You are currently viewing SemiWiki as a guest which gives you limited access to the site. To view blog comments and experience other SemiWiki features you must be a registered member. Registration is fast, simple, and absolutely free so please, join our community today!
It seems there are a lot of mysteries out there when it comes to Intel and Micron, first 3dXpoint memory's status and now this Automata computer chip. Any information would be appreciated. If this even lives up to a fraction of its promise, it will not only shake up tech, but society and commerce. Even just thoughts or comments would be welcome.
I did not dig into it, but from what i have seen, it is "putting logic closer to memory".
But i don't undestand, why it is revolutionary. It is something, what every compute chips did for more that 20 years. Just instead of putting logic into memory, they used memory blocks embedded closer co computing core. Such as register files, caches. Or "Ultra RAM" and "Block RAM" in FPGA. Same story. For example, Last Nvidia GPU has more than 20MB of register file.
Question is only ratio between memory and logic.
My opinion: As i see Intel involved, I expect presentation about 1000x performance improvement (hype), then years of waiting, and then noncompetitive products.
My understanding is that it's a very mature DRAM with small processing elements on-die, and it can be programmed almost like an FPGA. What I didn't know until recently is that this is Intel's canceled Larrabee GPU project mashed together Micron's DRAM expertise.
Micron + UVA have some kind of Automata program to see what the thing can do in the hands of engineers.
I think it's an interesting approach to neuromorphic hardware. The problem is that DRAM transistors are horrible, so the actual computing is slow and inefficient. The upshot is a high bandwidth link to Gb's without ever having to go off-chip.
As with all new, radical, chip architectures it takes time for computer scientists to re-write their traditional code into a format that is optimized for the AP. The web site at AP Portal talks about University papers as examples, so we are kind of at the starting point, looking for some domain that is willing to port their ideas onto the new hardware to see if it yields benefits over the traditional CPU and GPU approaches from the past decades.
I had lunch with an ex Intel friend and here is the gist of our conversation on Intel:
Intel makes poor business decisions outside their domain of expertise over and overdue to many factors:
They promote internal guys with little or no experience in new domain
They have too many employees and sites like AZ, Folsom And Austin, to some extent, are not making significant impact, from revenue point of view.
They rarely hire senior guys from outside.. instead, they hire junior guys and train them do the same old way of doing things to keep internal competition to the minimum.
They are behind in technology (they are in denial) and architecture. They don't know what and how to fix anything. Operations people are running the show, it just doesn't work for product teams.
it's really is sad to see Intel following into the future versus leading like years gone by.
BK came from Ops so it's not surprising. There is a balance between optimization and innovation that is extremely hard to manage. Ops is focused on optimization to the extreme, but when you do that it limits the flexibility of product development, which is about "moving fast and breaking things" and maximizing variety to address every possible customer segment. Mature companies with established product lines usually find focusing on incrementally improving their existing product lines and optimizing their operations is the easiest way to increase profits, and so they do, at the expense of future innovation. This can be mitigated to a degree if you are a good acquirer and successfully buy innovation, but Intel is a notoriously bad acquirer.
This guy makes a claim the contradicts the previous. Twitter
So politics as usual with different campuses hating each other.
The things I found interesting major production will start in 2018. Closer ties with India which makes sense because both countries share the same economic concerns as well as security concerns with Isis. And autonomous cars will begin in the next decade if that happens that will be a big change. The guy in charge Yaniv Garity seems to do what he wants and doesn't seem to be anyone's subordinate.
Intel will have problems that will only get worse until they find another Andy Grove that they are willing to give a free hand. I met someone that worked with him while on vacation and she said he had a low BS and FLUF tolerance (my words). Getting things done was the priority.
