Skip to content

Feature - New VM sizes#71

Merged
JaimieWi merged 8 commits into
mainfrom
feature/new_vm_sizes
Jun 23, 2026
Merged

Feature - New VM sizes#71
JaimieWi merged 8 commits into
mainfrom
feature/new_vm_sizes

Conversation

@JaimieWi

@JaimieWi JaimieWi commented Jun 19, 2026

Copy link
Copy Markdown
Contributor

Run ID: e20aee7f

What is being addressed

Introduce new VM sizes to aid in the battle against compute quota in Azure

How is this addressed

Current offering

VM Size vCPU RAM (GB) GPU GPU Memory (GB) Price/Hour Azure SKU
CPU-Only VMs
B-Series 2 4 None - £0.05 Standard_B2s
Dsv5 Series 2 8 None - £0.15 Standard_D2s_v5
D4as_v5 Series 4 16 None - £0.30 Standard_D4as_v5
Dsv5 Series 4 16 None - £0.30 Standard_D4s_v5
F4s_v2 Series 4 16 None - £0.30 Standard_F4s_v2
Dsv5 Series 8 32 None - £0.65 Standard_D8s_v5
Easv4 Series 8 64 None - £0.75 Standard_E8as_v4
Dsv5 Series 16 64 None - £1.25 Standard_D16s_v5
GPU VMs - A10 Series
NV6ads_A10_v5 Series 6 55 1/6 A10 4 £0.61 Standard_NV6ads_A10_v5
NV12ads_A10_v5 Series 12 110 1/3 A10 8 £1.22 Standard_NV12ads_A10_v5
NV18ads_A10_v5 Series 18 220 1/2 A10 12 £2.05 Standard_NV18ads_A10_v5
NV36ads_A10_v5 Series 36 440 1 A10 24 £4.11 Standard_NV36ads_A10_v5
NV72ads_A10_v5 Series 72 880 2 A10 48 £8.60 Standard_NV72ads_A10_v5
GPU VMs - H100 Series
NC40ads_H100_v5 Series 40 320 1 H100 NVL 94 £7.70 Standard_NC40ads_H100_v5
NC80adis_H100_v5 Series 80 640 2 H100 NVL 188 £15.90 Standard_NC80adis_H100_v5
  • New CPU quota will be offered for standard compute
  • GPU will be explored

Notes during development

Fasv6 (Standard_F*as_v6) sizes have been removed from the VM size options. The Fasv6 series requires NVMe disk controller support in the OS image. The current OUH images (Windows 10, Windows 11, and the OUH Server 2019 Data Science VM) were built with SCSI disk controllers and are not compatible, causing a 400 InvalidParameter error on deployment.

Expands the CPU VM size options available to researchers across both the Windows and Linux VM bundles.

The Dasv5 series (AMD EPYC) has been extended to cover 4, 8, 16, and 32 CPU sizes, providing an AMD-based alternative across the full size range and spreading quota load away from the Intel Dsv5 series.

Issues

Current GPU quota position in UK South:

A10 (NVadsA10_v5): Quota exhausted. No new A10 VMs can be deployed until quota frees up or is increased.
T4 (NCasT4_v3): Quota request unsuccessful.
H100 (NCadsH100_v5): Available. Retained in the bundle for essential workloads only due to cost (£7.70–£15.90/hr).
NVadsV710 (AMD ROCm): Available but requires unique VM image and set up and pre-qualification of researcher workload; not suitable as a general GPU offering.

Updates

A10 quota request — outcome

  • A10 (NVadsA10_v5): approved — 350 vCPUs granted in UK South. This should give us meaningful headroom for researcher GPU workloads.
  • T4 (NCasT4_v3): partially approved — Microsoft can support a maximum of 50 cores in UK South. We have decided to proceed with this and will add the 4 CPU (NC4as_T4_v3) and 8 CPU (NC8as_T4_v3) sizes to the bundle. Drivers will be brought in via airlock by researchers as needed.
  • Fsv2 quota request: declined — Microsoft advised that Fsv2 and other older series (F, Fs, B, Av2, Lsv2 etc.) do not run on their latest generation infrastructure and are directing customers to newer equivalents. The direct successor would be Fasv6, however this requires NVMe disk controller support which our current OUH images do not provide. Ddsv5 has the same constraint. Until images are rebuilt with NVMe support, Dasv5 is the closest viable alternative for quota spreading and compute-heavy workloads. Image rebuild is not currently planned.

Next actions:

  • Confirm A10 and T4 vCPU limits are reflected in the subscription
  • Add T4 sizes to Windows and Linux porter bundles and template schemas
  • Monitor GPU VM utilisation as researchers deploy

New VM sizes

VM Size vCPU RAM (GB) GPU GPU Memory (GB) Windows Price/hr Linux Price/hr Azure SKU
CPU-Only — B-Series
B-Series 2 4 None - £0.05 £0.05 Standard_B2s
CPU-Only — Dsv5 Series (Intel)
Dsv5 Series 2 8 None - £0.15 £0.10 Standard_D2s_v5
Dsv5 Series 4 16 None - £0.30 £0.20 Standard_D4s_v5
Dsv5 Series 8 32 None - £0.65 £0.35 Standard_D8s_v5
Dsv5 Series 16 64 None - £1.25 £0.65 Standard_D16s_v5
CPU-Only — Das_v5 Series (AMD)
Das_v5 Series 4 16 None - £0.30 £0.15 Standard_D4as_v5
Das_v5 Series 8 32 None - £0.60 £0.30 Standard_D8as_v5
Das_v5 Series 16 64 None - £1.20 £0.60 Standard_D16as_v5
Das_v5 Series 32 128 None - £2.40 £1.20 Standard_D32as_v5
CPU-Only — Fs_v2 Series (compute-optimised)
Fs_v2 Series 4 8 None - £0.30 £0.15 Standard_F4s_v2
Fs_v2 Series 8 16 None - £0.60 £0.30 Standard_F8s_v2
Fs_v2 Series 16 32 None - £1.15 £0.60 Standard_F16s_v2
CPU-Only — Easv4 Series (memory-optimised)
Easv4 Series 8 64 None - £0.75 £0.45 Standard_E8as_v4
Easv4 Series 16 128 None - £1.50 £0.90 Standard_E16as_v4
GPU — T4 Series
NCas_T4_v3 Series 4 28 1× T4 16 £0.60 £0.50 Standard_NC4as_T4_v3
NCas_T4_v3 Series 8 56 1× T4 16 £1.00 £0.70 Standard_NC8as_T4_v3
GPU — A10 Series
NV6ads_A10_v5 Series 6 55 1/6 A10 4 £0.61 £0.45 Standard_NV6ads_A10_v5
NV12ads_A10_v5 Series 12 110 1/3 A10 8 £1.22 £0.85 Standard_NV12ads_A10_v5
NV18ads_A10_v5 Series 18 220 1/2 A10 12 £2.05 £1.50 Standard_NV18ads_A10_v5
NV36ads_A10_v5 Series 36 440 1× A10 24 £4.11 £3.00 Standard_NV36ads_A10_v5
NV72ads_A10_v5 Series 72 880 2× A10 48 £8.60 £6.15 Standard_NV72ads_A10_v5
GPU — H100 Series
NC40ads_H100_v5 Series 40 320 1× H100 NVL 94 £7.70 £6.55 Standard_NC40ads_H100_v5
NC80adis_H100_v5 Series 80 640 2× H100 NVL 188 £15.90 £13.15 Standard_NC80adis_H100_v5

@JaimieWi

Copy link
Copy Markdown
Contributor Author
Screenshot 2026-06-19 at 12 48 55

@GamblingFool GamblingFool left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Great work on this PR!

The updates are well-structured and clearly address the issues outlined in the previous discussions. I appreciate the attention to detail and the thorough testing that has been conducted.

@GamblingFool GamblingFool left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Great work on this PR!

The updates are well-structured and clearly address the issues outlined in the previous discussions. I appreciate the attention to detail and the thorough testing that has been conducted.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants