VM Size Guide¶
This guide describes the VM types available in the TVS SDE and provides guidance on selecting the right size for your workload.
General principles¶
- Start small and resize up if needed. VMs can be resized within the same series without being deleted (e.g. Dsv5 2 CPU → Dsv5 4 CPU). Resizing across series requires deleting and redeploying the VM.
- GPU VMs consume shared quota across all workspaces on the platform. Quota is only released when a VM is deleted, not when it is stopped. Please delete GPU VMs as soon as your work is complete.
- If you are unsure which size to choose, the Dsv5 2 CPU / 8GB RAM is a reasonable default starting point for most workloads.
CPU-only VMs¶
B-Series — Standard_B2s¶
| vCPU | 2 |
| RAM | 4 GB |
| Windows price | £0.05/hr |
| Linux price | £0.05/hr |
The B-Series is a burstable VM type, meaning it accumulates CPU credits during periods of low activity and spends them during bursts of higher demand. The baseline CPU performance is low.
Use when: You need a low-cost VM for light tasks such as text editing, reviewing documents, or running the portal. Not suitable for data analysis, statistical computing, or any sustained CPU workload.
Do not use when: You are running R, Python, or any analytical workload — even a moderately sized script will exhaust CPU credits quickly and the VM will throttle significantly.
Dsv5 Series (Intel) — Standard_D2s_v5 to Standard_D16s_v5¶
| Size | vCPU | RAM | Windows price | Linux price |
|---|---|---|---|---|
| Dsv5 2 CPU | 2 | 8 GB | £0.15/hr | £0.10/hr |
| Dsv5 4 CPU | 4 | 16 GB | £0.30/hr | £0.20/hr |
| Dsv5 8 CPU | 8 | 32 GB | £0.65/hr | £0.35/hr |
| Dsv5 16 CPU | 16 | 64 GB | £1.25/hr | £0.65/hr |
The Dsv5 series is a general-purpose Intel-based VM with a balanced CPU-to-memory ratio (8 GB per vCPU). It supports Premium SSD storage and is suitable for most research workloads.
Use when: Running R or Python scripts, exploring datasets, fitting statistical models, or doing general-purpose data analysis. The 2 CPU size is a good default for most researchers. Step up to 4 or 8 CPU if your workload is slow or if you are parallelising across cores.
Use the 8 or 16 CPU size when: Running parallel processing (e.g. parallel, foreach in R, or multiprocessing in Python), large model fitting, or working with datasets that require more memory than the smaller sizes provide.
Resizable within series: Yes — you can resize from Dsv5 2 CPU to Dsv5 4 CPU without deleting the VM.
Das_v5 Series (AMD) — Standard_D4as_v5 to Standard_D32as_v5¶
| Size | vCPU | RAM | Windows price | Linux price |
|---|---|---|---|---|
| Das_v5 4 CPU | 4 | 16 GB | £0.30/hr | £0.15/hr |
| Das_v5 8 CPU | 8 | 32 GB | £0.60/hr | £0.30/hr |
| Das_v5 16 CPU | 16 | 64 GB | £1.20/hr | £0.60/hr |
| Das_v5 32 CPU | 32 | 128 GB | £2.40/hr | £1.20/hr |
The Das_v5 series is AMD-based and otherwise comparable to the Dsv5 in terms of workload suitability. It offers the same balanced CPU-to-memory ratio and Premium SSD support. In practice, performance differences between AMD and Intel at this generation are minimal for typical research workloads.
Use when: Your workload is the same as Dsv5 but Dsv5 quota is constrained, or when you need more than 16 CPU cores without moving to a GPU VM. The 32 CPU / 128 GB size is the largest CPU-only option available and is suitable for very large parallel workloads or large in-memory datasets.
Use the 32 CPU size when: You are running highly parallelised workloads (e.g. large bootstrapping runs, genome-wide association studies, simulation models) and need more cores than the Dsv5 16 CPU provides.
Resizable within series: Yes — you can resize within the Das_v5 series without deleting the VM.
Fs_v2 Series (compute-optimised) — Standard_F4s_v2 to Standard_F16s_v2¶
| Size | vCPU | RAM | Windows price | Linux price |
|---|---|---|---|---|
| Fs_v2 4 CPU | 4 | 8 GB | £0.30/hr | £0.15/hr |
| Fs_v2 8 CPU | 8 | 16 GB | £0.60/hr | £0.30/hr |
| Fs_v2 16 CPU | 16 | 32 GB | £1.15/hr | £0.60/hr |
The Fs_v2 series is compute-optimised with a lower memory ratio than the D-series (2 GB per vCPU vs 8 GB per vCPU). It offers higher CPU clock speeds relative to its price point, making it suited to workloads that are CPU-bound rather than memory-bound.
Use when: Your workload makes heavy use of CPU but does not require large amounts of RAM — for example, text processing, certain simulation workloads, or per-vCPU licensed software where you want to minimise the number of cores while maximising CPU performance.
Do not use when: Your workload loads large datasets into memory, fits large models, or requires more than 2 GB RAM per CPU. In those cases the Dsv5 or Das_v5 will serve you better.
Resizable within series: Yes — you can resize within the Fs_v2 series without deleting the VM.
Easv4 Series (memory-optimised) — Standard_E8as_v4 to Standard_E16as_v4¶
| Size | vCPU | RAM | Windows price | Linux price |
|---|---|---|---|---|
| Easv4 8 CPU | 8 | 64 GB | £0.75/hr | £0.45/hr |
| Easv4 16 CPU | 16 | 128 GB | £1.50/hr | £0.90/hr |
The Easv4 series is memory-optimised with a high memory-to-CPU ratio (8 GB per vCPU). It is AMD-based and supports Premium SSD storage. It is suited to workloads where the bottleneck is available RAM rather than CPU speed.
Use when: You are working with very large datasets that need to be held in memory simultaneously, running in-memory databases, working with large dataframes in R (data.table, dplyr on large data), or fitting models that require substantial RAM (e.g. large mixed-effects models, certain Bayesian methods).
Do not use when: Your workload is primarily CPU-bound with moderate memory requirements — in that case the Dsv5 or Das_v5 will give you more CPU for the same price.
Resizable within series: Yes — you can resize within the Easv4 series without deleting the VM.
GPU VMs¶
GPU VMs consume shared quota across all workspaces. Quota is only released when a VM is deleted, not when it is stopped. Please delete your GPU VM as soon as your work is complete.
When do you actually need a GPU?¶
Not all workloads benefit from a GPU. Before requesting a GPU VM, consider:
- Training deep learning models (neural networks, transformers, CNNs) — GPU will give substantial speedup.
- Running inference on large models (e.g. LLMs, image classification) — GPU recommended.
- GPU-accelerated visualisation (e.g. 3D rendering, certain GIS workloads) — GPU required.
- Standard statistical analysis, regression, survival models, mixed-effects models — CPU is sufficient; a GPU will not help.
- R and Python scripting without explicit GPU libraries — CPU is sufficient.
If you are unsure, start with a CPU VM. If your workload is taking hours and you are using a GPU library (PyTorch, TensorFlow, CUDA), then a GPU VM is likely worthwhile.
T4 Series — Standard_NC4as_T4_v3, Standard_NC8as_T4_v3¶
| Size | vCPU | RAM | GPU | GPU Memory | Windows price | Linux price |
|---|---|---|---|---|---|---|
| NCas_T4_v3 4 CPU | 4 | 28 GB | 1× T4 | 16 GB | £0.60/hr | £0.46/hr |
| NCas_T4_v3 8 CPU | 8 | 56 GB | 1× T4 | 16 GB | £0.93/hr | £0.66/hr |
The T4 is an NVIDIA Tesla T4 GPU with 16 GB VRAM. It is a mid-range data centre GPU well suited to inference workloads and moderate training tasks. It supports CUDA and is compatible with PyTorch and TensorFlow.
Use when: Running inference on small to medium models, fine-tuning smaller language models, GPU-accelerated data processing, or training moderately sized neural networks. A good cost-effective entry point for GPU workloads.
Do not use when: Training large models that require more than 16 GB VRAM, or running very large batch inference — in those cases consider the A10 or H100.
Note: NVIDIA drivers must be airlocked in by the researcher. GPU quota in UK South is limited to 50 vCPUs across this series.
A10 Series — Standard_NV6ads_A10_v5 to Standard_NV72ads_A10_v5¶
| Size | vCPU | RAM | GPU | GPU Memory | Windows price | Linux price |
|---|---|---|---|---|---|---|
| NV6ads A10 | 6 | 55 GB | 1/6 A10 | 4 GB | £0.61/hr | £0.45/hr |
| NV12ads A10 | 12 | 110 GB | 1/3 A10 | 8 GB | £1.22/hr | £0.85/hr |
| NV18ads A10 | 18 | 220 GB | 1/2 A10 | 12 GB | £2.05/hr | £1.50/hr |
| NV36ads A10 | 36 | 440 GB | 1× A10 | 24 GB | £4.11/hr | £3.00/hr |
| NV72ads A10 | 72 | 880 GB | 2× A10 | 48 GB | £8.60/hr | £6.15/hr |
The A10 is an NVIDIA A10 GPU. The smaller sizes (NV6ads, NV12ads, NV18ads) are GPU partitions — fractions of a physical A10 sharing the underlying hardware. The NV36ads gives you a full A10 with 24 GB VRAM, and the NV72ads gives two full A10s.
Use when: GPU-accelerated ML inference, visualisation workloads, or training small to medium neural networks. The partitioned sizes (4–12 GB VRAM) are suitable for inference and light training. The full A10 (24 GB) is suitable for training medium-sized models or running larger inference workloads.
Choosing a partition size: Match VRAM to your model size. As a rough guide — models up to ~1B parameters can typically fit in 4–8 GB; models up to ~7B parameters with quantisation may fit in 12–24 GB.
Note: A10 quota is shared across all workspaces (350 vCPU total). Delete your VM when work is complete.
H100 Series — Standard_NC40ads_H100_v5, Standard_NC80adis_H100_v5¶
| Size | vCPU | RAM | GPU | GPU Memory | Windows price | Linux price |
|---|---|---|---|---|---|---|
| NC40ads H100 | 40 | 320 GB | 1× H100 NVL | 94 GB | £7.70/hr | £6.55/hr |
| NC80adis H100 | 80 | 640 GB | 2× H100 NVL | 188 GB | £15.90/hr | £13.15/hr |
The H100 NVL is NVIDIA's highest-performance data centre GPU currently available on the platform, with 94 GB VRAM per card. It is significantly faster than the A10 for training workloads and supports the latest CUDA features.
Use when: Training large models (>7B parameters), running large-scale inference, or any workload where A10 performance or VRAM is insufficient. H100 should be considered a last resort after confirming the A10 cannot meet your requirements.
H100 quota is extremely limited. Deployment requires justification and should be for the minimum period necessary. Delete this VM immediately when your work is complete.
Note: The OUH Server 2019 image is not currently compatible with H100 VMs. Use Windows 10, Windows 11, or the OUH Ubuntu image instead.