Read more.Gradual evolution should start with HBM, progress to Computing in Memory (CIM).
Read more.Gradual evolution should start with HBM, progress to Computing in Memory (CIM).
I don't understand the technical ins and outs of this but I do think they are on the right lines.
I can easily see a situation where a large system doing specialised tasks needs to do a specific, simple operation on a packet as it hits the CPU which takes up a cycle no matter the simplicity. If that simple operation can be applied as it goes through the memory with less of a hit, it could take a meaningful load off the CPU. Take that to its logical conclusion and why would you need separate memory and CPU systems? Memory on CPU or Computing in Memory - depends who is looking at it.
I'm probably talking tripe.
No you're spot on. The compute operations for things like neural nets are actually very simple, you just need to do a lot of them as close to the data as possible. Embedding compute in memory (or expanding memory operations to include some basic compute) can net (ho ho) much better efficiency gains than using more complex CPUs.
If only logic and memory were made on the same process. One of them has to be compromised.
At least they aren't throwing some analogue in the mix.
Edit: OTOH, FPGAs are basically memory devices. They don't actually have programmable logic, they use sram look up tables to emulate it.
I know its not related, but just imagine if politicians was upgraded every 2-3 years and actually became better and more efficient.
What a wonderful world it would be.
Sadly the progression for those are on a totally other slope.
2 things are upgrading fast - memory and viruses.
As long as we don't bring about Skynet, everything is good.
SK conveniently forgets that each single CPU-to-DRAM connection requires a memory controller, which is quite a big piece of logic.
There are currently 1 users browsing this thread. (0 members and 1 guests)