A18 Pro … your thoughts?

Cmaier said:
“It’s complicated.” There are at least two effects that come to mind. First, as temperature increases, the transistor threshold voltage decreases because the fermi level decreases due to more carriers being elevated into the conduction band by thermal kinetic energy. By itself, this would actually increase transistor speed (but cause other problems like noise sensitivity) because it takes less voltage swing to “switch” the transistor. That doesn’t necessarily make a circuit faster, though, because now it’s more sensitive to noise spikes and may take longer to settle into the final result. Anyway, the “toggle frequency” of a transistor would decrease linearly with increasing temperature. (or at least pretty linearly. I’d have to think about all the various factors that play into threshold voltage and think about whether any others are sensitive to temperature).

The bigger factor, though, is channel resistance. As temperature increases, channel resistance (between the source and drain) increases exponentially, and the slope will be much bigger than the threshold voltage slope. Note that there are a bunch of competing effects here, too. Increased carrier concentration decreases resistance with increasing temperature, but the carrier concentration increase due to temperature in the channel is dwarfed by the free carriers present due to doping, so other effects - thermal scattering - dominate and increase the resistance.
Click to expand...
Hm. OK, let me ask a more limited question then: In practice, does cooling a chip to subarctic temperatures change the timing?

I have been wondering for a while *why* it's possible to run Intel x86 chips at dramatically higher clocks - over 9GHz, in at least one case. The two competing possible answers I can see are:
1) Intel has left tons of complexity on the table in each pipeline stage (and everywhere else in the chip)- or put it another way, each stage is done with its work with tons of time to spare, which seems wasteful, or
2) Cooling the chip that much reduces transistor latency (I'm sorry, there's a term of art for this that is escaping me at the moment) enough to allow for this extreme overclocking.

... or is there a third option I'm missing?
 
NotEntirelyConfused said:
Hm. OK, let me ask a more limited question then: In practice, does cooling a chip to subarctic temperatures change the timing?
Click to expand...
Yes :-)

If you can make the chip cooler, then you can run at a higher frequency (at least for any modern CMOS-type CPU.)

I have been wondering for a while *why* it's possible to run Intel x86 chips at dramatically higher clocks - over 9GHz, in at least one case. The two competing possible answers I can see are:
1) Intel has left tons of complexity on the table in each pipeline stage (and everywhere else in the chip)- or put it another way, each stage is done with its work with tons of time to spare, which seems wasteful, or
2) Cooling the chip that much reduces transistor latency (I'm sorry, there's a term of art for this that is escaping me at the moment) enough to allow for this extreme overclocking.

... or is there a third option I'm missing?
Click to expand...

It‘s (2). „Ft“ (sorry - my ipad has decided I am typing german for some reason) is the „toggle frequency“ of the transistor, which is the maximum speed with which it can toggle back and forth between on and off. This frequency increases as temperature decreases, largely due to decreases in the channel resistance. So exotic cooling allows you to increase the clock frequency (which reduces the cycle time) because the transistors in a given cycle need less time.
 
Die shots from both the A18 and A18 Pro if anyone is interested.

chipwise.tech

Apple A18 & A18 pro die shot - ChipWise

Apple 18 and A18pro die shot analysis
chipwise.tech chipwise.tech

A18
1727829547954.jpeg


A18 Pro
1727829521819.jpeg


Edit. Yeah so I pasted the wrong image for the A18. Sorry about that. Correct now.
 
Last edited:
NotEntirelyConfused said:
Interesting. They say "The Neural Engine in the A18 Pro also sees a boost, with more cores and higher throughput". Did we know that? I don't remember that from the Apple event.
Click to expand...
Yes I am confused by this also. I thought they performed the same.

Edit. Double checking on Geekbench AI the scores are similar, so I assume it’s a mistake by the writers.
 
Jimmyjames said:
Yes I am confused by this also. I thought they performed the same.

Edit. Double checking on Geekbench AI the scores are similar, so I assume it’s a mistake by the writers.
Click to expand...
could be another situation where the extra cores are there but not used? it’s happened before.
 
leman said:
Is it identical die but mirrored image and different contrast?
Click to expand...
I extracted links to the two best (IMO) images in the article. Open these in two tabs of the same browser window and flip between tabs to make the differences clear.


The overall layout is similar and some blocks are the same, but you can also see real differences. As expected, A18 Pro's big SRAM arrays are bigger than A18. I also see that A18 has only 5 physical GPU cores (GPU cores are on the left in these pictures). There's lots of differences in I/O layout at the edges, and in what I think are miscellaneous non-CPU/non-GPU blocks (media encode/decode and so on).

Worth noting: they scaled both of these images to 600pix wide, with different heights. They are almost certainly not the same scale as each other. The A18 Pro die should be larger, probably in both dimensions.

This is definitive proof that they're different tapeouts.
 
Images overlayed, attempted to match up size, made black and white and layered then diffed. the darker, the image the more similar.

Yeah, they are different.
Apple_A18_Compare.jpg
 
That is interesting indeed! I am surprised that it made economic sense for Apple to go this way.
 
leman said:
That is interesting indeed! I am surprised that it made economic sense for Apple to go this way.
Click to expand...
They sell a LOT of phones. Their volumes of either model are way above anything we ever saw at AMD.
 
It is likely sensible from Apple's perspective because they both needed to hit their battery life goals (disabling one GPU core does reduce power usage a bit) and also the need to more clearly differentiate the regular and Pro lines.
 
Joelist said:
It is likely sensible from Apple's perspective because they both needed to hit their battery life goals (disabling one GPU core does reduce power usage a bit) and also the need to more clearly differentiate the regular and Pro lines.
Click to expand...
sure, but of course they could have accomplished both of these goals by using the A18 Pro and disabling circuits that they don’t want to use in the non-Pro phones.
 
Cmaier said:
sure, but of course they could have accomplished both of these goals by using the A18 Pro and disabling circuits that they don’t want to use in the non-Pro phones.
Click to expand...

Precisely. Producing two chips means they believe they can save money on the volume difference. From what I understand finalizing these layouts and masks can cost up to hundreds of millions. Really puts things in perspective.
 
leman said:
Precisely. Producing two chips means they believe they can save money on the volume difference. From what I understand finalizing these layouts and masks can cost up to hundreds of millions. Really puts things in perspective.
Click to expand...
Now that we have an area comparison, it seems pretty obvious that this is the way to go. The difference is ~15%. How many non-Pro chips will they ship? A *lot*. In the first year, they'll sell well over 100 million just in iPhones. But then they will continue to sell millions every quarter for (most likely) two more years as they continue to sell iphone 16 after introducing the 17 and 18. Then there's whatever ipads, homepods, appleTVs, and whatever else they sell with the A18. Saving 15% on each of those chips adds up to a *lot* of wafers.
 
leman said:
Precisely. Producing two chips means they believe they can save money on the volume difference. From what I understand finalizing these layouts and masks can cost up to hundreds of millions. Really puts things in perspective.
Click to expand...
I always found the “hundreds of millions” figure suspect, but i guess it’s possible. I would have guessed closer to $50million now, taking into account a team of 100 people earning double what senior people used to get paid, with benefits doubling the cost a second time. So it’s possible. Software tools add to that, of course, as do servers, etc. [However: keep in mind that if it costs $200million to tape out A18, it does NOT cost another $200million to tape out A18 Pro. A ton of the work on the one is applicable to the other.]

But the money saved by getting more die on a wafer must more than make up for that.
 

Similar threads

T
Apple open sources on device AI models - your thoughts?
Replies
6
Views
478
throAU
T
Back
Top