By: Mark Roulo (nothanks.delete@this.xxx.com),
Room: Moderated Discussions
--- (---.delete@this.redheron.com) on May 14, 2024 12:14 pm wrote:
> Philip Turner on Twitter pointed out a striking fact that,
> once you see it, can help but modify your thinking.
>
> "
> Going from M1, M2, M3, M4, the transistor count has increased exactly by 4 billion per year.
> Next chip will have 32 billion, then 36, then 40. We're in the era where transistor count
> scales linearly, not exponentially. Just like clock speed when Dennard Scaling stopped.
> "
>
> He's right.
> M1 16B
> M2 20B
> M3 25B
> M4 28B
>
> This certainly gives one explanation for why Apple hasn't been able to move IPC much,
> and has pivoted (at least for these 4 chips) to alternative ways, so far... successful,
> to boost performance. I'd phrase this more in terms of linear reductions in density,
> but that's harder to present in a way that hits home like the above does.
>
> Apple is leading edge, but the way to bet is that everyone else will hit the same wall. Will be interesting
> to track this sort of metric for Intel over the next few years w/ i4, i3, A20; though it's tougher
> to agree on what counts as "the same" chip for Intel vs Apple – is there a commonly agreed upon "Intel
> Entry Level" chip that plays the role of M and is much the same from year to year?
>
> Superficially chiplets may provide a way out of this – but that depends on the full cost of chiplets, AND
> the extent to which they actually help. (Sure you CAN move the GPU to a separate chiplet and have twice the
> space for the CPU, but if you don't improve density, does that help much? You can thereby provide 2x the cores,
> but we're adults here and we care about single threaded throughput, not cinebench dick measuring).
Transistors aren't getting cheaper with process shrink. They used to get cheaper with newer processes, but they don't any more. You can put more of them in a given area, but the cost per transistor is fairly flat so 2x the density would provide 2x the transistors in the same area but also 2x the cost from TSMC.
> Philip Turner on Twitter pointed out a striking fact that,
> once you see it, can help but modify your thinking.
>
> "
> Going from M1, M2, M3, M4, the transistor count has increased exactly by 4 billion per year.
> Next chip will have 32 billion, then 36, then 40. We're in the era where transistor count
> scales linearly, not exponentially. Just like clock speed when Dennard Scaling stopped.
> "
>
> He's right.
> M1 16B
> M2 20B
> M3 25B
> M4 28B
>
> This certainly gives one explanation for why Apple hasn't been able to move IPC much,
> and has pivoted (at least for these 4 chips) to alternative ways, so far... successful,
> to boost performance. I'd phrase this more in terms of linear reductions in density,
> but that's harder to present in a way that hits home like the above does.
>
> Apple is leading edge, but the way to bet is that everyone else will hit the same wall. Will be interesting
> to track this sort of metric for Intel over the next few years w/ i4, i3, A20; though it's tougher
> to agree on what counts as "the same" chip for Intel vs Apple – is there a commonly agreed upon "Intel
> Entry Level" chip that plays the role of M and is much the same from year to year?
>
> Superficially chiplets may provide a way out of this – but that depends on the full cost of chiplets, AND
> the extent to which they actually help. (Sure you CAN move the GPU to a separate chiplet and have twice the
> space for the CPU, but if you don't improve density, does that help much? You can thereby provide 2x the cores,
> but we're adults here and we care about single threaded throughput, not cinebench dick measuring).
Transistors aren't getting cheaper with process shrink. They used to get cheaper with newer processes, but they don't any more. You can put more of them in a given area, but the cost per transistor is fairly flat so 2x the density would provide 2x the transistors in the same area but also 2x the cost from TSMC.