Overview
Fluidstack Atlas provides high-performance compute instances designed to meet the needs of inference and fine-tuning workloads. All compute instances are either bare-metal servers or highly tuned kvm virtual machines.
Instance Types
Fluidstack Atlas supports a variety of CPU and GPU-enabled instance types.
HGX Instance Types
Accelerate multi-GPU inference and training with 8 NVIDIA GPUs connected by a high-bandwidth NVLink fabric. All HGX instances include balanced CPU, RAM, and front-end network bandwidth for fast data loading, pre-processing, and checkpointing. For more specific details about the compute capabilities of the NVIDIA HGX platform, check out the official docs here.
| Instance Type | Description |
|---|---|
| b300.8x | 8 NVIDIA B300 GPUs with 288 GB memory each. |
| b200.8x | 8 NVIDIA B200 GPUs with 192 GB memory each. |
| h200.8x | 8 NVIDIA H200 GPUs with 141 GB memory each. |
| h100.8x | 8 NVIDIA H100 GPUs with 80 GB memory each. |
| a100-hgx-80gb.8x | 8 NVIDIA A100 GPUs with 80 GB memory each. |
CPU Instance Types
Use small CPU-only instance types as a bastion hosts or control plane node or application specific general purpose compute.
| Instance Type | Description |
|---|---|
| cpu.2x | 2 CPUs and 4 GiB RAM |
| cpu.4x | 4 CPUs and 8 GiB RAM |
| cpu.8x | 8 CPUs and 16 GiB RAM |