Specifications
Form Factor
10U rackmount, fully modular chassis
GPU Platform
8x AMD Instinct MI350X (or MI325X) GPU OAM modules
GPU Interconnect
4th-Gen AMD Infinity Fabric, fully connected
GPU Memory
Up to 2.3 TB total HBM3E
Peak Memory Bandwidth
Up to 48 TB/s aggregate (64 TB/s per AMD platform reference)
Peak AI Performance
Up to 161 PFLOPS FP4 / FP6 at platform level
GPU TDP
1,000 W per accelerator
CPU
2x AMD EPYC 9005 series processors
Memory
Up to 24 DIMM slots, DDR5-6400 at 1 DIMM per channel
PCIe Slots
10x PCIe 5.0 x16 FHHL + 2x PCIe 5.0 x16 HHHL (12 total)
Design Standard
Industry-standard OCP design for rapid deployment
Power Distribution
12 V and 54 V partitioned power rails
Power Redundancy
N+N partitioned PSU to CPU nodes plus N+1 redundant supplies
Cooling
Air cooled with 20 fans; front-cabling layout cuts I/O temperature 20%
Cabling
Front-cable configuration reduces rear cabling and improves airflow
Target Workload
Exascale AI training, generative AI, agentic AI, HPC simulation
Maintenance
Modular chassis for field service and scalability
Overview
The Aivres KR9298 is a 10U heterogeneous accelerated AI server built around eight AMD Instinct MI350X (or MI325X) GPU OAM modules connected in a fully-connected topology over 4th-Generation AMD Infinity Fabric links, mounted on an industry-standard OCP design.
Two AMD EPYC 9005 processors feed up to 24 DDR5-6400 DIMM slots, and the eight-OAM accelerator complex delivers up to 2.3 TB of total HBM3E memory with up to 48 TB/s of peak aggregate memory bandwidth. AMD's MI350 series platform reference cites up to 161 PFLOPS of FP4/FP6 peak performance for the 8-GPU configuration.
I/O expansion runs to twelve PCIe slots — ten full-height half-length PCIe 5.0 x16 plus two half-height half-length — for NICs, DPUs and storage HBAs. Power is partitioned 12 V / 54 V with N+N redundancy toward CPU nodes and N+1 redundant supplies toward the GPU complex. Twenty fans and a front-cabling layout (which Aivres says lowers I/O component temperatures by about 20%) keep the machine air-coolable for standard data halls.
Key Benefits
2.3 TB of HBM3E across eight MI350X OAM modules. Full 4th-Gen Infinity Fabric mesh for all-to-all GPU bandwidth. 12 PCIe 5.0 slots for scale-out networking. OCP-standard design for faster deployment and serviceability.
Applications
Exascale AI training and HPC simulation, large-model fine-tuning, generative and agentic AI serving, scientific computing, and analytics clusters replacing older MI300X/MI325X nodes.
Request a Quote — AIVRES KR9298 — 10U 8-OAM AI SERVER FOR AMD INSTINCT MI350 SERIES
QS Compute — global B2B supply of AI computing hardware, edge AI systems and accelerators. Volume pricing, 15-day sample lead time.
Get Your Quote →