hckrnws

The value proposition has never quite gotten clear. There's some places where things have gotten better, with opennic and various firmware expectations. The chassis and power supply specifications are pretty generally usable.

But in terms of actual computing hardware? Hell no. Meta and others keep releasing system designs every year or so, but they're deeply abstract designs. None of the hard engineering work is provided; it's largely abstract requirements that do practically nothing to ease system design. The hard work of building a motherboard still has to be done de-novo; it's just extra difficult for it has to meet difficult constraints too.

I still love and respect OCP. The Evenstar work contributed, for example, represents a very high flying state of the art cellular system that could should & ought totally revamp a large chunk of what we do, if we had any sense, but the conservative large-industry demand side isn't picking up this possibility. Theres so much interest & possibility in so many places. But almost never is the real good stuff given away or really detailed. Google giving away 75% of the work for a extreme efficiency DC-DC converter, STC (switched tank converter), is the best OpenCompute seems to be able to deliver, yet requires custom systems to use. More typically they just specify demands & provide no path & no help to getting there: we want this, build it, and in that case the competition remains deeply proprietary & guards it's offerings closely.

The hard work of building a motherboard still has to be done de-novo...

I think the idea is that you contact the same ODM that Meta is using and they reuse the existing design. Of course they usually don't tell you which ODM that is so you have to guess. Overall the selective openness is still pretty annoying.

> The chassis [snip] specifications are pretty generally usable.

IMHO it's a bit stupid they made the OCP chassis/rack spec incompatible with the standard EIA 19" rack (Yes, there's the Open19 stuff etc., but talking about the "main" OCP racks here). If you're going to break with the established standard, at least make something substantially different with scope for significant improvement. As it is, the OCP rack design is just subtly different from the 19" to be annoying but doesn't bring any huge improvements, fracturing the market for very little benefit.

Open Rack has totally different power distribution than industry-standard racks and it uses OUs instead of Us. How much more difference do you want?

Well standard racks don't really tell anything about how power distribution is handled. So why not make the OCP rack an extension of the EIA 19" rack? As it is, it's slightly wider, and as you mention the OU is slightly different from the EIA RU height. That would have enabled vendors to use the same sheet metal boxes, and MB dimensions etc. etc. for standard 19" gear and OCP gear. Or if you're going to be different, at least make it sufficiently different to be worthwhile. Is a 19-21" rack with ~RU/OU sized shelves really the optimal, or can we do something completely different that would be better (for some definition of better)?

Good stuff is happening here, such as security https://www.opencompute.org/wiki/Security and open source https://www.opencompute.org/projects/open-system-firmware

Microscaling formats for machine learning: https://www.opencompute.org/documents/ocp-microscaling-forma...

This is a big deal for increasing performance past fp8. We'll see Nvidia/AMD/Intel etc adopt this in their products in the next year.