OpenCAPI: Google and IBM Lead Tech Consortium To Speed Data Centre Performance (thestack.com)
An anonymous reader writes: IBM is leading a prestigious consortium of tech players in the open development of a new framework that, the company says, can speed data centre performance by a factor of 10. Participants in the OpenCAPI group include IBM, Google, Nvidia, Mellanox, Hewlett Packard Enterprise, Micron and Xilinx. Chris Johnson, a Principal Engineer at Google commented 'Google is committed to open standards and we are excited to contribute to the cross-industry use and development of OpenCAPI'. Google's collaboration with RackSpace on the Zaius server will include IBM's forthcoming POWER9 processor technology, which is built around OpenCAPI. Tom Eby, vice president of Micron's compute and networking business said:"While memory has always been an essential building block for computing, it is quickly becoming the critical technology to unlocking next-generation performance."
i want to talk about scared grits girls and linix
It could be a bust, but I think it looks like HyperTransport all over again. Intel misses out on a good standard. Honestly, what is up over there?
Someone had to do it.
Why no acronym expansion? That's not a well-known acronym.
It's an "open" version of IBM's Coherent Accelerator Processor Interface which is basically a specification for how programs can interface with specialized "accelerators" (usually FPGAs) by writing to designated sections of RAM. This method minimizes the number of alternations that need to be made to motherboards, allows the use of standard CPUs and standardizes the unifies the changes that need to be made to the kernel.
TL;DR: supercomputing nerd stuff
Anons need not reply. Questions end with a question mark.
Zaius Servers rock!
Generational bandwidth increases aside, the major features of this specification are virtual addressing to prevent malicious devices from scraping host/peer memory and reduce host kernel software overhead, and a coherency protocol to keep peer caches properly synchronized. This is a big win as more types of processing get offloadedd from host CPU.
Seriously, while not a direct competing standard per se, NVlink and Gen-Z stuff is getting interesting enough to give OpenCAPI a run for its money.