OCP-Solutions-With-NVIDIA-GPUs-For-New-Gen-AI-Presented-by-Aivres

James Zo, research fellow at Aivres, presents the company's OCP-conformant AI server portfolio at the OCP Global Summit 2025 in San Jose. Aivres offers three solution tiers — an OCP Full Stack rack solution (KRS8500) with 72 NVIDIA Blackwell GPUs for trillion-parameter models, an OCP OAM Module solution (KR9288/KR5288) for very large models, and a PCIe Card GPU solution for large models under 10B parameters. The talk covers hardware architecture, liquid cooling design, Open BMC-based diagnosis automation, and a full-stack software platform co-developed with NVIDIA.

Key Takeaways

Aivres KRS8500 OCP Full Stack rack packs 72 NVIDIA GB300 NVL72 Blackwell GPUs with integrated liquid cooling, networking, storage, and Aivres Diagnosis Automation — targeting trillion-parameter models like GPT-4 (1.8T parameters).
Aivres liquid cooling secondary loop uses comprehensive three-tier leakage detection, redundant temperature sensors, and real-time conductivity/turbidity/pH monitoring — addressing first-time deployment concerns common among enterprise customers.
Aivres adopts Open BMC with Redfish to unify server diagnosis automation across pre-boot, booting, and post-boot stages, replacing scattered bus protocols (I2C, USB, BIOS SSIF) with a single API for state registers, telemetry, and firmware update.
The Aivres full-stack AI software platform layers BIOS/firmware/OS/driver, a converged NVIDIA + Aivres resource scheduling framework, and AI application development tools (model development, training, inference, fine-tuning) — with Ubuntu 22.04 as OS and NeMo Megatron among supported DL frameworks.
Aivres commits to delivering guaranteed cluster application performance at the criteria agreed at order time, and ships systems pre-configured with optimal hardware settings determined through internal benchmark and open-source model tuning.
Aivres AI infrastructure management tool achieves a 92% diagnostic accuracy rate and supports unified management of servers, storage, networks, and facilities, with compatibility across multiple server vendors.