SOIN Technology Resmi Meluncurkan SOINAWE-RVE1300, Server Inferensi AI Performa Tinggi Kelas Enterprise, dengan Mesin AWE dan Ekosistem Xinchuang sebagai Fondasi Digital AI Perusahaan

SOIN Technology (SOINTECH) has officially launched the SOINAWE-RVE1300, an enterprise-grade high-performance AI inference server co-developed with ESWIN Computing. The system integrates four EBC7702-P02 accelerator cards (eight dedicated AI processors in total), delivers up to 1520 TFLOPs of FP8 compute, and builds a 96 GB unified shared memory pool over an Ethernet Mesh interconnect, capable of running large language models with up to 80 billion parameters or several 20-billion-parameter models concurrently. Deeply integrated with SOIN's self-developed AWE (Agentic Workflow Engine), fully compatible with Xinchuang (IT innovation/localization) AI accelerator cards as well as mainstream domestic databases and middleware, and delivered as a fully private deployment where data never leaves the enterprise domain, the RVE1300 provides automotive and other security-conscious enterprises with a ready-to-use AI digital foundation.

Product Overview: 4U Rack-Mount, Built for Inference

The SOINAWE-RVE1300 adopts a 4U rack-mount design, carrying four EBC7702-P02 accelerator cards with eight dedicated AI processors in total. The Ethernet Mesh interconnect forms a unified shared memory pool, backed by dual Intel Xeon host CPUs, 512 GB DDR4-3200 ECC memory and 4 TB PCIe 4.0 NVMe storage. The flexible Ethernet grid topology scales shared memory up to 96 GB, significantly reducing quantization requirements and improving model accuracy. Supporting both single-user/single-model and multi-user/multi-model scenarios, a single unit can carry enterprise-scale LLM inference workloads.

Key Specifications

Measured LLM Inference Performance

Measured inference performance of mainstream open-source LLMs on the SOINAWE-RVE1300 (Batch=32, TP=8, all eight AI processors engaged):

The results show sustained throughput above 500 tokens/s even on 72B-class models, and over 729 t/s on 32B-class models — enough to confidently serve hundreds of concurrent users across intelligent Q&A, coding assistance and knowledge retrieval applications.

Six Product Advantages

· High compute: up to 1520 TFLOPS (FP8) of deep-learning compute on the open RISC-V instruction set architecture, freely modifiable and maintainable;

· Secure and reliable: a fully open-source software stack with no closed-source black boxes — code can be audited independently to meet security review and compliance requirements;

· Flexible scaling: standard-Ethernet-based interconnect breaks proprietary interconnect barriers, scaling from 1 to thousands of nodes to build a true Exa-scale AI supercomputer;

· Software flexibility: a complete software stack from high-level frameworks down to the metal, balancing development convenience with deep hardware optimization;

· High performance: the self-developed BLOCKFP format delivers accuracy close to traditional floating point while cutting memory bandwidth and storage requirements by 50%, greatly accelerating data processing on identical hardware;

· High cost-effectiveness: high compute density, energy efficiency and flexible horizontal scaling reduce hardware, power and full-lifecycle management costs of large-scale AI deployment.

Deep Integration with the SOIN AWE Engine: Agents Out of the Box

The SOINAWE-RVE1300 ships with SOIN's self-developed AWE (Agentic Workflow Engine) pre-integrated. AWE turns an enterprise's industry knowledge graphs, business process logic and decision patterns into reusable agent technology assets. Combined with the RVE1300's high-throughput inference, enterprises can rapidly build agent applications across R&D, manufacturing and business analysis — without building AI infrastructure from scratch — delivering an integrated "compute foundation + agent engine" stack.

Full Xinchuang Ecosystem Compatibility

For enterprises operating in Xinchuang (IT application innovation) environments, the SOINAWE-RVE1300 supports flexible configuration of Xinchuang AI accelerator cards and has completed compatibility adaptation with mainstream domestic databases, application servers and messaging/caching middleware:

· Databases: Oracle, DM (Dameng), KingbaseES, GBase, Shentong (OSCAR), Huawei GaussDB;

· Application servers: TongWeb, BES, Apusic;

· Messaging middleware: TongLINK/Q, Apache Kafka, RabbitMQ;

· Caching and distributed transactions: Redis, Seata.

These adaptations allow the RVE1300 to slot seamlessly into an enterprise's existing Xinchuang IT architecture, advancing domestic substitution and intelligent upgrading on both fronts without rebuilding silos or duplicating investment.

Enterprise Data Security and Compliance: Data Never Leaves the Domain

For automotive, financial and public-sector customers with strict data sovereignty requirements, no matter how capable the model, the value is lost once data leaves the domain. The SOINAWE-RVE1300 answers at the architecture level:

· Data stays in-domain: deployed fully on-premises in the enterprise's own data center, all model inference, vector retrieval and knowledge-base matching complete inside the domain — R&D drawings, process parameters and business data never cross the enterprise boundary;

· Open-source and auditable: built on the open RISC-V instruction set and a fully open-source software stack with no closed-source black boxes, enabling independent code audits and removing dependence on foreign cloud services and uncontrollable supply chains;

· Full-stack autonomy: chips, servers, engine and platform are all independently controlled, and combined with the domestic database and middleware ecosystem, meeting security review and Xinchuang compliance requirements of critical infrastructure industries;

· Fine-grained permissions and audit: combined with the AWE engine's permission system and audit logs, full-chain control over who can use which model, access which knowledge, and leave which traces.

Built Around Automotive Scenarios: Keeping AI Inside the Enterprise

Many of SOIN Technology's customers come from the automotive industry, and the SOINAWE-RVE1300 is purpose-designed for automotive R&D and operations scenarios:

· R&D knowledge base and intelligent Q&A: vehicle R&D involves tens of thousands of technical standards, design specifications, test reports and legacy vehicle documents. Deployed inside the automaker's intranet, the RVE1300 powers a local R&D knowledge-base agent — engineers retrieve standard clauses, compare historical designs and draft review documents in natural language, while core R&D data stays entirely inside the enterprise network;

· Intelligent coding and R&D efficiency: combined with the AWE engine and an intelligent coding assistant, 72B-class code models serve the R&D team on-premises with code generation, review and refactoring suggestions, with measured throughput supporting organization-wide concurrent use;

· Multimodal quality inspection and design review: complex AI models fusing text, image and audio run smoothly on the RVE1300, supporting appearance defect detection and multi-dimensional intelligent design review;

· AI video generation and marketing content: generating high-quality dynamic content from text or images for vehicle promotion and marketing material — unreleased designs and campaigns never leave the domain;

· Structured governance of massive drawings and documents: the AWE engine crystallizes automotive knowledge graphs and business process logic into reusable agent assets, so expertise accumulates with the platform and is reused across the organization.

Three Core Application Scenarios

· LLM inference: high-performance inference for large language models at the trillion-parameter scale, providing a solid compute foundation for intelligent dialogue, content creation and knowledge Q&A;

· Multimodal fusion models: efficient processing of complex AI models fusing text, image and audio, unifying multi-dimensional understanding and intelligent analysis;

· AI video generation: smooth running of the latest video generation and processing models, rapidly producing high-quality dynamic content from text or images for film, television and marketing.

The launch of the SOINAWE-RVE1300 completes a key piece of SOIN Technology's three-in-one enterprise AI digital foundation: "AI compute hardware + AWE agent engine + Xinchuang software ecosystem". SOIN Technology will continue working with ecosystem partners such as ESWIN Computing to provide automotive and other industry customers with secure, compliant and independently controlled intelligent infrastructure — so enterprise LLM applications truly run in their own data centers, on their own data, within their own boundaries.