Don’t miss out on NVIDIA Blackwell! Join the waitlist.
Contact sales
scalar

The Ultimate GPU Server for Deep Learning

Now available with NVIDIA H100 NVL Tensor Core GPUs
Get pricing
hpc-social-sc23-hgx-h200-2048x2048 1-1

Join the waitlist for NVIDIA H200 NVL GPUs

Lambda Scalar servers will soon be available with up to 8x NVIDIA H200 NVL GPUs for a total of 1128 GB VRAM. NVIDIA H200 NVL supports 141 GB HBM3e, peak memory bandwidth of 4.8TB/s, and 4-way NVLink.
Get started
Microsoft_logo Intuitive_logo Amazon_logo Anthem_logo Raytheon_logo Argonne_logo Sony_logo IBM_logo Google_logo Caltech_logo Berkeley_logo Netflix_logo

Lambda Scalar powered by NVIDIA H100 NVL GPUs

Lambda Scalar servers come with the new NVIDIA H100 NVL Tensor Core GPUs and deliver unprecedented performance, scalability, and security for every workload. NVIDIA H100 NVL GPUs feature fourth-generation Tensor Cores and the Transformer Engine with FP8 precision, further extending NVIDIA’s market-leading AI leadership with faster training and inference speedup on large language models.
Get started
scalarh100

Engineered for your workload

Tell us about your research and we'll design a machine that's perfectly tailored to your needs.

Contact sales
4U 1

8

GPUs from NVIDIA

192

cores and 384 threads

8192

GB of memory

253

TB of NVMe SSDs

Easily scale from server to cluster

As your team's compute needs grow, Lambda's in-house HPC engineers and AI researchers can help you integrate Scalar and Hyperplane servers into GPU clusters designed for deep learning.

  • Compute
    Scaling to 1000s of GPUs for distributed training or hyperparameter optimization.
  • Storage
    High-performance parallel file systems optimized for ML.
  • Networking
    Compute and storage fabrics for GPUDirect RDMA and GPUDirect Storage.
  • Software
    Fully integrated software stack for MLOps and cluster management.
Learn more
easily_scale_from_server_to_cluster

Service and support by technical experts who specialize in machine learning

Lambda Premium Support includes:

  • Up to 5 year extended warranty with advanced parts replacement
  • Live technical support from Lambda's team of ML engineers
  • Support for ML software included in Lambda Stack: PyTorch®, Tensorflow, CUDA, CuDNN, and NVIDIA Drivers
Learn more
service_by_technical_experts-Oct-29-2024-11-07-29-4416-PM

Plug in. Start training.

Our servers include Lambda Stack, which manages frameworks like PyTorch® and TensorFlow. With Lambda Stack, you can stop worrying about broken GPU drivers and focus on your research.

  • Zero configuration required
    All your favorite frameworks come pre-installed.
  • Easily upgrade PyTorch® and TensorFlow
    When a new version is released, just run a simple upgrade command.
  • No more broken GPU drivers
    Drivers will "just work" and keep compatible with popular frameworks.
Learn more

Your servers. Our data center.

Lambda Colocation makes it easy to deploy and scale your machine learning infrastructure. We'll manage racking, networking, power, cooling, hardware failures, and physical security. Your servers will run in a Tier 3 data center with state-of-the-art cooling that's designed for GPUs. You'll get remote access to your servers, just like a public cloud.

Learn more

Fast support

If hardware fails, our on-premise data center engineers can quickly debug and replace parts.

Optimal performance

Our state-of-the-art cooling keeps your GPUs cool to maximize performance and longevity.

High availability

Our Tier 3 data center has redundant power and cooling to ensure your servers stay online.

No network set up

We handle all network configuration and provide you with remote access to your servers.

Explore our research

ICCV 2021

Multiple Pairwise Ranking Networks for Personalized Video Summarization

We propose a model for personalized video summaries by conditioning the summarization process with predefined categorical labels.

Learn more

ICCV 2019

HoloGAN: Unsupervised Learning of 3D Reps. from Natural Images

We propose a novel generative adversarial network (GAN) for the task of unsupervised learning of 3D representations from natural images.

Learn more

NeurIPS 2018

RenderNet: A Deep ConvNet for Differentiable Rendering from 3D Shapes

We present a differentiable rendering convolutional network with a novel projection unit that can render 2D images from 3D shapes.

Learn more

SIGGRAPH Asia 2019

Adversarial Monte Carlo Denoising with Conditioned Aux. Feature Modulation

We demonstrate that GANs can help denoiser networks produce more realistic high-frequency details and global illumination.

Learn more
View all research

Technical Specifications

GPUs

Up to 8 dual-slot PCIe GPUs. Options include:

  • NVIDIA H100 NVL: 94 GB of HBM3, 14,592 CUDA cores, 456 Tensor Cores, PCIe 5.0 x16
  • NVIDIA L40S: 48 GB of GDDR6, 18,176 CUDA cores, 568 Tensor Cores, PCIe 4.0 x16
  • NVIDIA RTX 6000 Ada Generation: 48 GB of GDDR6, 18,176 CUDA cores, 568 Tensor Cores, PCIe 4.0 x16
  • NVIDIA RTX 5000 Ada Generation: 32 GB of GDDR6, 12,800 CUDA cores, 400 Tensor Cores, PCIe 4.0 x16
  • NVIDIA RTX 4500 Ada Generation: 24 GB of GDDR6, 7,680 CUDA cores, 240 Tensor Cores, PCIe 4.0 x16
  • NVIDIA RTX 4000 Ada Generation: 16 GB of GDDR6, 6,144 CUDA cores, 192 Tensor Cores, PCIe 4.0 x16

Processors

2 AMD EPYC or Intel Xeon Processors...

  • AMD EPYC 7004 (Genoa) Series Processors with up to 192 cores total
  • Intel Xeon 4th Gen (Sapphire Rapids) Scalable Processors with up to 112 cores total

System Memory

Up to 8 TB of 4800 MHz DDR5 ECC RAM in 32 DIMM slots

  • Up to 8 TB of 4800 MHz DDR5 ECC RAM in 32 DIMM slots

Storage

Up to 491.52 TB of storage via 16 hot-swappable U.2 NVMe SSDs...

  • Up to 491.52 TB of storage via 16 hot-swappable U.2 NVMe SSDs
  • Up to 61.44 TB of storage via 8 hot-swappable 2.5" SATA SSDs

Networking

2 RJ45 10 Gbps BASE-T LAN ports...

Built-in networking:

  • 2 RJ45 10 Gbps BASE-T LAN ports
  • 1 RJ45 1 Gbps BASE-T LAN out-of-band management port

Optional high-speed NIC. Options include:

  • NVIDIA ConnectX-7 400 Gb/s NDR InfiniBand Adapter, OSFP56, PCIe 5.0 x16
  • NVIDIA ConnectX-7 200 Gb/s NDR200 InfiniBand Adapter, OSFP56, PCIe 5.0 x16
  • NVIDIA ConnectX-7 200 Gb/s NDR200 InfiniBand/VPI Adapter, QSFP112, PCIe 5.0 x16
  • NVIDIA ConnectX-6 200 Gb/s HDR InfiniBand/VPI Adapter, QSFP56, PCIe 4.0 x16
  • NVIDIA ConnectX-6 100 Gb/s HDR100 InfiniBand/VPI Adapter, 1x QSFP56, PCIe 4.0 x16
  • NVIDIA ConnectX-6 Dx EN 200 Gb/s Ethernet Adapter, QSFP56, PCIe 4.0 x16
  • NVIDIA ConnectX-6 Dx EN 100 Gb/s Ethernet Adapter, QSFP56, PCIe 4.0 x16
  • NVIDIA ConnectX-5 EN 100 Gb/s Ethernet Adapter, QSFP28, PCIe 3.0 x16

Power

4 hot-swappable 2000 watt 80 PLUS Titanium PSUs...

  • 4 hot-swappable 2000 watt 80 PLUS Titanium PSUs
  • 2 + 2 redundancy
AC input:
  • 220-240 Vac / 10-9.8A / 50-60 Hz
Max output per PSU:
  • 220-240Vac / 2000 watts

Front I/O

Power button...

  • Power button
  • Reset button
  • Power LED
  • ID button with LED
  • Information LED
  • Power supply failure LED
  • 2 LAN activity LEDs, one for each RJ45 1
  • Gbps BASE-T LAN port
    Storage drive activity LED

Rear I/O

VGA port...

  • VGA port
  • 2 USB 3.0 ports
  • 2 RJ45 10 Gbps BASE-T LAN ports
  • 1 RJ45 1 Gbps BASE-T LAN out-of-band management port

Accessories

Rackmounting kit...

  • Rackmounting kit
  • 4 configurable C19 power cables

Physical

Form factor: 4U rackmount...

  • Form factor: 4U rackmount
  • Width: 17.2 inches (437 mm)
  • Height: 7.0 inches (179 mm)
  • Depth: 29.0 inches (737 mm)
  • Server weight: 73 lbs (33 kg)
  • Rackmounting kit weight: 5 lbs (2.3 kg)
  • Total weight with packaging: 99 lbs (45 kg)

Scalar datasheet

Download the Scalar Datasheet