Product Brochure

AI

Hyperscale Training Platform

KR6288V2

KR6288V2, the 6U hyperscale training platform equipped with dual 4th or 5th Gen Intel Xeon Scalable Processors or AMD

EPYCTM 9004 Series Processors and 1x NVIDIA HGX-Hopper-8GPU module, features industry-leading performance, ultimate

I/O expansion, and ultrahigh energy efficiency. The precisely optimized system architecture with 4x CPU to GPU bandwidth,

up to 4.0Tbps networking bandwidth, 8TB system memory, and 300TB massive local storage can fully satisfy the communication and capacity demands of multi-dimensional parallelism training for giant-scale models. 12 PCIe expansions can be

flexibly configured with CX7, OCP3.0, and multiple SmartNICs, making it an ideal solution for both on-premises and cloud

deployment. It is built to handle the most demanding AI computing tasks like trillion-parameter Transformer model training,

massive recommender systems, AIGC, and Metaverse workloads.

Overview

Features

■ Unprecedented Performance

·Powered by 8* NVIDIA latest GPUs in a 6U chassis, TDP up to 700W.

·Support 2x 4th or 5th Gen Intel Xeon Scalable Processors or AMD

EPYCTM 9004 Series Processors.

·Industry-leading performance with 16PFlops AI performance by 3

times enhancement. The Transformer Engine significantly accelerates

the training speed of GPT large model.

■ Optimized Energy Efficiency

·Extremely low air-cooled heat dissipation overhead, less fan,

higher power efficiency.

·54V, 12V separated power supply with N+N redundancy reducing

power conversion loss.

·Intelligent regulation of heat dissipation in different layers to

reduce the power consumption and noise

■ Leading Architecture Design

·Lightning-fast intra-node connectivity with 4x CPU to GPU

bandwidth improvement.

·Ultra-high scalable inter-node networking with up to 4.0Tbps

non-blocking bandwidth.

·Cluster-level optimized architecture, GPU : Compute Network :

Storage Network = 8:8:2.

■ Multi-scenarios Adaptation

·Full modular design and extremely flexible configurations

satisfying both on-premises and cloud deployment.

·Easily harness large-scale model training, such as GPT-3, LLaMA

and Stable diffusion.

·Diversified SuperPod solutions accelerating the most

cutting-edge innovation including AIGC, AI4Science and Metaverse.

KR6288V2 Air Cooling

01

Specifications

Height

GPU

Processor

Memory

Storage

M.2

PCIe Slot

RAID

Front I/O

Rear I/O

OCP

Management

TPM

Fan

Power

Size

Weight

Environmental

Parameters

Model KR6288-E2-A0-R0-00

2x AMD EPYCTM 9004 Series Processors, Max cTDP 400W

24x DDR5 DIMMs, up to 4800MT/s

2x Onboard NVMe M.2 (optional)

KR6288-X2-A0-R0-00

6U

1x NVIDIA HGX-Hopper-8GPU module, TDP up to 700W per GPU

2x 4th or 5thGen Intel Xeon Scalable Processors, TDP 350W

32x DDR5 DIMMs, up to 5600MT/s

24x 2.5’ SSD, up to 16x NVMe U.2

2x Onboard NVMe/SATA M.2 (optional)

Support 10x PCIe Gen5 x16 slots. One PCIe Gen5 x16 slot can be replaced with two x16 slots (PCIe Gen5 x8 rate).

Optional support Bluefield-3, CX7, and various SmartNICs

Optional support RAID 0/1/10/5/50/6/60, etc., support Cache super capacitor protection

1x USB 3.0, 1x USB 2.0, 1x VGA

2x USB3.0, 1x MicroUSB, 1x VGA, 1x RJ45

Optional support 1x OCP 3.0, support NCSI

DC-SCM BMC management module with Aspeed 2600

TPM 2.0

GPU region: 6x 54V hot-swap fans with N+1 redundancy

CPU region: 6x 12V hot-swap fans with N+1 redundancy

2x 12V 3200W and 6x 54V 2700W, Titanium CRPS PSU with N+N redundancy

Width: 447mm, Height: 263mm, Length: 860mm

Net weight 92kg（Cross weight: 107kg）

Working temperature：10℃~35℃; Storage temperature：-40℃~70℃

Working humidity：10%~80% R.H.；Storage humidity：10%~93% R.H.

KR6288V2 Liquid Cooling

Features

■ Unprecedented Performance

·Powered by 8* NVIDIA latest GPUs in a 6U chassis, TDP up to 700W.

·Support 2x 4th or 5th Gen Intel Xeon Scalable Processors or AMD

EPYCTM 9004 Series Processors.

·Industry-leading performance with 16PFlops AI performance by 3

times enhancement. The Transformer Engine significantly accelerates

the training speed of GPT large model.

■ Optimized Energy Efficiency

·54V, 12V separated power supply with N+N redundancy reducing

power conversion loss.

·Direct liquid cooling design with more than 80% cold plate

coverage, PUE≤1.15.

·Support water inflow up to 45°C, and the primary side supports

natural cooling.

■ Leading Architecture Design

·Lightning-fast intra-node connectivity with 4x CPU to GPU

bandwidth improvement.

·Ultra-high scalable inter-node networking with up to 4.0Tbps

non-blocking bandwidth.

·Cluster-level optimized architecture, GPU : Compute Network :

Storage Network = 8:8:2.

■ Multi-scenarios Adaptation

·Full modular design and extremely flexible configurations

satisfying both on-premises and cloud deployment.

·Easily harness large-scale model training, such as GPT-3, LLaMA

and Stable diffusion.

·Diversified SuperPod solutions accelerating the most

cutting-edge innovation including AIGC, AI4Science and Metaverse.

02

Specifications

Height

GPU

Processor

Memory

Storage

M.2

PCIe Slot

RAID

Front I/O

Rear I/O

OCP

Management

TPM

Fan

Power

Size

Weight

Environmental

Parameters

Model KR6288-E2-C0-R0-00

2x AMD EPYCTM 9004 Series Processors, Max cTDP 400W

24x DDR5 DIMMs, up to 4800MT/s

2x Onboard NVMe M.2 (optional)

KR6288-X2-C0-R0-00

6U

1x NVIDIA HGX-Hopper-8GPU module, TDP up to 700W per GPU

2x 4th or 5th Gen Intel Xeon Scalable Processors, TDP 350W

32x DDR5 DIMMs, up to 5600MT/s

24x 2.5’ SSD, up to 16x NVMe U.2

2x Onboard NVMe/SATA M.2 (optional)

Support 10x PCIe Gen5 x16 slots. One PCIe Gen5 x16 slot can be replaced with two x16 slots (PCIe Gen5 x8 rate).

Optional support Bluefield-3, CX7, and various SmartNICs

Optional support RAID 0/1/10/5/50/6/60, etc., support Cache super capacitor protection

1x USB 3.0, 1x USB 2.0, 1x VGA

2x USB3.0, 1x MicroUSB, 1x VGA, 1x RJ45

Optional support 1x OCP 3.0, support NCSI

DC-SCM BMC management module with Aspeed 2600

TPM 2.0

Supports GPU, NVSwitch, CPU cold plate cooling, and supports high temperature water inlet up to 45℃

GPU region: 5x 54V hot-swap fans with N+1 redundancy

CPU region: 6x 12V hot-swap fans with N+1 redundancy

2x 12V 3200W and 6x 54V 2700W, Titanium CRPS PSU with N+N redundancy

Width: 447mm, Height: 263mm, Length: 860mm

Net weight 90kg（Cross weight: 105kg）

Working temperature：10℃~35℃; Storage temperature：-40℃~70℃

Working humidity：10%~80% R.H.；Storage humidity：10%~93% R.H.

Open Accelerator Al Server

KR6298V2

KR6298V2, the 6U open accelerator platform equipped with dual 4th or 5th Gen Intel Xeon Scalable Processors and 8x Intel

Gaudi2 OAMs, features industry-leading performance, ultimate I/O expansion, and ultrahigh energy efficiency. The precisely

optimized system architecture with 4x CPU to GPU bandwidth, up 4.0Tbps networking bandwidth, 8TB system

memory, and 300TB massive local storage can fully satisfy the communication and capacity demands of multi-dimensional

parallelism training for giant-scale models. 12 PCIe expansions can be flexibly configured with CX7, OCP3.0, and multiple

SmartNICs, making it an ideal solution for both on-premises and cloud deployment. It is built to handle the most demanding

AI computing tasks like LLM training.

Overview

Features

■ Unprecedented Performance

·Powered by 8x Intel Gaudi2 OAMs in a 6U chassis, TDP up to

600W.

·Support 2x 4th or 5th Gen Intel Xeon Scalable Processors.

·Industry-leading performance with up to 896GB/s bidirectional

communication between node chips.

■ Optimized Energy Efficiency

·Extremely low air-cooled heat dissipation overhead, less fan,

higher power efficiency.

·54V, 12V separated power supply with N+N redundancy reducing

power conversion loss.

·Intelligent regulation of heat dissipation in different regions to

reduce the power consumption and noise.

■ Leading Architecture Design

·Lightning-fast intra-node connectivity with 4x CPU to OAM

bandwidth improvement.

·Ultra-high scalable inter-node networking with up to 4.0Tbps

non-blocking bandwidth.

·Cluster-level optimized architecture, OAM: Compute Network:

Storage Network = 8:8:2.

■ Multi-scenarios Adaptation

·Full modular design and extremely flexible configurations

satisfying both on-premises and cloud deployment.

·Easily harness large-scale model training, such as GPT-3, LLaMA,

and stable diffusion.

·Diversified SuperPod solutions accelerating the most

cutting-edge GenAI innovation.

03

Specifications

Height

GPU

Processor

Memory

Storage

M.2

PCIe Slot

RAID

Front I/O

Rear I/O

OCP

Management

TPM

Fan

Power

Size

Weight

Environmental

Parameters

Model KR6298-X2-A0-R0-00

6U

8xIntel Gaudi2 OAMs, TDP up to 600W per OAM

2x 4th or 5th Gen Intel Xeon Scalable Processors, TDP 350W

32x DDR5 DIMMs, up to 5600MT/s

24x 2.5' SSD, up to 16x NVMe U.2

2x Onboard NVMe/SATA M.2 (optional)

Support 10x PCIe Gen5 x16 slots. One PCIe Gen5 x16 slot can be replaced with two x16 slots (PCIe Gen5 x8 rate).

Optional support Bluefield-3, CX7, and various SmartNICs

Optional support RAID 0/1/10/5/50/6/60, etc., support Cache super capacitor protection

1x USB 3.0, 1x USB 2.0, 1x VGA

2x USB3.0, 1x MicroUSB, 1x VGA, 1x RJ45

Optional support 1x OCP 3.0, support NCSI

DC-SCM BMC management module with Aspeed 2600

TPM 2.0

GPU region: 6x 54V hot-swap fans with N+1 redundancy

CPU region: 6x 12V hot-swap fans with N+1 redundancy

2x 12V 3200W and 6x 54V 2700W, Titanium CRPS PSU with N+N redundancy

Width: 447mm, Height: 263mm, Length: 860mm

Net weight 92kg（Cross weight: 107kg）

Working temperature：10℃~35℃;  Storage temperature：-40℃~70℃

Working humidity：10%~80% R.H.；Storage humidity：10%~93% R.H.

Powered by

Intel Processors

04

Versatile Computing

and Flexible Expansion AI Server

KR4268V2

KR4268V2 is the latest product of KAYTUS KR4268 series server. It is a new generation artificial intelligence server, that

provides excellent versatile computing performance and extremely flexible architecture, eight high-performance GPUs and

supports switching based on application scenarios. Equipped with 2x 4th or 5th Gen Intel Xeon scalable processors,

KR4268V2 provides up to 128 processor cores, 8TB system memory and 300TB local high-speed storage. It is optimized to

complex application scenarios such as deep learning, metaverse, AIGC, AI+Science, etc., providing the most adaptable

platform in the Intelligent computing era.

Overview

Features

■ Enhanced architecture

·Upgraded PCIE Gen5.0 Architecture, supports up to 600W power

consumption.

·Support latest generation GPU cards.

·Latest CXL1.1 interconnect technology support storage-level

memory space expansion and memory space sharing.

·Support latest 400G NDR Infiniband and 400G OCP 3.0.

■ Excellent performance

·Support 2x 4th or 5th Gen Intel Xeon Scalable Processors, TDP 350W.

·Built-in 8TB system memory and 300TB local high-speed storage

to improve data access speed.

·4 kinds of CPU-GPU topologies to flexibly match different AI

application scenarios.

·Up to 4 times PCIE bandwidth improvement meets high-throughput data transmission demands.

■ Flexible configuration

·Modular design supports flexible component-level free combination.

·Powerful versatile platform adapts to the latest AI accelerator

cards of various brands.

·25% improvement in cooling performance, 5% reduction in noise.

·Support N+N redundant 3000W platinum/titanium PSU.

■ Ecosystem

·Abundant global AI partners, AMD, Intel, etc.

·Leading deep learning frameworks, TensorFlow, PyTorch,

PaddlePaddle, etc.

·Powerful MotusAI artificial intelligence platform, algorithm

platform and application optimization service.

·Robust Meta-brain ecology delivers comprehensive application

algorithms and tailored solution services across all industries and

scenarios.

Specifications

CPU

GPU

Chipset

Memory

Internal PCIe

Front I/O

Rear I/O

Storage

RAID

OS

Cooling

Power

Size (W*H*D)

Temperature

Full load weight

Model KR4268-X2-A0-R0-00

2x 4th or 5th Gen Intel Xeon Scalable Processors, TDP 350W

Supports 8x Dual-slot FHFL PCIe interface GPU cards, and supports ≥ 5 PCIe 5.0 x16 slots

Intel Emmitsburg PCH C740

up to 32x DDR5 5600MHz RDIMM

Supports for 2x PCIE 5.0 expansion

2x USB, 1x VGA, 1x USB Type-C

1x Console, 2x USB 3.0, 1x RJ45, 1x VGA, 1x OCP 3.0 (support NC-SI)

24x 2.5 or 12x 3.5-inch SAS/SATA drive bays in the front, supports up to 16x NVME or E3.S

Built-in 2x M.2 NVME/SATA SSD

Supports RAID0, 1, 10, 5, 50, 6, 60, etc.

Supports Cache super capacitor protection, provide RAID state transition, RAID configuration memory

Supports mainstream operation system, Microsoft Windows Sever, Red Hat Enterprise Linux, Ubuntu Linux, CentOS, etc.

N+1 redundant system fans

4x 1600W/2000W/2200W/3000W 80Plus Platinum/Titanium PSUs, supports N+N redundancy

447mm × 174.5mm × 850mm

5 - 35°C / 41°F - 95°F

≤87kg

Powered by

Intel Processors

05 Multi-computing All Scenario AI Server

KR4268V2

KR4268V2 is the latest generation of KAYTUS KR4268 series server. It is a new generation of artificial intelligence server with

excellent multi-computing performance and flexible application in all scenarios. It is equipped with two AMD 4th Generation

processors with 5nm advanced process in 4U space, providing multiple up to 256 processor cores. Equipped with up to 10

GPU accelerator cards, it supports the industry's best Al accelerator card, and multiple PCIe topologies to ensure the most

efficient distribution of computational resources. Cooperating with MotusAI artificial intelligence platform, KR4268V2 creates

the most adaptable multi-computing power platform in the era of intelligent computing.

Overview

Features

■ Enhanced architecture

·Upgraded PCIE Gen5.0 architecture, supports up to 600W power

consumption.

·Support latest generation GPU cards.

·Latest CXL1.1 interconnect technology supports storage-level

memory space expansion.

·Support latest 400G NDR Infiniband network and 400G OCP 3.0.

■ Excellent performance

·Two latest 5nm AMD 4th Generation EPYC™ processors, up to 256

cores.

·Support 10 GPUs with powerful computing and efficient parallel

processing capabilities.

·Support 16 NVME or E3.S hard drives, greatly improving the data

access rate.

·The PCIE topology structure with up to 64GB/s bandwidth channel

meets high-throughput data transmission demands.

■ Flexible configuration

·Modular design concept supports flexible component-level

configuration and free combination.

·Powerful multi-computing platform adapts to the latest AI

accelerator cards of various brands.

·Three-level security system based on hardware security,

firmware security and software security.

·Support N+N redundant 3000W platinum/titanium power supply,

extreme system stability.

■ Ecosystem

·Abundant global AI partners such as AMD, Intel, etc.

·Leading deep learning frameworks including TensorFlow,

PyTorch, PaddlePaddle, etc.

·Cooperate with the powerful MotusAI artificial intelligence

platform, algorithm platform and application optimization services.

·Robust Meta-brain ecology delivers comprehensive application

algorithms and tailored solution services across all industries and

scenarios.

Specifications

CPU

GPU

Memory

Internal PCIe

Front I/O

Rear I/O

Local storage

RAID

Operation system

System heat dissipation

Power

Size (W*H*D)

Working temperature

Full weight

Model KR4268-E2-A0-R0-00

2x AMD 4th Generation EPYC™ processors, cTDP 400W

Up to 10x full-height full-length double-width PCIe interface GPU cards

24x DDR5 4800MHz RDIMM

Supports up to 2x PCIE 5.0 expansion

2x USB, 1x VGA, 1x USB Type-C

1x Console, 2x USB 3.0, 1x RJ45, 1x VGA, 1x OCP 3.0 (support NC-SI)

24x 2.5 or 12x 3.5-inch SAS/SATA drive bays in the front, supports up to 16x NVME or E3.S

Built-in 2x M.2 NVME SSD

Supports RAID0, 1, 10, 5, 50, 6, 60, etc.

Supports Cache supercapacitor protection, provide RAID state transition, RAID configuration memory

Supports mainstream operation system, Microsoft Windows Sever, Red Hat Enterprise Linux, Ubuntu Linux, CentOS, etc.

N+1 redundant system fans

4x 1600W/2000W/2200W/3000W 80Plus Platinum/Titanium PSUs, supports N+N redundancy

447mm × 174.5mm × 850mm

5 - 35°C / 41°F - 95°F

≤87kg

Powered by

AMD Processors

06

Artificial Intelligence Platform

MotusAI

MotusAI is a management system developed in house by Kaytus for the AI model development scenario. It is dedicated to

helping enterprises to build efficient deep learning development platforms, manage and schedule AI computing resources in a

unified manner, and effectively improve the utilization of computing resources. MotusAI equips AI development engineers

with a comprehensive AI development software stack and streamlined development process, significantly enhancing their

R&D productivity.

Overview

·Uniform monitoring, O&M and scheduling of AI resources and development services to ensure the AI platform operates

efficiently and continuously.

·Central management of development data to balance read speed and security.

·One-stop AI development environment to reduce deployment time and improve efficiency of AI development engineers.

·Automatic programming and management of AI training tasks to accelerate AI development and shorten development cycle.

Features

Advantages

■ Fine-grained scheduling of GPU

GPU shared scheduling strategy realizes single-card reuse of

GPU resources and supports the reuse of up to 64 tasks per

card. Allocation and isolation at any granularity are supported,

and users can dynamically request GPU resources based on the

video memory.

■ Severe data silos

The strategies of \"zero-copy\" transmission, multi-thread fetch,

incremental data update and affinity scheduling for training

data greatly shorten the data cache cycle and improve the

efficiency of model development and training.

■ Efficient distributed trainings

Support the extension of distributed trainings through MPI

Allreduce in TensorFlow, PyTorch and other mainstream

frameworks and provides standard UI operations, so that users

can submit distributed trainings through simple GPU computing

resources and training script configuration.

■ Fault tolerance mechanism

Provide fault tolerance for training tasks, enabling the platform

to effectively ensure continuous training of tasks and reduce

the waste of time in case of server crash or GPU failure.