AI Training & Inference
Optimized for trillion-scale models, transformer networks, and real-time inference.
KRS8500V3 is Aivres’ next-generation L11 rack-scale platform, integrating 36 NVIDIA Grace CPUs and 72 NVIDIA Blackwell Ultra GPUs into a fully liquid-cooled architecture. Designed for trillion-parameter LLM training and inference, it delivers unprecedented performance density, scalability, and energy efficiency, setting the foundation for large-scale AI, HPC, and data analytics in data centers worldwide.
Compared to the previous generation KRS8000V3 based on NVIDIA GB200 NVL72, KRS8500V3 achieves up to 1.5x faster training and ~1.4x higher inference throughput, establishing new MLPerf records and redefining the standard for hyperscale AI computing.
1.5x
LLM Training Speedvs GB200
1.4x
LLM Inference Throughputvs GB200
Supporting Extreme Performance Applications
 |
Ultra Tensor Cores2x attention-layer acceleration, 1.5x AI compute FLOPS uplift Optimized for trillion-scale LLM and mixture-of-experts (MoE) workloads |
|
|
|
 |
Expanded HBM3e Memory288GB per GPU, 1.5x larger vs. previous gen Improved context length support and inference efficiency |
|
|
|
 |
NVIDIA ConnectX-8 SuperNICDual-device IO module delivering 800 Gb/s per GPU Enables high-throughput, low-latency scale-out networking |
|
|
|
 |
Rack-Scale Liquid CoolingModular CDU options (In-Row, In-Rack, Sidecar) Full lifecycle DLC service for deployment and operations |
KRS8500V3 Specifications
| NVIDIA® GB300 NVL72 | |
|---|---|
| Configuration | 72x Blackwell Ultra GPUs, 36x Grace CPUs |
| NVLink Bandwidth | 130TB/s |
| Fast Memory | Up to 40 TB |
| GPU Memory | Bandwidth | Up to 21 TB | Up to 576 TB/s |
| CPU Memory | Bandwidth | Up to 18 TB LPDDR5X | Up to 14.3 TB/s |
| CPU Core Count | 2,592 Arm® Neoverse V2 cores |
| FP4 Tensor Core | 1,080–1,400 PFLOPS (rack-level) |
| FP8/16 Tensor Core | 720 PFLOPS |
| INT8 Tensor Core | 23 PFLOPS |
| FP16/BF16 Tensor Core | 360 PFLOPS |
| TF32 Tensor Core | 180 PFLOPS |
| FP32 Tensor Core | 6 PFLOPS |
| FP64 / FP64 Tensor Core | 100 TFLOPS |
| Rack Specifications | |
|---|---|
| Dimensions (W x H x D) | 600mm x 2285mm x 1200mm (23.62” x 89.96” x 47.24”) |
| Weight | 1,360–1,590 kg (3,000–3,500 lbs) |
| NVL Config | 72x 1 |
| NV OOB Switch | Option 1: 2x SN2201_M Option 2: 3x SN2201_M Option 3: 4x SN2201_M |
| NVL Cartridge | 4 |
| Rack Type (Per Rack) | 9x 1U NVlink Switch Trays 18x 1U Compute Trays 8x 1U Power Shelves |
| Power-Shelf | (6+2)x 33kW |
| Busbar | 1,400A |
| Rack Manifold | Option 1: 44RU, BF Option 2: 44RU, TF |
| CDU | Option 1: L2L In-Row Option 2: L2L In-Rack Option 3: L2A Sidecar |
| Compute Tray | |
|---|---|
| CPU/GPU | 2x Grace CPUs + 4x Blackwell Ultra GPUs |
| Count | 18 per Rack |
| Cooling | CPU/GPU/CX8 liquid-cooled, others air cooling |
| Storage | 8x E1.S NVMe SSDs |
| M.2 | 1 x M.2 NVMe SSD |
| Front I/O | 1x USB 3.0, 1x Mgmt I/O , 1x RJ45, 1x mini display port |
| N-S Networking | 1x FHFL PCIe Gen 5 X16 (BF3) |
| E-W Networking | 2x CX7 Mezzanine on board (2x CX8 per Mezzanine) for 4x 800G OSFP Conn |
| Fan | CPU region: 8x 12V 4056 hot-swap fans with N+1 redundancy |
| Management | DC-SCM BMC management module |
| TPM | Supports TPM 2.0 |
| Switch Tray | |
|---|---|
| Type | 2x NVLink N5600_LD |
| Count | 9 per Rack |
| Bandwidth | 14.4TB/s |
| Cooling | 1U liquid cooled |
| Front IO | 2x RJ45, 1x USB, 1x UART |
