Apple family GPUs

[awesomedeluxe]

Anandtech did an in-depth analysis:

The Apple iPhone 11, 11 Pro & 11 Pro Max Review: Performance, Battery, & Camera Elevated

www.anandtech.com

Have in mind that SPECint and SPECfp not only measure a single CPU core but also RAM and compiler speed.

I've read that several times and somehow apparently never came to the correct conclusions. Including reservations about RAM (I always struggle to estimate RAM power usage), that still puts the lightning core power usage higher than I expected.

Of course, it goes without saying the cores can and will use less power as more of the system comes on line. A13 clock speeds drop to 2.56GHz when a second CPU core turns on. I'd like to know if they drop further when the GPU cores are fully in use - but one upside of a dual perf core design is that your CPU cores don't become too much of a liability.

Looking again at the Bloomberg design, and assuming now that Apple would need to give 4W a piece to sustain performance equivalent to the iPhone's peak, then an 8 firestorm core part is already demanding 32W for its CPU cores, plus a nominal 2W for four icestorm cores. Apple will invariably want to increase cache sizes for its mac parts, putting further pressure on the APU. Ouch.

That makes pushing clocks higher a nonstarter for the MBP14, which is going to have to pare back watts / clock speeds as more of the APU comes online and may also need a couple firestorm cores lasered off. I think Apple can still get to the point where they can advertise 3GHz (on N5P), but multicore will sustain at lower than that. I also still think it's fine to leave around 12-16 GPU cores on the die - I'd prefer to see an APU with 6 firestorm cores and 16 GPU cores.

The MBP16 is more comfortable throwing 4 or 5W at 8 cores to sustain speeds of up to 3.2GHz on N5P. It can still use the same part as the MBP14, prioritizing good CPU yield and being perfectly happy to disable 4 GPU cores. I'm absolutely against it trying to accommodate all of its GPU needs on a monolithic die now - it needs a separate GPU part, or at a minimum, a separate GPU die.

 

[Unregistered 4U]

I think Apple can still get to the point where they can advertise 3GHz (on N5P), but multicore will sustain at lower than that.

I don’t think they’ll even advertise a GHz speed rating, though, and will do what they’ve done on iOS. Because, they know that what’s far more important to potential buyers is how much faster it is than their prior Mac.

 

[burgerrecords]

My guess is when Apple kicked Intel CPUs back to the curb, they also gave notice to AMD graphics.

that doesn’t necessarily mean it’s going to perform better in absolute terms, it may mean that it is just going to be “better” (better performance on battery power and better margins for Apple)

 

[s66]

a Mac CPU on HP node doesn't make much ROI.

Apple is going to be looking at a bigger picture and a more long term one, where any research in GPU, CPU and all the surrounding stuff would (eventually) happen anyway for their phones and tablets. (While apple does look like any stock market funded company at the short term for results, they have a remarkable ability to do long term focussed investments and research as well).
Moreover they will consider the investment as one into their ecosystem, not one into a CPU or GPU business (a business they are not interested in being in), and as such the ROI is always going to be there as it will allow them to gain a competitive advantage for their entire product line.
And the final one is being dependent on roadmaps from intel and AMD (which are not as confidential as Apple likes it to be) and then them failing to deliver on their promises (which hurts Apple's plans), as well as the lack of Intel to continue to do the special treatment Apple got when Jobs moved them from the G5 to Intel ... They will want to break with those 3rd parties they cannot control for reasons of getting unique advantages (that are not easy to steal or copy) as well as secrecy and full control over a roadmap. Those advantages are worth a lot of money in the end.

 

[leman]

I'm absolutely against it trying to accommodate all of its GPU needs on a monolithic die now - it needs a separate GPU part, or at a minimum, a separate GPU die.

Why? A 70-80 watt SoC is trivial to cool nowadays. What would be the advantage in splitting the components up? You’ll only end up with a larger, less performant and less energy efficient system.

 

[ksec]

Apple is going to be looking at a bigger picture and a more long term one, where any research in GPU, CPU and all the surrounding stuff would (eventually) happen anyway for their phones and tablets. (While apple does look like any stock market funded company at the short term for results, they have a remarkable ability to do long term focussed investments and research as well).
Moreover they will consider the investment as one into their ecosystem, not one into a CPU or GPU business (a business they are not interested in being in), and as such the ROI is always going to be there as it will allow them to gain a competitive advantage for their entire product line.
And the final one is being dependent on roadmaps from intel and AMD (which are not as confidential as Apple likes it to be) and then them failing to deliver on their promises (which hurts Apple's plans), as well as the lack of Intel to continue to do the special treatment Apple got when Jobs moved them from the G5 to Intel ... They will want to break with those 3rd parties they cannot control for reasons of getting unique advantages (that are not easy to steal or copy) as well as secrecy and full control over a roadmap. Those advantages are worth a lot of money in the end.

1. Even without HP Node they are still going to have the advantage. Just not as much.
2. At $200 to $500M on a redesign, and a small volume of those targeted market, ( <1M ), that is a cost of *minimum* $200 to $500 per unit alone, *excluding* wafer cost.

Could they afford it? Of course. Apple gets $10B from Google *every year* just for default search engine placement.

Would they do it? Not so sure. Apple has a culture of tight budget even in Steve Jobs era.

 

[awesomedeluxe]

Why? A 70-80 watt SoC is trivial to cool nowadays. What would be the advantage in splitting the components up? You’ll only end up with a larger, less performant and less energy efficient system.

We’re still talking about the MBP 16 right? Can you explain the source of your optimism... why do you consider cooling a 70-80W SoC inside a notebook the size of the MBP 16 (or smaller, in the likely event Apple opts for a thinner redesign) “trivial”?

 

[leman]

We’re still talking about the MBP 16 right? Can you explain the source of your optimism... why do you consider cooling a 70-80W SoC inside a notebook the size of the MBP 16 (or smaller, in the likely event Apple opts for a thinner redesign) “trivial”?

Because a) Apple already does it (they cool a 45W CPU and a 50W GPU) and b) there are laptops of similar dimensions that host 80+W GPUs.

But let's look it it in terms of available facts. The chassis definitively has the capacity of dissipating 80+ watts. The question is then whether it can be efficiently dissipated from a relatively small surface area. I am far from being an engineer and my physics days are long gone, but professional sources suggest that modern cooling elements have thermal flux in excess of 100W per cm2 (ant potentially much higher than that). Assuming the Mac SoC die will have area between 1cm2-4cm2, it sounds absolutely feasible to me.

 

[awesomedeluxe]

I have failed to keep heat/area in my thoughts and prayers. I do not believe in heat/area. The existence of the Razer Blade proves that thermodynamics aren't real.

I assume you are talking about the Razer Blade's GeForce 2070 Super Max-Q. The Razer Blade uses your vapor chamber technology, and its GPU does use about 80W in a chassis that is only a hair bigger than the Macbook Pro. It is not trivial to cool. The Blade struggles to cool this device, its fan is hot and loud and runs incessantly, and reviewers hate the thing because honestly what were they thinking?

The kicker, though, is again heat/area. The GeForce 2070 Super Max-Q is 445mm2. An ~80W Apple APU with a 16+4-and-42 design would be around 235mm2. The same power consumption in half the space of the GeForce, which is only the start of our problems since our cotenants are no longer other well-behaved GPU cores operating in parallel. Our GPU cores now live with CPU cores that like to self-immolate whenever work gets hard.

Basically, what you've seen is a device the size of a MBP16 barely succeed at cooling a 445mm2 GPU that uses 80W by using vapor chamber tech. This does not mean that by using vapor chamber tech we can cool a 235mm2 APU using 80W. It is true that the MBP will have a cooler interior since there is not an outstanding CPU somewhere, but that's not going to stop it from thermal throttling. We can either separate the CPU cores and GPU cores out somehow or we can just settle for an APU that uses less power.

 

[leman]

Basically, what you've seen is a device the size of a MBP16 barely succeed at cooling a 445mm2 GPU that uses 80W by using vapor chamber tech. This does not mean that by using vapor chamber tech we can cool a 235mm2 APU using 80W.

First of all, I didn’t write that text you are quoting, so there might have been some creative editing involved.

Second, I think you are confusing cooling (dissipation) with heat transfer. Vapor chamber does not cool anything. It’s purpose is to transfer the heat from a hot component to the radiator (the actual cooler), where that heat gets distributed over a large surface area and then expelled from the system via convection.

So there are two questions: a) can the laptop house a cooling system (radiator) big enough to remove the heat and b) can the laptop effectively transfer the heat from the hot area to the radiator. If the radiator is not large enough, the entire system will overheat. If the heat is not transferred fast enough (insufficient heat flux), the heat will stay with the hot components and overheat them. Modern heat transfer technology (e.g. a vapor chamber) should not have any issues transferring 100W of power from a component that only measures 2-4cm2. I don’t think that the Blade has any issues with heat transfer, but it might certainly have an issue with heat dissipation. After all, it has to dissipate more than 100W (CPU+GPU) in that compact chassis.

To sum it up: the problem with the Blade is not getting the heat away from the components but getting it out of the system. The MBP won’t have that problem. We know for a fact that 16” MBP can confidently dissipate 80Watts. Again, I don’t think that transferring that heat will be a problem, as modern heat transfer components are supposed to have heat flux in excess of 100W/cm2

 

[Waragainstsleep]

I do suspect that Apple would prefer its laptops to run cooler than they currently do.

 

[awesomedeluxe]

First of all, I didn’t write that text you are quoting, so there might have been some creative editing involved.

Second, I think you are confusing cooling (dissipation) with heat transfer. Vapor chamber does not cool anything. It’s purpose is to transfer the heat from a hot component to the radiator (the actual cooler), where that heat gets distributed over a large surface area and then expelled from the system via convection.

So there are two questions: a) can the laptop house a cooling system (radiator) big enough to remove the heat and b) can the laptop effectively transfer the heat from the hot area to the radiator. If the radiator is not large enough, the entire system will overheat. If the heat is not transferred fast enough (insufficient heat flux), the heat will stay with the hot components and overheat them. Modern heat transfer technology (e.g. a vapor chamber) should not have any issues transferring 100W of power from a component that only measures 2-4cm2. I don’t think that the Blade has any issues with heat transfer, but it might certainly have an issue with heat dissipation. After all, it has to dissipate more than 100W (CPU+GPU) in that compact chassis.

To sum it up: the problem with the Blade is not getting the heat away from the components but getting it out of the system. The MBP won’t have that problem. We know for a fact that 16” MBP can confidently dissipate 80Watts. Again, I don’t think that transferring that heat will be a problem, as modern heat transfer components are supposed to have heat flux in excess of 100W/cm2

Just a bit of levity. I think we will just have to agree to disagree on this one. Although I want to correct that a 235mm2 part would not be "a component that only measures 2-4cm2." It would be smaller. e.g., a 1.5cm2 part would be 15mm * 15mm = 225mm2.

I do understand better why you believe an 80W SoC would be able to transfer heat quickly enough that it would not be an issue. I just have strong doubts that this is the case. I think a 1.5 centimeter squared part using 80W of power in a cage the size of the MBP16 will probably start to fail very shortly after warming up and that nothing short of magic will prevent that from happening. I think it could throttle its way to survival in an iMac and it might be able to reach its full potential in an iceberg.

Frankly, whether Apple debuts a 50W or 80W APU, it will necessarily be the first appearance of an APU with this level of power density. So we may just have to see what is possible.

 

[leman]

I think we will just have to agree to disagree on this one.

I’m perfectly fine with that After all, we are just speculating, with no hard facts on our hands. We can always revisit it once the products is out to see who’s prediction was correct if we do want.

Although I want to correct that a 235mm2 part would not be "a component that only measures 2-4cm2." It would be smaller. e.g., a 1.5cm2 part would be 15mm * 15mm = 225mm2.

15x15mm does not make 1.5cm2, you’d want something like 10x15mm for that. 1cm2 equals 100mm2, so 225mm2 is 2.25cm2

 

[awesomedeluxe]

Oh, you're right. I backed out of the ^2 to the width/length and then went the wrong way back into ^2 area. Not sure why I did that. But yes - I look forward to looking back on these conversations when we finally have some real info and seeing what we got right and wrong.