The article appears to misunderstand the point of the project, and mostly just describes what zswap already does (writing compressed pages back into ram), rather than talking about cram's hardware-offloaded compression that allows cacheline-level access rather than page-level. (And they appear to be aware that they don't really understand it: “My explanation of CRAM might not be completely correct”)
The phoronix article is better, and I say that as someone who usually detests the quality of phoronix's technical writing.
tancop 21 hours ago [-]
It's not really about hardware offload, the biggest problem with zswap is that it's swap. The rest of the kernel treats it like a fast SSD (which is still incredibly slow compared to RAM) instead of slower memory that needs a bit of special handling on writes.
NUMA maps a lot closer to what compressed RAM actually is. The subsystem is more aware of CRAMs specifics so it can make better decisions about where to put allocations and everything gets faster. And it's less overhead because swap is not really optimized for frequent direct access but for NUMA it's the most basic function.
ahartmetz 14 hours ago [-]
Whoa, building it on top of the NUMA logic is pretty clever - also, finally all that complicated code does something useful on normal computers!
yonatan8070 10 hours ago [-]
Is the hardware support this needs common? Can it be used on regular x86-64 desktops/laptops or is it for specialized hardware?
madduci 24 hours ago [-]
Finally we can run frontier models locally
sroussey 19 hours ago [-]
We brute force AI models right now because a) we don’t know better, and b) it’s premature optimization.
I beg to differ on point b, but no one is delaying their next model just so they can concentrate on optimization.
Another is separating the the intelligence part of the model from the known facts part of the model (which can be better compressed)
djmips 8 hours ago [-]
The Chinese breakthroughs are often around optimization. And I gather it's because they are constrained and very bright so it's a natural progression.
ranger_danger 2 hours ago [-]
IMO Restrictions are exactly what sparks creativity in people... whereas too many options leads to choice paralysis, which is what I fear is a big problem in the FOSS world.
Don't get me wrong... I think having choices is still good, I just think that having too many choices is not equally as good, or necessarily better.
tintor 19 hours ago [-]
Compression doesn't really work for model weights.
Model quantization and model distillation are two techniques to reduce model size.
https://winworldpc.com/product/connectix-ram-double/windows-...
Nice to see its back, with better performance.
History repeated, with refinement.
Whoa, new options! I'm downloading the Quantum RAM with Haptic Feedback.
https://www.youtube.com/@LinuxPlumbersConference
not sure where it falls in the schedule here either https://lpc.events/event/20/timetable/#all
https://www.youtube.com/live/OPRciCSsdS4?t=2031
The phoronix article is better, and I say that as someone who usually detests the quality of phoronix's technical writing.
NUMA maps a lot closer to what compressed RAM actually is. The subsystem is more aware of CRAMs specifics so it can make better decisions about where to put allocations and everything gets faster. And it's less overhead because swap is not really optimized for frequent direct access but for NUMA it's the most basic function.
I beg to differ on point b, but no one is delaying their next model just so they can concentrate on optimization.
It’s coming though.
One example: https://siliconangle.com/2026/07/28/ai-model-compression-sta...
Another is separating the the intelligence part of the model from the known facts part of the model (which can be better compressed)
Don't get me wrong... I think having choices is still good, I just think that having too many choices is not equally as good, or necessarily better.
Model quantization and model distillation are two techniques to reduce model size.