#5950353

cucaulay malkin

2021-09-29 17:20

lol what that's 3000% efficiency improvement in 5 years what are they talking about

#5950370

kapu

2021-09-29 18:52

I they ment 30%

#5950392

Denial

2021-09-29 20:22

kapu:

I they ment 30%

They meant 30x but it's not as challenging when you're pretty far behind in most of those categories. It's not like they are saying Epyc CPUs itself will be 30x more efficient - it's a combination of efficiency improvements in hardware abd software with very specific worksets which can massively benefit from fixed acceleration (for example tensor operations).

#5950394

cucaulay malkin

2021-09-29 20:32

kapu:

I they ment 30%

30% would be awful

#5950412

vestibule

2021-09-29 22:18

Remember when M$ experimented with submarine type vessels for there servers for cooling and saving on energy. Damn good idea I thought.

#5950499

RED.Misfit

2021-09-30 09:34

They did achieve a superior than 25x improvement (31.7x according to them) in mobile CPUs through 2014-2020. (okay, they were coming from far probably at that time with their "architecture & nodes")

#5950520

Horus-Anhur

2021-09-30 11:20

This 30X better efficiency, is probably just in certain taskloads. They'll probably add some specialized units and instructions to accelerate some types of calculations.

#5950800

tunejunky

2021-10-01 05:38

this sounds more extreme than the reality of their progress with TSMC and their improvements in uArch. if i'm putting dozens (or tens of dozens in future) of cores on an Epyc chip, and that chip itself goes to a sub 3n process, more cores at a higher (12 - 20 %) efficiency just by node alone. not to mention apps and A.I. and given that there are two Epyc node shrinkages in that time frame i say highly doable.