstruction, and automatic repair during routine inspections.

Application Fields

• Intelligent Computing Center Infrastructure

Supports AI training acceleration, massive data writing, and large-scale parallel computing.

• Extreme Performance Industry Acceleration

Applied to HPC clusters and cloud application acceleration, delivering ultimate performance.

DPU Storage Server Solution

Solution Overview

Equipped with BMC policy configuration, dual main controller redundancy, DPU hardware-level offloading, and high-speed PCIe Switch interconnection.

With an extreme structural design, the 1U chassis supports 24 NVMe drives with end-to-end transmission, featuring high space utilization. Supports automatic failover of faulty links and redundancy of key nodes, reducing O&M difficulty and costs. Implements storage-computing separation and flexible expansion, with file system performance linearly improving as nodes increase.

Key Indicators

Single-machine bandwidth: 288GB/s (1M read) Single-machine IOPS: 64.8 million (4k rand read) Access latency: <20us Total bandwidth: 2.4Tb (6*DPU)

Application Fields

• Chip Design

• High-Performance Computing (HPC)

• Financial High-Frequency Trading

• Life Sciences

• Internet Real-Time Recommendation

PCIe Switch System Solution

Solution Overview

Sudotech high-performance Fabric switch consists of a main control module, a switching module and a BMC module. Adopting a 1U standard rack-mounted design, it integrates one PCIe Switch and provides 16 QSFP-DD ports. It delivers native memory semantic expansion for super nodes and enables fast configuration of computing and switching networks. Supports on-demand scaling of 16/32/64-card super nodes, realizing flexible resource expansion and building a pooled system.

Sudotech PCIe switch effectively reduces end-to-end latency with lower power consumption and improved energy efficiency. The overall solution features “low latency & high bandwidth + flexible expansion + intelligent scheduling + high integration & low power consumption”.

Application Fields

• Resource pooling for cloud service providers

• Full-cycle development of trillion/hundred-trillion parameter AI large models

• High-performance computing cluster expansion

• Scientific computing & AI for Science

Edge Intelligent Computing Appliance System Solution

Solution Overview

The Edge Intelligent Computing Appliance enables rapid on-demand adaptation to edge computing power requirements and empowers efficient real-time data processing at the edge.

Featuring a professional VPX chassis design, it supports flexible and dynamic plug-in and free recombination of various blade modules. With high-performance PCIe Switch as the core interconnection, it seamlessly integrates independent nodes into an integrated system. It realizes pooling and elastic scheduling of computing and storage resources, improving utilization and scalability.

Application Fields