4 ms·
Can’t we get content addressable memory already? That would make memcpy moot in many cases.
by actionfromafar 4y ago
Can’t we get content addressable memory already? That would make memcpy moot in many cases.
- warkdarrior 4y agoContent addressable memory with built-in copy-on-write, fine granularity, and security tags? Not sure it is possible.
- actionfromafar 4y agoWould be amazing though? I’m thinking lots of tiny compute near the RAM chips.
- rsaxvc 4y agoSounds like Belt.
- meindnoch 4y agoOk, I’ll bite: in which cases does content addressable memory obsolete memcpy? It’s just a lookup table in hardware…
- actionfromafar 4y agoContent addressable memory is the ultimate "zero copy architecture". Imagine a "CPU instruction" or library call which determines that, "no, I have already seen this address" (where address isn't an address as much as a unique hash) and can pass that hash along to the destination instead of running a copy. Overhead increases by effectively removing pointers as we know them, replacing them with hash+length, but the reward is lower memory pressure and better cache locality. It also carries the promise of making a "buffer overrun" not only impossible but unthinkable. In such a world, incrementing "one extra" would not step into some other heap or stack, because everything would be segmented.
- rwmj 4y agoSomething similar in this paper, putting an FPGA between the processor and DRAM so you can arbitrarily remap bits of it: https://arxiv.org/abs/2109.14349 https://arxiv.org/abs/2109.14349