Hacker News new | comments | show | ask | jobs | submit login
HPE Powers Up The Machine Architecture with ARM (nextplatform.com)
62 points by jcbeard 5 days ago | hide | past | web | 10 comments | favorite





Wasn't "The Machine" (whoever came up with that name should be whipped with a wet noodle) supposed to have memristors? I skimmed the description, but didn't see any mention of this having memristors, just conventional DIMMs.

Actually it has a very large shared non-volatile memory. HP felt memristors would be a great way to make that memory but as it is pointed out in the article you can make it with regular DRAM and leaving the power on, or with Intel/Micron's cross point (Optane) NV-Dimm tech.

From what I have read, the memristors were the goal but they just aren't ready yet. Rather than wait, they implemented The Machine with what is available now.

It was. Then again, it's been a few years. In the meantime, other forms of non-volatile memory became cheaper.

The article, like most, makes a big deal about the (not yet really extant) memristers but the other interesting technical advance in this device is the optical interface (optical-to-the-die interface) which is only discussed in passing in this article. That alone could potentially carry HPe, which is good since what else could?

They spend 9 paragraphs discussing it, and other aspects of the optical system.

I think it's pretty neat. I'm mostly skeptical because commodity compute has led HPC for so long. Novel architectures are often quite interesting, but feel like they only have a performance advantage for short periods of time.

It looks like a neat system though, and I'm sure many of the ideas will trickle down into commodity hardware eventually. It's great that HP are still doing this kind of stuff.


Every company is working on that optical stuff and it keeps getting delayed. It could only save HPE if they get it working and their competitors don't.

My idea of a vacation would be playing with this thing. In looking at variations in systems architecture I see this as the opposite extreme of Google's "shared nothing" architecture. There is probably a PhD thesis in developing a calculus that can transform an algorithm P from one to the other architecture.

> There is probably a PhD thesis in developing a calculus that can transform an algorithm P from one to the other architecture.

Far more than a single thesis. That is a grand challenge problem.


Since it seem that they're showing working prototype, does it mean the wider world will get some hand on too ? (not the layman of course...) Would be a fun toy to tinker with !



Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | DMCA | Apply to YC | Contact

Search: