Intel, which hit a brick wall about 10 years ago when the competition disappeared due to AMD's Bulldozer, is finally responding to both of those chips from last year and while it has been able to pull some decent performance improvements for once it definitely dipped into the power bucket for a lot of that gain. I wouldn't count that as proof x86-64 is unable to compete in high performance high efficiency computing anymore - I'd hold my judgement until Zen 4 and "M2" (or whatever it will be called) release around this time next year as it seems each will be a really big update for each chip. Who knows, maybe Intel will have something new by then too now that they have actual competition again.
Size (mm^2) TDP (w)
Apple APU
M1 119 18
M1 Pro 251 ??
M1 Max 425 90 (peak -- not TDP)
AMD CPU/APU
4800U 156
5800U 180
5800X 206
5950X 286
Epyc Rome 1008
Intel CPU/APU
Ice Lake 123
Tiger Lake 146
Comet Lake 206
Rocket Lake 276
Alder Lake S 209
Alder Lake M 215
AMD GPU
Vega 20 331 300
Navi 14 158 150 (80 mobile)
Navi 10 251 235 (120 mobile)
Navi 23 237 160 (100 mobile)
Navi 22 335 230 (145 mobile)
Navi 21 520 300 (no mobile)
Nvidia GPU
TU106 445 175
TU104 545 250
TU102 754 280
GA106 276 170 (115 mobile)
GA104 392 290 (150 mobile)
GA102 628 350 (no mobile)
Nothing comes close to the M1 when looking at the combination of performance, TDP, and die area.M1 Pro has worse multi-threaded integer performance than a 5950, but better floating point performance. It does this in a 251mm^2 die vs a 286mm^2 die AND while most of the die is used up by the GPU.
M1 Max has the equivalent of GA104 tacked on to its 8 CPU cores (plus 2 little cores). AMD's equivalent would be a 5800H plus a Navi 22 mobile. Together, those would use more than double the power while being 100mm^2 larger.
- GPU on the M1 Pro/Max is really impressive but irrelevant
- M1 Pro/Max came out a year later than the CPUs they are being compared to here
- The M1 takes 133% the transistor count on a generation newer TSMC node to get as far past the 5800 as it does
So from a CPU architecture perspective I really can't write x86 as having hit a brick wall to the point it will never be able to keep up with ARM, that's just not what the M1 has shown.
From an overall perspective though yes the M1, particularly Pro/Max with their larger GPUs (can't wait for mine to arrive!), are hands down the best overall systems of their classes right now I just don't think that has as much to do with the CPU being ARM instead of x86 as GP made it out to be.
It makes their performance look good relative to M1 for now. But what happens when Apple push out their next chip design? Is Intel going to up the watt usage by a huge amount again?
Nah , this is a dead end. Alder Lake is a blind alley.
A15 performance cores had significant gains and really improved integer performance all while reducing power consumption slightly.
A15 efficiency cores are even more amazing. They got something like 30-40% faster. They are now roughly as fast per clock as the big cores in x86 systems while still using only a fraction of the power of the big cores. This roughly translates into 8 full-size x86 cores, but with 4 of those being even faster.
Then there's the massive cache (32mb on mobile and probably 48/64 on M2) to further push down power consumption. They'll also be bumping the GPU by 30% (more if it doesn't throttle).
AMD really needs to either pick back up their ARM designs or start heavy investment into RISC-V designs.
And I'll believe that they're off of their node formally known as 10nm when I see it. It's been close to ten years straight that Intel has been slipping their public deadlines for bringing a new node up.
So, yeah, for a company that used to be a full node ahead of everyone, there's some serious stagnation going on.
[1] https://www.extremetech.com/computing/311275-intel-doesnt-wa...
Because the M1 Max is not less than half
https://browser.geekbench.com/v5/cpu/compare/11136167?baseli...
A15 has already been out and spoiled what's coming in M2.
Size
M1 is 119mm^2 and M2 is probably at most around 130-140mm^2 (depending on cache and RAM controllers) which is still smaller than 5800U at 180mm^2 and much smaller than Alder Lake mobile at 215mm^2.
Cache
A14 and M1 have 16mb of SLC (system level cache). A14 has 8mb L2 while M1 has 12mb. A15 doubled to 32mb SLC and jumped from 8 to 12mb L2. I suspect M2 keeps the 32mb. 12 or 16mb of cache will depend on the performance increase, but I suspect they'll stick with 12mb.
5800U has 4mb of L2 cache and 16mb of L3 cache. Alder Lake desktop has 14mb L2 cache and 30mb L3 cache (dropping to 9.5mb L2 and 20mb L3 on the 12600 and 7.5mb L2 and 18mb L3 on the leaked 12400).
Power
Peak power for M1 is around 15-20w absolute max power. A15 improved power efficiency 17% over A14, so we could actually see power usage improve. Those massive caches may slightly increase chip power, but they decrease memory accesses which radically decrease total system power. The chip running at far lower peak frequencies further helps to keep cache power consumption under control.
5800U can hit upwards of 50-60w peak power to keep the stated turbo speeds. Current Intel chips do this too and I suspect that Alder Lake won't be any different here.
CPU Performance
A15 P-core increases 8-10% (up to 37% on some things) and a lot of that was integer improvements. A15 E-core increased performance by about 30%. Their actual performance is close to Zen 2. Overall, performance in A15 was around 20% better than A14. GPU performance also jumped 32%.
M1 already handily beats 5800U in pretty much everything. I'd bet heavily that M1 beats the little Alder lake. I'm less sure about the bigger alder lake because I don't know what frequencies it will be able to sustain (I'd bet it still beats out M2 though).
GPU Performance
M1 shares SLC between the CPU and GPU. This works much like AMD's Infinity fabric to reduce necessary bandwidth and allows a pretty big increase in GPU performance over what would otherwise be possible.
AMD hit the iGPU wall a few years ago. They've slowly increased GPU resources as DDR bandwidth has increased, but that's about it. I'm really looking forward to Zen 4/5 + RDNA 2/3 + Infinity Cache, but I don't know when that will actually arrive. I've seen iGPU tests of the U770 on Alder Lake desktop and it's not even as fast as last-gen Iris Xe. The mobile variant will have 3x the GPU resources, but it still only looks to be 30% faster than last-gen and then only if it isn't bandwidth limited. This seems possible with DDR5, but I'd guess it's capped at last-gen performance on DDR4 systems.
As much as I dislike Apple's overall business practices, they are 100% correct about ARM being the future of high performance, high efficiency computing.
And we need high efficiency in a world that is becoming starved of energy as more and more people get connected every day.