Choosing an instance type
Overview
ECI offers three instance families — GPU, NPU, and CPU. Picking the type that matches your workload lets you keep costs in check.
Instance family comparison
| Family | Hardware | Primary use | Example codes |
|---|---|---|---|
| GPU | A100 PCIe, NVIDIA H100 SXM, B200 SXM | AI training and inference, parallel HPC | G-NHHS-640, G-NAHP-320, G-NBTHS-1440 |
| NPU | Rebellions ATOM Plus/Max, Furiosa Warboy/Renegade | AI inference only (low-power, efficient) | N-FR-48, N-FW-16, N-FW-32, N-FW-64 |
| CPU (general) | General vCore (memory = vCore × 2 GiB) | Data preprocessing, web servers, general | C-2, C-4, C-8, C-16, C-32 |
| CPU (memory-optimized) | General vCore (memory = vCore × 4 GiB) | Memory-heavy workloads (in-memory DB etc.) | M-2, M-4, M-8, M-16 |
Infrastructure > Instance Types lists the CPU, GPU, and memory specs for every type currently offered. Available types vary by zone, so some codes may only show up in specific zones.
Recommendations by workload
AI model training
Use a GPU instance. Pick based on training scale:
- Single GPU: small models and PoCs (e.g. 1 × A100)
- Multi-GPU on a single VM: medium-scale training (e.g. 8 × H100)
- Virtual cluster (InfiniBand): large-scale distributed training with tens of GPUs or more
AI inference servers
- GPU: large LLMs and high-throughput inference
- NPU: inference-only, more power-efficient than GPU. Supported devices: Rebellions ATOM Plus/Max, Furiosa Warboy/Renegade. Check each NPU's SDK and the list of compatible models before deciding
Data preprocessing / general servers
Use a CPU instance. Running tasks that don't need a GPU on a GPU instance is a waste of money.
Reading the instance type code
GPU and NPU codes follow the pattern {family}-{device code}-{total memory} (G- for GPU, N- for NPU).
G-NHHS-640: H100 SXM × 8
| Part | Meaning |
|---|---|
G | GPU family |
NHHS | NVIDIA H100 80GB SXM device |
640 | Total VRAM (GB): 80 GB × 8 |
The same NHHS series ranges from G-NHHS-80 (1 GPU) to G-NHHS-640 (8 GPUs).
G-NAHP-320: A100 PCIe × 4
| Part | Meaning |
|---|---|
G | GPU family |
NAHP | NVIDIA A100 80GB PCIe device |
320 | Total VRAM (GB): 80 GB × 4 |
G-NBTHS-1440: B200 SXM × 8
| Part | Meaning |
|---|---|
G | GPU family |
NBTHS | NVIDIA B200 180GB SXM device |
1440 | Total VRAM (GB): 180 GB × 8 |
The same NBTHS series ranges from G-NBTHS-180 (1 GPU) to G-NBTHS-1440 (8 GPUs).
C-{n} / M-{n}: CPU and memory
n is the vCore count. The CPU family provides vCore × 2 GiB of memory; the memory-optimized family provides vCore × 4 GiB.
Next steps
- Pricing model: on-demand vs reserved vs spot
- Virtual cluster: connecting multiple VMs over InfiniBand
- Operating spot VMs: how to cut costs