Discussion in 'Article Discussion' started by Gareth Halfacree, 20 Jun 2016.
13.1W, 1.78GHz average peak clock.
A thousand cores and still no auto-spell-checker... "have created whan they claim..."
It's great on paper, but mentions nothing about performance-per-watt, precision, or net IPC. Seeing as this is a specialized processor, it would likely only be used in servers. I don't really see a situation where a server would use between 1-99% of the cores, which makes the in-depth core control an unnecessary complexion. But, if that's just an inherent trait of this design, then that's pretty great.
I'd like to see where this goes, because clearly there are some advantages. But some of these advantages seem a little too obvious; stuff that GPUs could have done years ago. So I'm a little suspicious.
Going on the ucdavis.edu page about they say it has a 1.78 trillion IPS.
Although they also say "[Cores] transfer data directly to each other rather than using a pooled memory area" but Gareth says otherwise, have i misunderstood.
I went to fix this, but I was beaten to the punch by Matt - still, cheers for flagging it!
No, I think the students and I are using different definitions here: I'm saying it has 'pooled memory' in that the cores don't have to write their data to external memory to pass it to the next core, while the students are saying it *doesn't* have pooled memory for the same reason. Their definition is almost certainly more correct than mine, given that they're graduating from computer science courses and have designed a processor and I haven't done either of those things...
Isn't that a lot like a GPU, with thousands of stream processors/CUDA cores etc..?
I know GPU stream processors just divvy up the work load rather than be individually programmable, but isn't this just an evolution of that?
Wonder how much PPD it would get folding.
None, it would be too busy trying to run Crysis
Separate names with a comma.