AVX2 is dimwitted compared to AVX512

[mocore]

*puts ML hat on* 

older gpu/hw fails to solve mathematical riddles ...

https://github.com/google/gemma.cpp/issues/23

On a $10,000 AMD Ryzen 7995WX (znver4 avx512) Gemma 7b instruct sfp is able to solve mathematical riddles.

 

[Rich]

Hi mocore

... older gpu/hw fails to solve mathematical riddles ...

https://github.com/google/gemma.cpp/issues/23 ...

I'll admit a lot of the stuff they were talking about was over my head, but
to me it appeared the second machine was treating this as a linear problem
instead of an exponential one.

Exponential (volume doubles each day):
Algae growth rate = StartQuantity * ( 2 ^ Days )
Days to fill reservoir = 30
Days to fill half reservoir = 29

Linear (volume increases by equal amount each day):
Algae growth rate = StartQuantity * Days
Days to fill reservoir = 30
Days to fill half reservoir = 15

Of course it still could have been a subtle math error and the fact
that it answered 15 was just a coincidence.

[mocore]

hi rich

to me it appeared the second machine was treating this as a linear problem
instead of an exponential one.

interesting to hear you perspective on this
...  it appears you where paying attention in math class and have reattained the numeric reflex's

of course  *

Of course it still could have been a subtle math error and the fact
that it answered 15 was just a coincidence.

... many potential methods to get to the correct answer , (presumably?) even more for less correct ones

which reminds me of count down
https://www.youtube.com/watch?v=Lbhxv5QPErM


much ( if not all lol ) of the details are likely way over my head also

...however  wrt the (*wonders* of modern hardware  ) hw 

https://en.wikipedia.org/wiki/Advanced_Vector_Extensions

Downclocking
Since AVX instructions are wider and generate more heat, some Intel processors have provisions to reduce the Turbo Boost frequency limit when such instructions are being executed.

https://en.wikipedia.org/wiki/Single_instruction,_multiple_data

https://en.wikipedia.org/wiki/Flynn's_taxonomy

[mocore]

fwiw
happened to read this https://lists.gnu.org/archive/html/help-guix/2025-06/msg00081.html
attempt to  "get the exact same result on both AVX512 and non-AVX512
computers without much of an overhead"


and find this
 https://github.com/twosigma/libvirtcpuid/ - libvirtcpuid provides transparent CPUID virtualization, all in userspace.
 strange workaround mentioned