@PyTorch: Lablup (@lablupinc) has joined PyTorch Foundation as a Silver Member to advance open source AI! @lablupinc builds softw…

X AI KOLs Timeline 09/03/26, 12:00 AM News

open-source ai-infrastructure pytorch-foundation gpu-management software-tool lablup backend-ai

Summary

Lablup has joined the PyTorch Foundation as a Silver Member to advance open-source AI. They develop Backend.AI, a software tool for AI infrastructure that enables GPU orchestration and virtualization.

Lablup (@lablupinc) has joined PyTorch Foundation as a Silver Member to advance open source AI! @lablupinc builds software that powers AI—from desktops to data centers. With http://Backend.AI, anyone can choose, connect, and operate AI models at the right cost and scale. "Lablup has developed http://Backend.AI as open source since founding. PyTorch sits at the center of that work. Joining the Foundation is our way of contributing back to the ecosystem that has long supported us." — Jeongkyu Shin, CEO of Lablup

Original Article

View Cached Full Text

Cached at: 09/03/26, 02:04 AM

Lablup (@lablupinc) has joined PyTorch Foundation as a Silver Member to advance open source AI!

@lablupinc builds software that powers AI—from desktops to data centers. With http://Backend.AI, anyone can choose, connect, and operate AI models at the right cost and scale.

“Lablup has developed http://Backend.AI as open source since founding. PyTorch sits at the center of that work. Joining the Foundation is our way of contributing back to the ecosystem that has long supported us.”

— Jeongkyu Shin, CEO of Lablup

Backend.AI | AI Infrastructure OS: GPU Orchestration & Virtualization

Source: https://www.backend.ai/

AI infrastructure, One unified OS

You focus on creating amazing AI- We’ll handle the backend.

Our expertise

Transforming complex operation intosimplicity

Backend.AI enables you to build, train, and serve AI models of any type any size, any scale. Run parallel computing jobs, train deep learning models, and deploy inference services within the same unified environment at ease. From fractional GPU sharing to multi-node clusters with thousands of GPUs, optimize your infrastructure to get every last bit.

Transforming complex operation into simplicity

Buying GPUs is the easy part

The most common enterprise GPU infrastructure challenges and how Backend.AI solves them.

Not fully utilizing the deployed GPU

Billions invested in GPU infrastructure, yet utilization rarely exceeds 30%. Teams allocate 8 GPUs but only use 2, a pattern that repeats across the organization.

See our solution →

Fractional GPU: utilize every last unit

Partition physical GPUs into precise fractional slices for concurrent multi-user sharing. Compatible with all NVIDIA GPUs at less than 5% overhead.

Company A GPU utilization 30% → 85% (2.8×)

Difficult to share GPUs among multiple teams

Multiple teams share a GPU cluster without clear allocation policies. Urgent experiments stall while other teams hold idle resources.

See our solution →

4-tier multi-tenancy for fair allocation

Manage resources hierarchically across Organization, Department, Project, and User levels. Idle resources are automatically redistributed via dynamic allocation.

University B Automated resource management without administrator intervention

Cloud adoption is restricted

Handling sensitive information restricts the adoption of external cloud AI services. AI must be operated in air-gapped environments.

See our solution →

Fully independent air-gapped operation

Installation, model deployment, and updates all operate without internet connectivity. multi-layer security sandboxing meets finance and defense-grade requirements.

Bank C air-gapped LLM operation

Facing GPU pricing and supply shortage challenges

Rising GPU prices and supply shortages drive interest in AMD, Intel, and domestic NPUs, but platform dependencies make it difficult to break away from a specific ecosystem.

See our solution →

12+ vendors under one unified interface

The Hardware Abstraction Layer (HAL) manages NVIDIA, AMD, Intel, Google TPU, Rebellions, FuriosaAI, Tenstorrent, and more from a single control plane.

Company D to introduce domestically developed NPUs on Backend.AI for its GPUaaS business

Must keep up with growing GPU demands

AI service expansion demands more GPUs, but scaling from dozens to hundreds of nodes introduces complexity that existing tools cannot handle.

See our solution →

Sokovan Orchestrator scales with you

The Sokovan scheduler handles multi-node, multi-tenant workloads at any scale. GPU and storage infrastructure grows seamlessly without re-architecture.

University E to increase 256% of GPU Node

More teams, exponentially more chaos

As teams grow, priority conflicts and queue issues arise. Staff turnover also risks losing operational know-how.

See our solution →

Software that handles scalable infrastructure

Backend.AI automates GPU operations through policy-based management. Resource quotas, scheduling rules, and access controls are defined as policies, enabling efficient operations.

University F operates campus infrastructure with minimal staff

About our heart

Sokovan. Meet the real expert who knows AI workload

Introducing Sokovan, the most powerful containerized workload orchestrator for AI infrastructure, designed from the ground up with AI in mind. At its core is a manager-agent dual-layer scheduling system that maximizes hardware utilization, performance, and operational efficiency across multi-tenant environments.

Explore Sokovan→

Sokovan. Meet the real expert who knows AI workload

From idle to ideal

Container-levelGPU virtualization

Backend.AI’s patented technology intercepts CUDA API calls inside containers to precisely control GPU resources at the software level. Multiple users safely share a single physical GPU while maintaining complete workload isolation.

400%

GPU utilization increase

75%

Infrastructure cost reduction

Learn more→

GPU utilization comparison

Multi-user GPU sharing

User A

0.25GPU

User B

0.5GPU

User C

0.15GPU

User D

0.1GPU

Hardware freedom

Avendor-neutral platformthat supports 12+ AI accelerators

Backend.AI manages 12+ AI accelerator types, including NVIDIA Blackwell, Intel Gaudi, AMD Instinct, Rebellions ATOM+, and FuriosaAI RNGD, through a Hardware Abstraction Layer (HAL) under a single unified interface. Develop, train, and deploy with the same workflow regardless of the underlying hardware.

An intelligent platform that understands storage

Any storage,one interface.

Access data the same way regardless of the storage backend: VAST, WEKA, PureStorage, IBM Storage Scale, and more. With NVIDIA GPUDirect Storage support, transfer data directly from storage to GPU memory.

Learn more→

Mind the gap

Complete AI environments in air-gapped networks

Reservoir AI (model hub) + Reservoir (package repository)

Keep AI models and tools up to date, even in air-gapped networks. Reservoir completes the software supply chain for isolated AI infrastructure.

Explore Reservoir→

Air-Gapped Network Boundary

Backend.AI Control Plane

Control

API Gateway

Auth / RBAC

Scheduler

Data

Reservoir AI

Reservoir

GPU Nodes

H100H100A100A100B200B200B200

Intuitive web-based management

Lowering the barrierto GPU utilization and operations

Backend.AI WebUI lets you manage GPU clusters, monitor resources, and govern multi-tenant environments through a browser, with no CLI expertise needed. From session management to policy-based resource allocation, everything happens in one unified interface.

Explore WebUI→ Lowering the barrier to GPU utilization and operations

NVIDIA DGX-Ready Software

Built for NVIDIA DGX, certified to run at enterprise scale

As NVIDIA DGX-Ready Software, Backend.AI delivers seamless operation with NVIDIA DGX systems and Grace Blackwell. Container-level GPU virtualization combined with the Sokovan orchestrator optimizes utilization for AI and HPC workloads, in various scales.

Explore DGX-Ready→

Built for NVIDIA DGX, certified to run at enterprise scale

Works the same in any environment

Choose the deployment model that fits your organization’s security policies and infrastructure: on-premises, cloud, or hybrid.

On-Premises

All data and workloads are processed within your own data center. Operate independently even in air-gapped environments.

Full data sovereignty
Air-gapped support
Easy regulatory compliance

Cloud

Runs on GPU instances across public clouds like AWS, GCP, Azure, and OCI. Start immediately with zero upfront investment.

Start instantly with no upfront cost
Elastic scale up/down
Instant global region deployment

Hybrid

Manage on-premises and cloud from a single control plane. Keep sensitive data on-premises while bursting workloads to the cloud.

Unified single control plane
Policy-based workload placement
Protect existing on-prem investment

Dashboard & management

Monitor, manage, and control everything

all-smi — Cluster Overview

Cluster Overview

Nodes200/200

GPU Cores1600

Total VRAM220.3TB

Avg. Temp68°C

Total Power684kW

Total RAM320.0TB

Tabsnode-0001node-0002node-0003 ... node-0200

GPUModelUtilVRAMTempPowerGPUH200 14179.1%84.2/141GB78°C511/700WGPUH200 14128.8%90.0/141GB63°C420/700WGPUH200 14136.1%61.6/141GB63°C385/700WGPUH200 14166.2%69.3/141GB75°C502/700Wh:Help q:Exit c:CPU →:Tabs s:Scroll

The dashboard lets you monitor resource usage, session status, and infrastructure health at a glance. With All-SMI, you can monitor heterogeneous accelerators and servers including NVIDIA, AMD, Intel Gaudi, and Google TPU from a single interface. Team and user management features are built in for organizational control.

Learn about All-SMI→