3 ms·
at work I got a m1 pro with 16GB of RAM just because they offered the 14" managed device with 16GB only, while they offered 32GB on the 16" devices. When I comp
by gbil 4y ago
at work I got a m1 pro with 16GB of RAM just because they offered the 14" managed device with 16GB only, while they offered 32GB on the 16" devices. When I complained about this, the response was "but is 16GB unified memory" as if I was disputing the performance of the RAM and how it is handled within the M1 chip while my issue was filling the 16GB. I'm happy that at least in HN people seem to be more aware on RAM sizing
- Damogran6 4y agoMore ram = more performance if there's a cost in going across a bridge to keep the pipeline full. 16gb of _unified_ RAM that can be accessed by the CPU and GPU without a performance hit in going to the motherboard is a factor in some workloads. Whether you agree or not, it's an effective test comparing two architectures. So the question is: are you spec benchmarking, or are you interested in the performance of the workload, because it's the actual work that matters, not how the bits are stored under the keyboard.
- gbil 4y agoI'm interested on not swapping which is a simple matter of RAM size, regardless if that is unified or not - which I believe I touched in my post
- Damogran6 4y agoOnly if your datasets, in aggregate, extend beyond what can be compressed into 16Gb...an what's the performance of the device you're swapping to? There's a GREAT BIG DIFFERENCE between an RLL 80 Gb drive and What you're swapping to on the motherboard today... (In fact, you could consider the on-die RAM to be a really big Cache and the Flash storage to be REALLY REALLY big Non volatile addressable RAM.)