Microsoft gives Hyper-V ceilings a Herculean hike
- Reference: 1712928429
- News link: https://www.theregister.co.uk/2024/04/12/hyperv_maximums_windows_server_2005/
- Source link:
"Please read carefully. These are not typos," [1]wrote Jeff Woolsey, Microsoft's principal program manager for Azure Stack HCI, Windows Server and hybrid cloud.
That warning came in the first of a series of Xeets in which Woolsey revealed that VMs under Hyper-V in Windows Server 2025 can run 2,048 virtual CPUs and 240 TB of memory. A Microsoft [2]article adds a little more detail: each VM can support up to 256 SCSI disks, four SCSI controllers, and virtual disks can reach 64 TB apiece. You can also have 68 virtual network adapters, alongside four virtual fiber channel adapters.
[3]
Tragically, you'll have to make do with a single virtual floppy drive.
[4]
[5]
Hosts for Hyper-V reach even greater heights – with 4 PB of memory permitted.
Clusters can scale to 64 nodes and host 8,000 VMs.
[6]
Most of the above numbers stay the same under Windows Server 2022, other than the vCPU count, which drops to a mere 1,024. The 240 TB memory ceiling is a big lift for Windows Server 2022.
Some of the large numbers above can only be reached with Gen-2 VMs, which were designed to scale higher and include Trusted Platform Modules and the Unified Extensible Firmware Interface. Gen-2 VMs will become the default under Windows Server 2025.
[7]March Patch Tuesday sees Hyper-V join the guest-host escape club
[8]Microsoft veteran on how to blue screen your way to better testing
[9]Microsoft has made Azure Linux generally available. Repeat, Azure Linux
[10]Oracle VirtualBox 7.0 is here – just watch out for the proprietary Extension Pack
Microsoft's article in which the maximums are revealed instructs users to "plan for Hyper-V scalability in Windows Server" – an imperative that reflects the fact Windows Server 2025 is currently in preview.
But news that Microsoft-managed VMs can scale higher will be welcome for a couple of reasons.
One is that some workloads need monster VMs. SAP can require tens of terabytes of RAM. AI training workloads will need truckloads of storage and memory. It just makes sense for Microsoft to make sure its platform can handle those requirements.
[11]
The other reason is that Broadcom's acquisition of VMware has given virtualization shops a good reason to consider their options. While Hyper-V and its adjacent tools – like the System Center Virtual Machine Manager – are generally rated inferior to VMware's stack, Broadcom's new licensing plans have angered many users. And it's a rare IT operation that doesn't have some entitlement to Windows Server and therefore Hyper-V.
Woolsey also points out that Hyper-V is also present in Xbox – the gaming console uses it to isolate games and other workloads. The Register mentions this as idle speculation about the kind of entertainment experience that would become possible if Xboxes ever scaled anywhere near the hosts Hyper-V can now inhabit – with the added essential speculation about whether it could run Crysis. ®
Get our [12]Tech Resources
[1] https://twitter.com/WSV_GUY/status/1778373652611821889
[2] https://learn.microsoft.com/en-us/windows-server/virtualization/hyper-v/plan/plan-hyper-v-scalability-in-windows-server?pivots=windows-server-2025
[3] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/virtualization&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2ZhlamiDzx9FxbJkKV2C7rAAAAEQ&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[4] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/virtualization&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44ZhlamiDzx9FxbJkKV2C7rAAAAEQ&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[5] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/virtualization&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33ZhlamiDzx9FxbJkKV2C7rAAAAEQ&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[6] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/virtualization&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44ZhlamiDzx9FxbJkKV2C7rAAAAEQ&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[7] https://www.theregister.com/2024/03/13/patch_tuesday_march_2024/
[8] https://www.theregister.com/2024/02/21/raymond_chen_windows_notmyfault/
[9] https://www.theregister.com/2023/05/26/microsoft_azure_linux_container/
[10] https://www.theregister.com/2022/10/11/oracle_virtualbox_7/
[11] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/virtualization&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33ZhlamiDzx9FxbJkKV2C7rAAAAEQ&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[12] https://whitepapers.theregister.com/
I was about to post the same point!
The main benefit of virtualisation is being able to run multiple VMs on a host. If that level of power is required then it's probably going to be the only VM on there, so why not just cut out the virutalisation layer?
Because "stuff" can share that bare metal hardware. The over head of Hyper-V is minimal anyway. It was always one of the better (if not the best) scaling hypervisors for IOs vs bare metal.
If there's any capacity left to share, with a VM that size running, then it's going to be a bloody powerful piece of hardware!
Windows 11 doesn't run on bare metal, when you install it on your computer it installs a lite version of hyper-v as the base with Windows 11 running on top of that. Windows 10 on compatible machines does the same. On any modern system you are already running Windows in a VM.
OK, to an extent that's true but adding the full set of hypervisor services for user-managed VMs seems pointless at the sort of power level being discussed here. Bare metal (or as close to it as possible within the design of the OS) would normally make more sense.
You could have multiple VMs with just only one running at any given time, say you are renting the system out. Each VM would have it's own virtual TPM which would mean you don't need to wipe the system's TPM every time a different customer rents it.
If you are renting out capacity for that sort of VM, you are a major cloudy provider and probably aren't going to be hosting on Hyper-V.
Not quite. IF you enable Virtualisation Based Security and Hypervisor-Enforced Code Integrity (HVCI) then the Kernel runs "like" a VM but it's not Hyper-V and most of Windows is not "in a VM".
Correct response, I was just about to write the same. But why as AC?
A fair question, but the answer is 'portability'. The only glimpse the guest OS has of the real hardware is the CPU make & model. The rest is all an abstraction, so at any point in the future you can shift to a new/different system (even with wildly different hardware) and the VM will be blissfully unaware. You don't need to hope and pray that 1) the OS doesn't complain about X and Y changing, or 2) the admin updating config items doesn't miss one.
Because at some point in the future, this will likely be considered small. It was within living memory that 1MB of RAM was considered massive, or a 10MB HDD was considered to be like having a McMansion of storage.
So, instead of making small tweaks every couple of years, they just did about a decades worth in one go.
Mobility, independency from the hardware layer.
Intel network cards making problems? Change them. Don't need to change a thing in the host. Including the hardware-fingerprint activation of some special software.
it might be able to run Vista then
So it's a cloud option, right ?
Because there ain't no motherboard that can handle that level of hardware.
I shudder to think of the monthly cost for someone who actually signs up for 2000 CPUs and 240TB of RAM. Some beancounter somewhere is going to have a heart attack.
Re: So it's a cloud option, right ?
If you were creating a monster VM like that in Azure, it would be directly in Azure itself, not within a Server 2025 VM running on Azure.
Re: So it's a cloud option, right ?
There is no "in Azure itself", it is all Hyper-V there. Has been for a long time, including the machines where you can get 20% or 25% of an installed GFX card. Even if it is only one VM and it takes the whole machine, it is still Hyper-V.
Re: So it's a cloud option, right ?
I think you missed the point. I am well aware that Azure runs on a customised version of Hyper-V. My point was that this news article is about Hyper-V capacities within Server 2025. If you were running a monster VM like that on Azure you would be running it directly on Azure (which has been able to handle monster VMs for some time), not using nested virtualisation within Server 2025.
Re: So it's a cloud option, right ?
The cloud stuff is mostly optional after quite some complaints about "why is OneDrive, copilot and Azure-Arc installed and active by default? Available fine, but please not active!" complains, with a few (including me) pointing to trust issues on those.
"VMs under Hyper-V in Windows Server 2025 can run 2,048 virtual CPUs and 240 TB of memory."
Today the limit of Hyper-V. Tomorrow: minimum recommended spec for Windows 12.
I'll see myself out.
This is Microsoft's way of trying to remind people they still make a hypervisor, same way they beg you not to download Chrome when it's the only thing you use IE/Edge for. Most well-heeled Microsoft-y admins pay VMWare today for *good* virtualization, but with Broadcom pissing on everyone now, they need another easy button since being allergic to Linux usually. Anyone that isn't afraid of Linux will just use KVM or KVM-based solutions.
Immune
And of course Hyper -V never has any vulnerabilities that would require a restart after patching....
Re: Immune
There is no hypervisor which did not have a security issue. But check the CVE statistics. Hyper-V has less problems than VMware. As for the other hypervisor based virtualization beside VMware und Hyper-V: Where it the link to the Reg article with the statistics again....
“It just makes sense for Microsoft to make sure its platform can handle those requirements.”
We know what their normal QA is like. What do you suppose the QA on a 240TB RAM VM was? “Good news! It booted!”
How many systems that size could MS possibly own - 2 or 3?
That's an awful lot when you consider 640K is enough for anyone.
You have to direct that to IBM not Microsoft. They created that limit when designing those machines and setting the de-facto standard, for example which memory area is mapped to video and so on.
For SMEs, is the windows server cost per core having much of an effect on the server hardware that they buy?
i.e.buying dual 8 core CPUs rather than dual 24 cores, to avoid +10K in per core Datacenter licenses?
AFAIK it changed to CPU core count, so one fat 16 core is as expensive and two 8 core. And in quite some cases a 16 core CPU is faster than a dual 8 core.
But be aware: That might have already changed again....
It was Server 2012 or 2012 R2 (can't remember which) which moved server licenses from per physical box to per physical CPU. From Server 2016 onwards it's been per physical CPU core (with a minimum of 16 cores as I recall).
Maybe a dumb question... If you're so concerned about performance that you need two thousand CPUs and a couple of hundred TB of RAM wouldn't you be running on bare metal anyway, why is this necessary?