# Aivres KRS8000 datasheet (parsed)

Parsed from Aivres's PDF https://aivres.com/wp-content/uploads/KRS8000v3.1.pdf on 2026-10-15 (2 pages; OCR and table reconstruction by Reducto; tables are HTML inside this Markdown, images are figure crops from the PDF). Specifications belong to Aivres and are subject to change without notice; the PDF is the authority.

Products: [krs8000v3](https://aivres.upstation.io/p/krs8000v3)

---


<!-- page: 1 -->
*AIVRES*

DATASHEET

<b>KRS8000V3</b>

<b>AI Rack Based on NVIDIA GB200 NVL72</b>

<b>KRS8000V3 is an L11 AI rack based on NVIDIA GB200 NVL72, integrating 36 NVIDIA Grace™ CPUs and 72 NVIDIA Blackwell GPUs in a rack-scale, liquid-cooled architecture, achieving breakthrough performance in real-time trillion-parameter large language model (LLM) inference and training.</b>

<b>KRS8000V3 with GB200 NVL72 is poised to redefine performance benchmarks for AI, HPC, and data analytics, making it a pivotal component in next-generation computing infrastructure.</b>

<b>30X</b>
<b>LLM Inference</b>
vs. Nvidia H100
Tensor Core GPU

25X
Energy Efficiency
vs. H100

4X
LLM Training
vs. H100

<figure style="margin: 1rem 0; padding: 0.6rem; background: rgba(255,255,255,0.04); border-radius: 4px;">
  <img src="images/reducto/p001_figure_0011.jpg" alt="figure p1" style="max-width: 100%; max-height: 480px; display: block; margin: 0 auto;" />
  <figcaption style="font-size: 0.78rem; color: rgba(255,255,255,0.55); text-align: center; margin-top: 0.5rem;">KEE

XCE

XCE

KEE

EXCE

EXCE

EXCE

XCC

TEE</figcaption>
</figure>

18X
Data Processing
vs. H100

<b>KEY FEATURES</b>

<b>NVIDIA Blackwell Rack-Scale Architecture:</b>

• Connects 72 Blackwell GPUs via NVIDIA® NVLink™.

• Delivers 130 TB/s of low-latency communication bandwidth.

• Acts as a single massive GPU for efficient processing.

<b>Performance Enhancements:</b>

• Achieves 30X faster real-time trillion-parameter LLM inference compared to previous generations.

• 4X faster training for large language models using FP8 precision.

<b>Advanced Data Processing:</b>

• Includes a hardware decompression engine supporting LZ4, Deflate, and Snappy formats.

• Provides up to 800 GB/s decompression throughput.

• Achieves 18X faster performance for database query benchmarks compared to traditional CPUs.

<b>Memory and Bandwidth:</b>

• Offers 8 TB/s high memory bandwidth.

• NVIDIA Grace CPU NVLink®-C2C interconnect ensures high-speed data transfer.

<b>USE CASES</b>

<b>AI Training and Inference:</b>

• Optimized for training large-scale models and performing real-time inference.

• Particularly effective for transformer-based models and other resource-intensive AI applications.

<b>Big Data Analytics:</b>

• Enhances the efficiency of big data processing pipelines.

• Reduces storage costs and processing times for large datasets.

<b>Scientific and Engineering Simulations:</b>

• Speeds up simulations in various domains, including fluid dynamics, circuit design, and more.

<figure style="margin: 1rem 0; padding: 0.6rem; background: rgba(255,255,255,0.04); border-radius: 4px;">
  <img src="images/reducto/p001_figure_0037.jpg" alt="figure p1" style="max-width: 100%; max-height: 480px; display: block; margin: 0 auto;" />
  <figcaption style="font-size: 0.78rem; color: rgba(255,255,255,0.55); text-align: center; margin-top: 0.5rem;">3 x 1Gb Management Switch

QDQIQO 4 x 1RU Power Shelf 33KW

10 x Compute Trays

9 x Non Scalable NVLink Switch
Trays

8 x Compute Trays

4 x 1RU Power Shelf 33KW</figcaption>
</figure>

<u>www.aivres.com</u>

Copyright © Aivres. All Rights Reserved. The information contained herein is subject to change without notice.

Ver 3.0

<!-- page: 2 -->
<b>Technical Specifications</b>

<b>Model</b>: <b>NVIDIA® GB200 NVL72</b>
<b>Configuration</b>: 36 Grace CPU and 72 Blackwell GPUs
<b>FP4 Tensor Core2</b>: 1,440 PFLOPS
<b>FP8/16 Tensor Core2</b>: 720 PFLOPS
<b>INT8 Tensor Core2</b>: 720 POPS
<b>FP16/BF16 Tensor Core2</b>: 360 PFLOPS
<b>TF32 Tensor Core</b>: 180 PFLOPS
<b>FP32</b>: 6,480 TFLOPS
<b>FP64</b>: 3,240 TFLOPS
<b>FP64 Tensor Core</b>: 3,240 TFLOPS
<b>GPU Memory | Bandwidth</b>: Up to 13.39 TB HBM3e | 576 TB/s
<b>NVLink Bandwidth</b>: 130TB/s
<b>CPU Core Count</b>: 2,592 Arm® Neoverse V2 cores
<b>CPU Memory | Bandwidth</b>: Up to 17.28 TB LPDDR5X | Up to 18.4 TB/s

<b>Rack Specifications</b>

<b>Dimensions</b>: 600mm (23.6") W x 2236mm (88") H x 1200mm (47.2") L
<b>Weight</b>: 1,553.27 kg (3,434.37 lbs)
<b>NVL Config</b>: 72x 1
<b>NV OOB Switch</b>: Option 1: 3 x SN2201 DC
Option 2: 4 x SN2201 DC
<b>NVL Cartridge</b>: 4
<b>Rack Type (Per Rack)</b>: 9x 1U NVLink Switch Trays
18x 1U Compute Trays
8x 1U Power Shelves
<b>Power-Shelf</b>: 8x 33kW
<b>Busbar</b>: 1,400A
<b>Rack Manifold</b>: Option 1: 44RU, BF
Option 2: 44RU, TF
<b>CDU</b>: Option 1: L2L In-Row
Option 2: L2L In-Rack
Option 3: L2A Sidecar

<b>Compute Tray</b>

<b>CPU/GPU</b>: 2x Grace CPUs + 4x Blackwell GPUs
<b>Cooling</b>: 1U liquid cooled
<b>Storage</b>: 8x E1.S NVMe SSDs
<b>M.2</b>: 1x Onboard NVMe / SATA M.2
<b>Front I/O</b>: 1 x USB 3.0, 1x Mgmt I/O, 1x RJ45, 1x Mini Display Port
<b>N-S Networking</b>: 2x FHFL PCIe 5.0 x16 (BF3 or NIC Card)
<b>E-W Networking</b>: 2x Mezzanine card on board
4x HHHL PCIe 5.0 x16 with 400G bandwidth
<b>Fan</b>: CPU region: 8x 12V 4056 hot-swap fans with N+1 redundancy
<b>Management</b>: DC-SCM BMC management module
<b>TPM</b>: Supports TPM 2.0

<b>Switch Tray</b>

<b>Type</b>: 2x NVLink X-800
<b>Bandwidth</b>: 14.4TB/s
<b>Cooling</b>: 1U liquid cooled
<b>Front I/O</b>: 2x RJ45, 1x USB, 1x UART

<u>www.aivres.com</u>

Copyright © Aivres. All Rights Reserved. The information contained herein is subject to change without notice.

Ver 3.1
