News Researchers boost graphics performance through processing in-memory

65 percent improvements claimed.
https://www.bit-tech.net/news/hardware/2017/04/21/graphics-processing-in-memory/1

 

To clear up some misconceptions in the comments - and apologies if the article was unclear:

The technique works, as Ed and Corky have both mentioned, by adding a processing element to each memory stack which is capable of working directly on data stored in said memory. So, instead of the GPU having to read 8GB (or whatever) of data, do its thing, and write the 8GB back again, the processing happens on the memory directly - hence 'in-memory processing.' It's not a new idea, but it's the first time I've seen it applied to graphics processing with practical results.

As Ed mentioned, it's the exact opposite of bandwidth-dependent: the data doesn't go anywhere, so where the GPU can only work on the contents of the memory at the throughput of the memory bus the in-memory processing system can operate however quickly the memory itself works at - and in parallel, too, meaning if you've got eight stacks of memory you can do your processing eight times faster than if you had one stack of memory without worrying about saturating any buses.

You're limited in what you can do, though: the die space and power envelope for adding a logic layer to stacked memory are both way, way smaller than for a GPU - so you can't have anything general-purpose going on there. Hence the proof-of-concept: a logic layer that only does one thing, anisotropic filtering - something which is fairly simple computationally but that requires massive memory bandwidth. With that tiny bit of extra processing power, you're lightening the load on the GPU by a percent or two at most - but because you're no longer bottlenecked by the memory bus you're increasing the performance by 65 percent.

Step one of commercialisation: task offload acceleration, by adding anisotropic filtering logic to the memory stacks (or whatever task ends up making sense to offload - there may be something else that would give even bigger performance increases in modern gaming engines.)

Step two: add more logic layers. As well as your anisotropic filter logic layer, why not stick a - I don't know - bump-mapping layer on there? Keep adding layers until you can't fit any more on there.

Step ???: by now your GPU is basically just there to tell the memory stacks what they should be doing, so you've effectively created a fundamentally new architecture. Instead of an ultra-powerful GPU talking to dumb memory, your graphics card is now a dumb and lightweight central controller talking to ultra-powerful in-memory processors. Likely? Who knows; the technique has to survive the prior steps first.

As for the paper, I'll drop the guys an email and see if there's a timescale on public access - or, given that it's DoE funded, whether it'll ever be publicly accessible.

 

I'm Ron Burgundy?

 

Got in touch with Shuaiwen Leon Song at PNNL, and he's out travelling this week but he's going to swing back around next with with as much additional information as he can gather. Should be interesting!