This is what happens when AI generates your PR – FAQ. Describe what the product does versus just a data dump.

I’m usually in agreement with this sentiment, but they do say what it is in the first paragraph.

“NVIDIA Kubernetes® Engine (NVKE) provides the management plane that platform teams install to create and operate Kubernetes clusters for their customers.”

Ok, so say I have a platform team to install k8s or k3s. Why do I need NVKE? Also, how do I resell my kubernetes to customers? Like is there stripe integration?

I am more confused by its description than by just reading the product name and guessing.

That doesn’t say anything. It doesn’t answer a single question I would have.

maybe you're not the user

I am exactly the potential user of this, as I am currently setting up a small collocated cluster and investigating software for management. But this doesn’t tell me anything that would actually matter in that decision.

Are you planning on building an AI factory ?

Well one needs to read a bit more to understand what that meant...

NVKE is not a Kubernetes distribution and not a Cluster API flavour. It is a self-hosting factory for Kubernetes clusters: a throwaway seed cluster stands up a management cluster, and the management plane on it then provisions and lifecycle-manages many isolated workload clusters on demand.

A mention of something Kubernetes related reached the front page! A few years ago, BAI, there was one or more k8s headlines gracing HN's front page at all times.

Kinda glad that ended. k8s is fine and all, but it was all a bit much for a while there.

Did you just make an acronym that means "before AI"?

Yes, but apparently I didn't invent it. I thought maybe I had, so I checked, and no, I'm not the first: it's been bouncing around informally among academics for a while now. I hadn't previously seen it though, so I can at least say I came up with is independently.

Future history books will be like: ancient legacy technology conglomerate Google was founded around 24BAI

pretty sure "before wwiii" will take precedence

I would guess throwing around the word "AI" constantly and endlessly debating p(doom) will be considered pretty amusing like reading essays from the late 19th century that can't stop pontificating about living in a boundless technological singularity thanks to railroads and the telegraph. Current hackocalypse discourse already has shades of "the bomber always gets through" from the 20s.

I can imagine someone saying the same in the 60s: don't sign this nuclear test ban treaty, we have induction from 1300s that.. wait let's chose a time just after the Black Death. We have induction from the Renaissance that proves we always worry too much about things and nothing is different about this time.

Or something like 32BAGI

Can’t regular Kubernetes also deploy to GPUs?

Yes. https://kubernetes.io/docs/tasks/manage-gpus/scheduling-gpus...

I just got flashbacks of stuff I forgot long ago like OpenStack and EMC Rex-ray and so many things like that https://www.openstack.org https://github.com/rexray/rexray

[delayed]

Just deployed OpenStack by using the Kolla Ansible installer. Works great coupled to Ceph for storage. Thanks to Broadcom we are moving in the right direction. Proxmox as fallback plan but it looks like OpenStack is solid.

[delayed]

This is for tenant-isolated k8s clusters using NVIDIA's DSX reference stack, relying on NICo / BlueField to set up isolated networking, and wipes/re-attests each host between tenants. Mostly useful if you're chasing NVIDIA Cloud Partner status or need NVIDIA support across the whole stack.

Its great to see them open source this stack!

Lots of negativity from HN posters who usually manage 3 Kubernetes clusters in a single region via a managed service on AWS or GCP.

When I wrote a Kubernetes-compatible Rust service, I got a lot of hate too [1].

[1] https://github.com/debarshibasak/superkube

I got lot of hate too when I tried building a api server compatible engine in rust [1]

[1] https://github.com/debarshibasak/superkube

> Mostly useful if you're chasing NVIDIA Cloud Partner status or need NVIDIA support across the whole stack.

What if I don't care about the cloud partner status or support, but I still have a lot of Nvidia hardware I need to hand out to run isolated workloads? Seems like deploying straight k8s is the way to go?

I'm not super familiar with this tool, but I am a k8s administrator. I think it depends on how isolated you need those workloads to be. This seems like a good fit if you want cluster level isolation. There are other projects that implement the Cluster API (CAPI) though, which provide similar functionality, eg declaratively managing tenant clusters.

This seems to mix several (should-be) distinct concerns like CAPI, Kamaji and gpu-operator.

The only things I'd take from this is the kind-based bootstrapper for the management cluster and SPIFFE/TPM integration.

yeah i wish more visibility would go towards CAPI, Kamaji.

tbh i'm not sure from reading this why you wouldn't just roll rke2/k3s if the concern is an environment outside of the well worn hyperscaler managed k8s

This looks overly heavy and non-composable, locking in to one vendor on many levels. I would like to see tools which enrich existing infrastructure in a unix-way

Will this be contained in the new 2000$ GPU annual fee for Nvidia software or will it need a plus subscription?

https://www.tomshardware.com/pc-components/gpus/nvidia-discu...