The Future of Microprocessors • Sophie Wilson • GOTO 2024

[MT]

Sophie Wilson - Computer Scientist & Co-Creator of the ARM Processor:
The Future of Microprocessors - a talk about the history of microprocessors, how we got here and what might happen next. There will be two laws, one equation, some graphs and a particle beam weapon out of Star Trek.

[brucehoult]

Will (shortly) watch with interest. Wilson did great work on the original ARM, correctly analysing what made the 6502 so fast in practice and putting similar properties into the Arm ISA.

I'm not aware of any useful contributions after 1990 though. Wilson chose not to go to Arm Ltd, and was critical of Thumb and Thumb2, which were in fact what rocketed Arm to its current prominence.

[brucehoult]

And in fact next to nothing about the FUTURE of microprocessors, more like "how we got here". And a bit about the future of transistors.

[SiliconWizard]

Yes, nothing new to see really, merely a recap of what has happened until now.
The "future" is basically what we currently do, like "hybrid" multi-core chips.
"3D" transistors have been around for a while. One good point is that as we get into the finer process nodes the ratio number of transistors/cost may not be so attractive anymore. But I'm pretty sure we still have some margin.

My "fear" currently is that the focus on NPUs/TPUs may be putting a halt to the improvement of CPUs.

[MT]

And in fact next to nothing about the FUTURE of microprocessors, more like "how we got here". And a bit about the future of transistors.

yes very dry on the future but a nice walk trough of the past and the nanometer race.

[coppice]

Will (shortly) watch with interest. Wilson did great work on the original ARM, correctly analysing what made the 6502 so fast in practice and putting similar properties into the Arm ISA.

I'm not aware of any useful contributions after 1990 though. Wilson chose not to go to Arm Ltd, and was critical of Thumb and Thumb2, which were in fact what rocketed Arm to its current prominence.

I believe she developed embedded DSP architectures at Broadcom after ARM. Those don't get as much public visibility as something like the ARM.

[Simmed]

from the video :
you buy a 6Ghz computer
runs 6Ghz for 15s, hits thermal, backs down 5Ghz 4Ghz
cores start to turn off

i dont know why you bought it

 

[brucehoult]

from the video :
you buy a 6Ghz computer
runs 6Ghz for 15s, hits thermal, backs down 5Ghz 4Ghz
cores start to turn off

Well, actually, while my i9-13900HX laptop is limited to 5.4 GHz, not 6, it'll run one compute-bound thread [1] at 5.4 GHz indefinitely.

i dont know why you bought it

 

A little thing called "Amdahl's law". It's mentioned in the video.

Yup, it can only run all 24 cores at around 3.6 GHz continuously on the parallel parts of the computation, but when you come to the single-threaded parts it makes a significant difference if you can run those parts at 5.4 GHz instead of 3.6 GHz.

Or to look at it from the other angle ... even if you have to back the MHz off 33% going from single-core to 24 cores -- you're still running 24 cores...

I tried a cross-compile of a RISC-V Linux kernel two ways:

17m38.6s taskset -c 8 make  // 5.4 GHz all the time

 1m27.5s make -j32             // actually throttled back to 3.6 GHz after a while

So an actual speedup of 12x, but 18x compared to if the single-core MHz had been limited to the same 3.6 GHz as all-cores (which would be 26m28s).

Not bad for 8 big P cores and a bunch of small cheap E cores -- as described in the video.

[1] it turns out that on my particular chip, virtual cores 8&9 (the two hyperthreads of one P core) can run continuously at 5.4 GHz, while the other P cores will only do 5.2 GHz continuously. Fun with `taskset` !!

[coppice]

She has been giving this talk for years, gradually adding a bit more of the most recent stuff. Its has always been called "The Future of Microprocessors" and it has never included anything about the future. Its all history.