IBM claims new ‘in-memory’ computing architecture will speed up computers by 200 times

[Captain Gecko]

IBM Researchers believe this new prototype technology will enable ultra-dense, low-power, and massively parallel computing systems that are especially useful for AI applications. The researchers tested the new architecture using an unsupervised machine-learning algorithm running on one million phase change memory (PCM) devices, successfully finding temporal correlations in unknown data streams.

 

 Original article http://www.kurzweilai.net/ibm-scientists-say-radical-new-in-memory-computing-architecture-will-speed-up-computers-by-200-times 

[Spenser1337]

Doubtful 

[bcredeur97]

10/10 great post by the OP.

[Coaxialgamer]

2 minutes ago, Spenser1337 said:

Doubtful 

why ? Architecture plays a major role in computer performance . This is a radical departure from standard Von-Neumann/M.Harvard architectures .

[mynameisjuan]

200x increase....yeah in probably one specific task tailored for it. 

[Princess Luna]

Cool stuff sadly the thread will be moved since it does not attend the criteria for being posted in this section.

[Captain Gecko]

If someone wants to re-post it properly, they are more than welcome, I just thought it was interesting. 

[Princess Luna]

14 minutes ago, Sniperpaul296 said:

I just thought it was interesting. 

Nobody's bashing or any thing, it is interesting by all means... there simply is a posting rule that should be met for this particularly section [:

[Brooksie359]

39 minutes ago, Sniperpaul296 said:

If someone wants to re-post it properly, they are more than welcome, I just thought it was interesting. 

You could change it to fit the posting guidelines as well. 

[HomeBoi]

But will it improve crysis fps?

[Ryujin2003]

13 hours ago, mynameisjuan said:

200x increase....yeah in probably one specific task tailored for it. 

GeekBench!

[mr moose]

5 hours ago, HomeBoi said:

But will it improve crysis fps?

No, but it will give the AI bots the processing power they need to develop a virus for your PC that allows them to escape the confines of the game and reek havoc on the rest of the world.

[Damascus]

Can't wait for this to be implemented in 3 decades

[Herman Mcpootis]

Finally, with this maybe i can run crysis now...

[leadeater]

Threads Merged

[Ryujin2003]

2 hours ago, Leshya said:

So IBM, through a new architecture is claiming to have 200X the speed of current attainable performance. In other words this could me a game changer.

 

The theory behind his claim is quite interesting with the idea being memory isn’t used to its fullest potential.

I’m interested to see how this affects the wider populous and wheather this is simply tailored to one specific task as seen with intels refreshes claiming “up to” 15% improvements in certain tasks.

Examples of this performance is as follows;

Sauce:

http://www.kurzweilai.net/ibm-scientists-say-radical-new-in-memory-computing-architecture-will-speed-up-computers-by-200-times 

So are they saying it's faster if the ALU of used to help get the address for the next instruction from main memory and bypass the cache?

[AncientNerd]

The thing is that this has been know for a very long time, there were computers that implemented a non-von Neumann architecture back in the '50s and they are  still used in some specialized applications, this is not a non-von Neumann. When I was getting my BSCS we had lots of discussions of the various reasons that the von Neumann architecture was settled on for the core of generalized computers. nvN architectures require you to fetch instructions in a non-sequential way, and be able to process them out of order - the last few generations of x86 chips are actually moving in this direction more than this is due to the fact that both Intel and AMD seem to have solved some of the generalization issues that normally happen when dealing with out of sequence operations. What it appears IBM is doing is effectively drastically increasing the cache size for each individual control unit. So they move the bottleneck farther out to between the "computational memory" and the "conventional memory". So if I am understanding things correctly what they will be doing is trading chip real-estate on the CPU dies between CPU/Cache transistors and "computational memory" transistors. I question if this is a worthwhile trade off - it might be or it might not be. Can the CPU die size really afford several GB of memory on die to handle having enough of this new memory type to be worth the effort for general use?

[Beskamir]

Considering many tasks are IO bound I can actually see this having a noticeable effect. Maybe not as much as they say but it could help with better parallelization of tasks. Idk, I didn't actually read the full article so I'm not sure how they're implementing it aside from what I gathered from the title and diagram.