Premium
GPU Instances

Up to 8x on-demand NVIDIA® HGX B200 180GB GPUs using Blackwell Tensor Core technology.

This is the latest hardware from NVIDIA available on the market, purpose-built to accelerate inference for LLMs and MoE models.

We deploy the HGX platform with NVLink interconnect, offering up to 1.8TB/s P2P NVLink bandwidth.

Up to 8 NVIDIA® H200 141GB GPUs, each containing 16896 CUDA cores and 528 Tensor Cores.

We only deploy the H200 in the SXM5 form-factor, which offers a memory bandwidth of 4.8TB/s and up to 900GB/s P2P bandwidth.

The servers utilize the fourth generation AMD Genoa, 384 threads with a boost clock of 3.7GHz, offering a substantial uplift in performance vs. alternatives.

The 8H200 VM has the same number of vCPUs as the 4H200, but delivers higher CPU performance by utilizing only the physical cores, rather than the logical processors (hyper-threads).

Up to 8 NVIDIA® H100 80GB GPUs, each containing 16896 CUDA cores and 528 Tensor Cores.

We deploy the SXM5 NVLINK module, which offers a memory bandwidth of 3TB/s and up to 900GB/s P2P bandwidth.

Fourth generation AMD Genoa, up to 384 threads with a boost clock of 3.7GHz.

The name 8H100.80S.176V is composed as follows: 8x H100 SMX5, 176 CPU core threads & virtualized.

The 8H100 VM has the same number of vCPUs as the 4H100, but delivers higher CPU performance by utilizing only the physical cores, rather than the logical processors (hyper-threads).

Up to 8 NVIDIA® A100 80GB GPUs, each containing 6912 CUDA cores and 432 Tensor Cores.

We only use the SXM4 'for NVLINK' module, which offers a memory bandwidth of over 2TB/s and Up to 600GB/s P2P bandwidth.

Second generation AMD EPYC Rome, up to 192 threads with a boost clock of 3.3GHz.

The name 8A100.176V is composed as follows: 8x RTX A100, 176 CPU core threads & virtualized.

Up to 8 NVIDIA® L40S 48GB GPUs, each containing 18176 CUDA cores and 568 fourth-generation Tensor Cores.

Featuring 864GB/s memory bandwidth, 1466 Tflops Tensor performance and 212 Tflops RT Core performance, L40S boasts excellent capability for handling large ML models, single-GPU training, and other 32-16-8 bit operations.

The name 8L40S.160V is composed as follows: 8x L40S, 160 CPU core threads & virtualized.

Up to 8 NVIDIA® RTX6000 ADA [2022] GPUs, each containing 18176 CUDA cores, 568 fourth-generation Tensor Cores and 142RT cores.

Features 48GB of GDDR6 memory for working with the largest 3D models, render images, simulation and AI datasets.

The name 8RTX6000ADA.80V is composed as follows: 8x RTX6000ADA, 80 CPU core threads & virtualized.

Up to 8 NVIDIA® RTX A6000 [2021] GPUs, each containing 10752 CUDA cores, 336 Tensor Cores and 84RT cores.

Despite having less tensor cores than the V100, it is able to process tensor operations faster due to a different architecture.

Second generation AMD EPYC Rome, up to 96 threads with a boost clock of 3.35GHz.

PCIe Gen4 for faster interactions between GPUs.

The name 8A6000.80V is composed as follows: 8x RTX A6000, 80 CPU core threads & virtualized.

Up to 8 NVIDIA® Tesla V100 GPUs, each containing 5120 CUDA cores and 640 Tensor Cores.

Second generation Xeon Scalable 4214R CPUs [2020], up to 48 threads with a boost clock of 3.5GHz.

NVLink for high bandwidth P2P communication.

The name 4V100.20V is composed as follows: 4x V100, 20 CPU core threads & virtualized.

Second or Third generation AMD EPYC Rome or Milan.

All hardware is dedicated to your server for the best performance.

The name CPU.32V indicates the server runs on 32 virtualized core threads.