Ultra-Dense Peak Performance
The KR2288V3 platform features NVIDIA HGX Rubin NVL8 and is engineered to deliver over 5× performance improvement in FP4 precision inference, targeting next-generation AI workloads that demand extreme efficiency and scalability.
Built on an advanced HBM4 high-bandwidth memory architecture, the platform is optimized for long-context processing, million-token inference, and bandwidth-intensive AI scenarios, ensuring sustained throughput for large-scale generative models.
All components are designed with a rack-level, full liquid-cooling solution, enabling ultra-high density deployment with reduced power consumption and improved thermal efficiency.
 |
Next-Generation HBM4 ArchitectureEquipped with HBM4 memory, each GPU operates at up to 2.3 kW, delivering 5× overall performance improvement at NVFP4 precision, optimized for large-scale AI inference workloads. |
|
|
|
 |
Modular 1U Liquid-Cooled CPU ArchitectureCPU module features a fully liquid-cooled, modular design with upgrade-ready architecture, ensuring seamless platform evolution for future processor generations. |
|
|
|
 |
End-to-End Intelligent Liquid Cooling ManagementComprehensive monitoring across node → rack → CDU → intelligent valve, forming a fully integrated, closed-loop smart cooling management system. |
|
|
|
 |
Front-Accessible OS Drive with Cold-Swap SupportFront-side OS storage supports cold-swap replacement, improving serviceability and operational availability in large-scale deployments. |
|
|
|
 |
Enhanced Security with EROT ProtectionSupports EROT (Enhanced Root of Trust) functionality to prevent unauthorized firmware attacks, meeting Zero Trust platform security requirements. |
Specifications
| Model | KR2288-X3 |
|---|---|
| Form Factor | 2U rackmount |
| Processor | 2x Intel ® Xeon® 6 processors (6700P-Series) |
| Memory | 32x DDR5-6400 RDIMMs, up to 4TB DDR5 ECC |
| GPU | NVIDIA HGX Rubin 8-GPU |
| Storage | 8x NVMe E1.S + 2x NVMe M.2 |
| PCIe | 1x PCIe 6.0 x16 FH3/4L with BF4240L |
| Network Interface | 8x OSFP (IB XDR, ETH 800G) from CX-9 |
| Power | 54V Busbar |
| Cooling | In-Rack CDU, In-Row CDU, 34LPM Inlet 45C PG-25 |
| Rear I/O | MicroUSB x1, In-Band Management (I210), USB 3.0 x2, Mini DP x1, Out-Of-Band Management |
| Management | ASPEED AST2600 BMC Module |
| Security | TPM 2.0, ERoT |
| Power Consumption | 24KW |
| Dimensions | 448mm (W) x 87mm (H) x 800mm (D) |
| Weight | ~55Kg |
Performance & Acceleration
| Precision/Metric | NVIDIA Rubin (Single GPU) |
|---|---|
| FP4 Tensor Core | 50 PFLOPS (NVFP4) |
| FP8/FP6 Tensor Core | 17.5 PFLOPS |
| FP16/BF16 Tensor Core | 4 PFLOPS |
| FP32 Tensor Core | 130 TFLOPS |
| FP64/FP64 Tensor Core | 33 TFLOPS |
| INT8 Tensor Core | 2 PFLOPS |
| GPU Memory | 288 GB HBM3e |
| Memory Bandwidth | 15 ~20 TB/s |
| TDP | 2,300W per GPU |