Why OpenCAPI is a declaration of interconnect fabric war An OpenCAPI consortium has sprung into life, promising a new, open specification that can increase data center server performance up to 10x through the use of a new CPU-memory-IO adapter interconnect scheme – and it doesn't include Intel in its membership. The Open Coherent Accelerator Processor Interface represents, the …
Not well thought out, so far, is that operating systems, application software, and all the other myriad bits that go into systems today have written into the assumptions about the underlying architecture. We're in for several iterations on differing paths before we converge on a new, possibly unified (perhaps hybrid) design. Whatever else, it should prove interesting to full stack geeks like me.
Sure but as far as the OS, whatever new may come, it needs to be compatible with all the existing ones. Else what chance can it seriously have? As far as application software, well if the OS is compatible...
The PCI fight has been going on since the late 90's (That's just what I know of it). I thought that IBM was part creator of PCI, but the moment it took off they seemed to walk away after the words "IBM compatible" were uttered. Again, I could be wrong as I never really followed this debate, but surely somebody can come up with something faster than 2 or 4 or whatever channels we have rammed into our way of thinking.
P.S. I only posted because I read Intel and Google weren't in full favor of this project, so it must be good!
This post has been deleted by its author
You and I must be ancient, Hans 1.
I remember installing CAPI drivers and configuring them in 1990s, and again oddly in 2006 for early 4G gear (dial-up connection + Native networking, so they used CAPI driver architecture. All mobile is a little bit like ISDN, it's sort of dialup rather than always on)
@Mage +1
As for this article's comment section: I am saddened, this is actually revolutionary news in this day and age and the comment section is hopelessly empty.
Believe me or not, I was dreaming about something like this the other day ... ok, I had server-grade "pi's" in mind - this is actually all that is missing .... design a new pi with high-speed interconnect, get 100 of them, hook them up over a high speed interconnect, put linux on it, and you have a 400 core system, for half the price of a single Intel® Xeon® Processor E7-8893 v4 ...
>a 400 core system, for half the price of a single Intel® Xeon® Processor E7-8893 v4
Which is great if you can parallelise(sp?) your workload into 400 parts.
I love the idea and I think there is a lot of money "wasted" on vmware (for functionality which should be part of the OS) and cpu cycles wasted on hypervisors.
I'm hoping someone will come up with a more dynamic process migration and resource allocation. I'd like to see more management and less emulation and I suspect it needs to be open source because per-socket/core licensing leads to over-consolidation and latency. Excellent hardware designs will always fail if the software costs blow them out of the water and proprietary software hasn't got a handle on ARM licensing.
...as I look forward to restructuring an undergraduate "Computer Organization" course for the spring...
Should I de-emphasize the CPU organization? Is it time to pay more attention to how various elements are interconnected, than to what the elements are?
Even the processor? After all, what is a processor? A CPU? A GPU? Some application-specific FPGA? Does it matter, if the system has all three, or is their mutual interconnect more important?
Maybe I'll cop out, and just hide behind discussions of "Moore's Law".
What is all this obsession with cores ? It's even worse than the MHz thing that went before it.
Are people so F£$%^ing stupid these days they don't realise that unless you have something useful to do with them you're wasting money on many things including 1) Silicon 2) powering them 3) powering all the complicated cache-coherency logic 4) being slowed down because of interconnect complexity 5) being slowed by software locks
@jms22
Yup. Marketing droids in action, what did you really expect? Gotta push something "new and shiny" or people might just not spend money on things with them - things we don't need, but they need to sell.
And legacy software that could be re-written to use more than one thread/core, but isn't (yet, and maybe never).
It makes me sick to see many things that are trivially parallelizable sit and eat 100% of one core while the rest just sit there (and do cache coherency stuff etc etc, yes, I once designed this kind of thing). The only exceptions at this point appear to be video editing (yay) and some Java stuff (which for me, is a yuck, I don't write it though, thank heavens).
I do manage to do well enough when I design big data acquisition systems with databases and gui's, but I do think about how I'm going to "program into the strengths of the platform" before I do a design. BTW, that'd be scientific data aq, not advertising, which is disgusting.
They simply hit a mhz (ghz really) limit with affordable silicon processes, but had the need to keep the company going, so, the trick was, how can we use more transistors, since we have them, yet can't make the ones we had any big amount faster without severe cost increases and exotic materials. So we have what we have...
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
