Does the speed (bandwidth) of memory matter?

[CRquantum]

Thanks @JohnCampbell .

Yeah I mentioned in another thread,

Gauss Algorithm written in Fortran slower than written in Python for large matrices?

Thank you @rnoffke . Sorry for a lazy question, would you mind posting the whole Jupyter Notebook here? I tried the link of the Jupyter notebook book you posted, and I also pip install gauss in my anaconda 3. But when I run the Juypyter notebook I got an error message below, [image] I am using anaconda 3 on windows 10. By the way, I just bought a Macbook Air with 16GB ram. Here is what I found, for a memory heavy Fortran code I wrote, (it frequently operates big arrays and matrices, …

For a Fortran code I wrote which depend on heavy memory operations, I did found even the entry M1 with 68.5GB/s memory bandwidth run the code 2X faster than my Xeon-2186m laptop with DDR4 2666 which is 35 GB/s bandwidth.
I do not think M1 is really faster than intel’s CPU, but the memory bandwidth seems do make a difference. Perhaps that is why M1 may be good at video editing or something which heavily depend on the speed of memory. The high bandwidth memory may be a game changer.

I figure perhaps to match M1 Max’s 400GB/s bandwidth, on a regular PC motherboard which only support dual channel memory, you may need DDR5 30000, lol. Current world record is DDR5 8888.

I know the workstation mother board may support quad-channel memory. It perhaps is time for consumer PC to support quad-channel memory as well.

Most importantly, all the PC’s memory seems are all 64bit width, if they use 256bit width, even DDR4 2666 with on dual channel memory motherboard can achieve 35*4= 140GB/s. I guess M1 chip is mostly likely equipped with 128/256/512bit width memory, just like some GPU’s memory.