Bare-Metal GPU Servers

8x Intel Gaudi 2 — 768 GB HBM2E for LLM Training & Inference

Name: 8x Intel Gaudi 2 Bare-Metal AI Server
Brand: LeaderGPU

Eight Intel Gaudi 2 AI accelerators on bare metal, each with 96 GB of HBM2E at 2.45 TB/s, on a dual Xeon Gold 6338N host with 2 TB of RAM. PyTorch and Hugging Face models run through the Intel Gaudi software stack. No hypervisor. No setup fees.

See Pricing & Order

From €175/day (8 GPUs) · excl. VAT

ISO 27001 & NEN 7510 Certified Data Rooms Tier 3 Compliant Data Rooms 13+ Years on Market

Built for Large Language Models

96 GB of HBM2E per accelerator and 768 GB across the server keep large models and long sequences in device memory

LLM Training & Fine-Tuning

Pre-train and fine-tune Llama-, Mistral- and Qwen-class models with PyTorch, DeepSpeed and Hugging Face Optimum Habana. BF16 and FP8 run on the Matrix Multiplication Engines.

LLM Inference & Serving

Serve 70B-class models across the eight accelerators with vLLM and its Gaudi plugin, with room in 768 GB of HBM2E for long-context KV caches.

On-Chip Networking

Each Gaudi 2 carries 24 × 100 GbE RoCE ports on the chip itself: 21 link the eight accelerators all-to-all for collective operations in distributed training, and 3 are for scale-out.

Generative AI & Multimodal

Train and run Stable Diffusion, vision-language and speech models from the Hugging Face Hub with Gaudi-optimized pipelines.

Data-Heavy Pipelines

2 TB of system RAM and 64 Xeon cores keep tokenization, data loading and preprocessing ahead of the accelerators.

An Alternative to NVIDIA GPUs

768 GB of accelerator memory for PyTorch workloads in one server, at a fixed daily, weekly or monthly price.

Software Compatibility

Supported through the Intel Gaudi software suite and its PyTorch integration

Intel Gaudi Software

Driver & Graph Compiler

PyTorch

Deep Learning

PyTorch Lightning

Training Framework

Optimum Habana

Hugging Face

Transformers

Hugging Face

Diffusers

Generative AI

DeepSpeed

Training Optimization

vLLM

LLM Serving

Ray

Distributed Computing

Docker

Gaudi Containers

Jupyter

Notebooks

Intel Gaudi 2 Server Pricing

No setup fees. No hidden costs. All prices exclude VAT.

Configuration	GPU Memory	RAM	Storage	Daily	Weekly	Monthly	Per GPU/mo	Status
8x Gaudi 2 2x Xeon Gold 6338N	768 GB	2048 GB	1.92 TB NVMe	€175	€990	€3,500	€438		Order
CPU CPU Intel Xeon Gold6338N, 32cores / 64 threads @ 2.2GHz ~ 3.5 GHz RAM 2048 GB Storage 1.92 TB NVME Network 1 Gbit

Why LeaderGPU

Trusted by researchers and enterprises across Europe

Bare-Metal Performance

No hypervisor, no virtualization overhead. Direct hardware access for maximum GPU throughput.

Ready in 1 Business Day

Your server is provisioned within 1 business day of your order, with no setup fees and no hypervisor.

Tier 3 Compliant Data Rooms

Enterprise-grade data rooms with ISO 27001 and NEN 7510 certification.

Flexible Billing

Daily, weekly, or monthly rental. No setup fees. Scale your commitment to match your project timeline.

13+ Years / 6x FD Gazellen

Operating since 2012. Among only 78 companies to receive the FD Gazellen award in 2024.

Enterprise Connectivity

Peering via AMS-IX and NL-ix. Four redundant internet connections for maximum uptime.

Data Room Infrastructure

Netherlands-based Tier 3 compliant data rooms with enterprise-grade redundancy

Tier 3 compliant data rooms supplier
ISO 27001 and NEN 7510 data room standards
Each server connected to dual-power feeds (UPS)
Generators: N+1
Stable 24°C cooling, backup units (N+1)
Four redundant internet connections with peering via NL-ix and AMS-IX
VESDA very early smoke detection + gas fire extinguishing
24/7 security guards, fingerprint scanners, CCTV, intrusion detection

Supported Operating Systems

Ubuntu 22.04 LTS

Ubuntu 24.04 LTS

Custom OS (1 business day)

24/7 Support

Ordering & incidents around the clock. Technical support Mon-Fri, 9:00-18:00 CET.

Frequently Asked Questions

How quickly will my server be provisioned?

Your server is provisioned within 1 business day of placing your order. A configuration that is busy shows the date it is available from in the pricing table.

How much memory does the Intel Gaudi 2 have?

Each Gaudi 2 carries 96 GB of HBM2E with 2.45 TB/s of bandwidth, so the server has 768 GB of accelerator memory in total. That holds a 70B-parameter model in BF16 with room left for the KV cache, and lets you fine-tune large models without offloading to host memory.

Does Gaudi 2 run CUDA code?

No. Gaudi 2 is an Intel AI accelerator, not an NVIDIA GPU. PyTorch models run on it through the Intel Gaudi software suite and Hugging Face Optimum Habana; DeepSpeed runs through Intel's Gaudi fork and vLLM through its Gaudi plugin, so most training and inference scripts need only small changes. Custom CUDA kernels have to be ported.

How does Gaudi 2 compare with the NVIDIA A100?

Gaudi 2 has more memory per card than the A100 80 GB (96 GB of HBM2E against 80 GB), higher memory bandwidth (2.45 TB/s against about 2 TB/s), and FP8 support, which the A100 lacks. The A100 has the broader software ecosystem through CUDA. For PyTorch training and inference of transformer models, Gaudi 2 is a cost-efficient alternative.

Which CPU, RAM and storage ship with the server?

The eight accelerators are paired with two Intel Xeon Gold 6338N processors (32 cores / 64 threads each), 2048 GB of system RAM and 1.92 TB of NVMe storage.

Are there any setup fees?

No. There are zero setup fees for any rental period — daily, weekly, or monthly. You only pay the rental price listed. All prices are exclusive of VAT.

Which operating systems are supported?

Gaudi 2 servers ship with Ubuntu 22.04 LTS or Ubuntu 24.04 LTS. Custom OS installations are available on request and require 1 business day. All servers come with root access so you have full control over the software stack.

Where are the servers located?

Our servers are hosted in Tier 3 compliant data rooms in the Netherlands, upheld to ISO 27001 and NEN 7510 standards, with peering via AMS-IX and NL-ix for excellent European and global connectivity.