Apple GPU stumble reported

Cmaier

Site Master
Staff Member
Site Donor
Posts
5,215
Reaction score
8,264

I believe this for not one second. The reporting has discussions of ”early prototypes.” The reporting is based on this idea that CPU designers design something, send it out to make a prototype, then find out how much power will be consumed. At that point they have time to completely replace the GPU and try again.

Hilarious.

1) the power consumption is highly predictable at the design stage using software. Such software is commercially available, but it is so easy to predict that we did it within 5%-10% at AMD using a perl script I wrote. You can actually get pretty close by multiplying your total capacitance x your voltage squared x half your clock frequency.

2) you don’t make “prototypes.” When you tape out a chip, the expectation is that it will be the production chip. You may very well have to do additional spins to fix unforeseen problems. A cross-coupling issue your software simulations didn’t account for, a logic bug that didn’t turn up in verification, etc. But your fixes will not include completely ripping out the GPU and then fitting a new GPU into the same chip area (because you certainly aren’t going to completely change the size and shape of the chip - that would require revisiting the design of every other part of the chip). You move some metal around. Maybe, worst case, you have to add new logic gates.

3) If there WAS a power prediction problem, it would have affected the whole chip. The only way it could happen is if capacitance was wildly mispredicted, which would affect the CPU too, and would also mean that predicted clock speed would be way off.

4) there simply isn’t time in the design cycle to take multiple shots like this

What might have happened?

The design team found that to hit their performance goals with the architecture they were going with, they’d need too much power. Possibly because the new thing was originally intended for 3nm. They would discover this at the design stage, not via “prototypes.” So before they taped anything out, they put the new thing on the back burner and went with the old thing this time around.
 

leman

Site Champ
Posts
610
Reaction score
1,122
Yeah, what you write makes much more sense to me. Also explains your earlier comments that your sources tell that a new GPU is ready, and also why Apple published a bunch of new GPU patents recently/ It is possible that the article simply messed up the details, and that there indeed was a serious bug discovered at a late stage, but if this is indeed about power there is no way that the chip would go into production as is.
 

Cmaier

Site Master
Staff Member
Site Donor
Posts
5,215
Reaction score
8,264
Yeah, what you write makes much more sense to me. Also explains your earlier comments that your sources tell that a new GPU is ready, and also why Apple published a bunch of new GPU patents recently/ It is possible that the article simply messed up the details, and that there indeed was a serious bug discovered at a late stage, but if this is indeed about power there is no way that the chip would go into production as is.
I think the most likely thing is that the new GPU architecture they came up with simply isn’t practical on the node they got stuck on, and that we will be seeing it as soon as they shift to N3. Probably also where the ray tracing comes in.

It’s possible that it came as a surprise to them that they’d be on this node for A15, and very early in the design process they had to re-think things. But I seriously doubt this was any sort of last minute change of plans.
 

Roller

Elite Member
Posts
1,392
Reaction score
2,698
Not that Apple hasn't had its share of design miscues over the years, but the way this is being sensationally reported on sites like AppleInsider and MR, it's an indication of gross incompetence due to departures, which I don't buy. All high tech companies deal with loss of personnel to rivals large and small.
 

Cmaier

Site Master
Staff Member
Site Donor
Posts
5,215
Reaction score
8,264
Not that Apple hasn't had its share of design miscues over the years, but the way this is being sensationally reported on sites like AppleInsider and MR, it's an indication of gross incompetence due to departures, which I don't buy. All high tech companies deal with loss of personnel to rivals large and small.

They are pretending that Williams was responsible for simulating power consumption or something. It’s all nonsense.

It seems pretty clear that something has happened that put Apple behind where they hoped to be. But I believe it to be a fab-related hiccup.
 

Colstan

Site Champ
Posts
822
Reaction score
1,124
I think the most likely thing is that the new GPU architecture they came up with simply isn’t practical on the node they got stuck on, and that we will be seeing it as soon as they shift to N3. Probably also where the ray tracing comes in.
As good as Apple Silicon is, when it was first announced, I knew I'd be waiting until the M3 landed. I always like to wait a couple revisions before jumping on board. This does play into the idea that the M2 was a "stopgap" and the real fun starts with N3.

Hey @Cmaier, it sounds like there's a good chance we'll be getting a new GPU with Apple's next offering, do you think they'll pair it with new P-cores?
Not that Apple hasn't had its share of design miscues over the years, but the way this is being sensationally reported on sites like AppleInsider and MR, it's an indication of gross incompetence due to departures, which I don't buy. All high tech companies deal with loss of personnel to rivals large and small.
Every time a janitor leaves Apple, we get another story about "brain drain" at the company. It's funny that we never hear that about any of their competitors.
 

jbailey

Power User
Posts
164
Reaction score
180

I believe this for not one second. The reporting has discussions of ”early prototypes.” The reporting is based on this idea that CPU designers design something, send it out to make a prototype, then find out how much power will be consumed. At that point they have time to completely replace the GPU and try again.

Hilarious.

1) the power consumption is highly predictable at the design stage using software. Such software is commercially available, but it is so easy to predict that we did it within 5%-10% at AMD using a perl script I wrote. You can actually get pretty close by multiplying your total capacitance x your voltage squared x half your clock frequency.

2) you don’t make “prototypes.” When you tape out a chip, the expectation is that it will be the production chip. You may very well have to do additional spins to fix unforeseen problems. A cross-coupling issue your software simulations didn’t account for, a logic bug that didn’t turn up in verification, etc. But your fixes will not include completely ripping out the GPU and then fitting a new GPU into the same chip area (because you certainly aren’t going to completely change the size and shape of the chip - that would require revisiting the design of every other part of the chip). You move some metal around. Maybe, worst case, you have to add new logic gates.

3) If there WAS a power prediction problem, it would have affected the whole chip. The only way it could happen is if capacitance was wildly mispredicted, which would affect the CPU too, and would also mean that predicted clock speed would be way off.

4) there simply isn’t time in the design cycle to take multiple shots like this

What might have happened?

The design team found that to hit their performance goals with the architecture they were going with, they’d need too much power. Possibly because the new thing was originally intended for 3nm. They would discover this at the design stage, not via “prototypes.” So before they taped anything out, they put the new thing on the back burner and went with the old thing this time around.
A perfect example of Knoll's law or the Gell-Mann amnesia effect. Most people reading this report will just assume that the journalists did their job and verified that the anonymous sources and rumors had a basis in fact. But Cliff, being an expert in the field, can point out succinctly why the story doesn't make any sense from a technical perspective.

Unfortunately, we know how this is going to go. Any story that can be critical of Apple that has the appearance of an unbiased source is going to spread like fire.
 

Cmaier

Site Master
Staff Member
Site Donor
Posts
5,215
Reaction score
8,264
As good as Apple Silicon is, when it was first announced, I knew I'd be waiting until the M3 landed. I always like to wait a couple revisions before jumping on board. This does play into the idea that the M2 was a "stopgap" and the real fun starts with N3.

Hey @Cmaier, it sounds like there's a good chance we'll be getting a new GPU with Apple's next offering, do you think they'll pair it with new P-cores?

Every time a janitor leaves Apple, we get another story about "brain drain" at the company. It's funny that we never hear that about any of their competitors.
I’m sure there will be new P-cores, but I don’t know how different they‘ll be. I guess Arm v9 still needs to make an appearance? Since the SRAM scaling is not great from node-to-node, that, by itself, will probably result in some changes. For example, extra pipeline stages. Then you want to hide those stages by improving prediction, etc. At some point it might even make sense to do multi threading just to fill holes in the pipeline caused by memory reads that now take more cycles. In the immediate future, I don’t expect massive changes, but who knows.
 

dada_dave

Elite Member
Posts
2,067
Reaction score
2,046
As good as Apple Silicon is, when it was first announced, I knew I'd be waiting until the M3 landed. I always like to wait a couple revisions before jumping on board. This does play into the idea that the M2 was a "stopgap" and the real fun starts with N3.

Same boat. I have other reasons too, but that’s a big one.

Hey @Cmaier, it sounds like there's a good chance we'll be getting a new GPU with Apple's next offering, do you think they'll pair it with new P-cores?

Every time a janitor leaves Apple, we get another story about "brain drain" at the company. It's funny that we never hear that about any of their competitors.

To be fair people cited that a lot for Intel during its malaise.
 

Citysnaps

Elite Member
Staff Member
Site Donor
Posts
3,603
Reaction score
8,825
Main Camera
iPhone
It's been fun (for lack of a better word) watching all the experts at the other place pontificate about this-n-that on a story that I suspect is based on dubious facts.
 

theorist9

Site Champ
Posts
603
Reaction score
548
Not me. I never acknowledged that there were any brains at Intel in the first place
Plus they practice blatant age discrimination. Nearly all their open positions are for individuals who got their PhD within the last (IIRC) 3 years. It's a way they can discriminate based on age, yet get away with it, since the ovewhelming majority of new PhDs are in their late 20's - early 30's.

Of course, that may not make them different from any other tech company these days.
 

Cmaier

Site Master
Staff Member
Site Donor
Posts
5,215
Reaction score
8,264
Plus they practice blatant age discrimination. Nearly all their open positions are for individuals who got their PhD within the last (IIRC) 3 years. It's a way they can discriminate based on age, yet get away with it, since the ovewhelming majority of new PhDs are in their late 20's - early 30's.

Of course, that may not make them different from any other tech company these days.

I didn’t know that. That’s pretty dumb - heck, I was discriminated against by DEC because I had a PhD - two interviewers asked me “why‘d you get your PhD? That was dumb” or words to that effect.
 

theorist9

Site Champ
Posts
603
Reaction score
548
I didn’t know that. That’s pretty dumb - heck, I was discriminated against by DEC because I had a PhD - two interviewers asked me “why‘d you get your PhD? That was dumb” or words to that effect.
Several years ago I had some friends from grad school circulate my CV at Intel, and it ended up in the hands of a manager who said "It looks like you'd be perfect for this postion, but I can't touch you because you got your PhD more than three* years ago." [*I think it was three.]

I'm sure if they hired me their chips would now be sipping power and they'd have more EUV machines than TSMC 😄.
 

Colstan

Site Champ
Posts
822
Reaction score
1,124
While there are vague rumblings about Apple's alleged issues, Cliff's favorite company is about to have another Rocket Lake moment.



Intel source: "I don't know what marketing will pretend, but I wouldn't expect to see Meteor Lake laptops until Q3 2023 if we are lucky, and we'll be even luckier if we see Meteor Lake desktop at all..."
 

exoticspice1

Site Champ
Posts
298
Reaction score
101
Not me. I never acknowledged that there were any brains at Intel in the first place
Not me. I never acknowledged that there were any brains at Intel in the first place

What about the current AMD?

My take is that they are doing good.
Their 5nm CPUs are good but I would expect better efficieny from their 65 watts Zen 4 parts in Jan 2023.
Their RDNA 3 GPUs have buggy drivers but since RDNA 3 is chiplet based its bound to cause issues that will be there for a few months. Still not a total dissappointment like Intel's GPUs.

Intel not so much, still on old nodes and bad GPUs.
 

NT1440

Power User
Posts
194
Reaction score
216
Not that Apple hasn't had its share of design miscues over the years, but the way this is being sensationally reported on sites like AppleInsider and MR, it's an indication of gross incompetence due to departures, which I don't buy. All high tech companies deal with loss of personnel to rivals large and small.
I say this constantly to people. Tech “journalism” is not journalism by any means. Half tech-literate people handing in click bait articles get the completely tech illiterate audience to buy just about anything these days.

Given the logistics involved, as the first post of this thread laid out, this story is simply false.
 

Cmaier

Site Master
Staff Member
Site Donor
Posts
5,215
Reaction score
8,264
What about the current AMD?

My take is that they are doing good.
Their 5nm CPUs are good but I would expect better efficieny from their 65 watts Zen 4 parts in Jan 2023.
Their RDNA 3 GPUs have buggy drivers but since RDNA 3 is chiplet based its bound to cause issues that will be there for a few months. Still not a total dissappointment like Intel's GPUs.

Intel not so much, still on old nodes and bad GPUs.
i don’t pay attention to AMD for the same reason I don’t pay attention to what’s going on with SPARC processors. Making the best x86 chip today is a lot like making the best internal combustion engine today: being the best at a dying technology is not all that interesting to me.
 

exoticspice1

Site Champ
Posts
298
Reaction score
101
Making the best x86 chip today is a lot like making the best internal combustion engine today: being the best at a dying technology is not all that interesting to me.
you say that but x86 will stay alive in consumer, prosumer and server for a long time(more than 10 years) but time will give us an answer.
 
Top Bottom
1 2