There is no such thing as an " Apple A7", Apple just gets others to make something for them to their specs. They have no fabs and no expertise whatsoever in designing processors.
I don't think L1 are correctly placed.Tentative A7 analysis:
We publish this with the caveat that these are best guesses we have not done any real circuit extraction to confirm them. The dual-core CPU and cache make up ~17% of the die area, and the quad-core GPU and shared logic about 22%. The CPU itself is not packed the same way as the A6 (see below), it looks much more like a conventional automated layout;
I don't think L1 are correctly placed.
They quite obviously are not. I don't know how someone at Chipworks made this kind of mistake. The blocks just diagonal of the L2 and to the opposite side of core are the L1 caches. My guess is the one closer to the L2 is dcache. Stuff just to the left of those blocks can include tags, TLB entries, and buffers. Area in the center to the right of the L2 SRAM arrays would be L2 tags.
A basic question on the power of the Apple A7.
On Geekbench 3, the A7 scores around 1400 for a single core (@1.3GHz). The Intel Core i7-3840QM scores around 3,500 on a single core basis (@2.8Ghz). The TDP of A7 is 1.5-2W and the TDP of the core i7-3840QM is around 45W (but then people say that is only for the CPU, the integrated GPU is not included).
So, theoretically Apple can double the A7 frequency to 2.6GHz and the performance will increase to 2,800 per core with power quadrupling to 6-8W TDP, so basically getting 80% of core i7 score with 1/6th the power consumption. I realise, you can't just double the clock rate and there are other factors, but does that show you how close the A7 architecture is to the best that Intel has got. By the way, the Core i7-3840QM suggested retail price is $568.
If that is the performance that Apple gets from 28nm planar design, then clearly it can take the same design, ramp up the clock rate in 20nm next year and get a processor with single core performance equal to the best of Intel?
Is the above broadly correct?
If it is, then what stops apple from having two different (or 3) processor design teams geared to deliver cores for different TDPs. One for phones (lowest TDP), another for macbook pro and air and the third for the imacs (the highest power consumption levels)?
The problem, as already shown in other threads, is that geekbench is not a good way of showing cross platform performance. In short, the real world performance is lower on the A7 than geekbench shows vs for example x86. The A7 "magic" is already dead.
The problem, as already shown in other threads, is that geekbench is not a good way of showing cross platform performance. In short, the real world performance is lower on the A7 than geekbench shows vs for example x86. The A7 "magic" is already dead.
do you have a link to the tdps? I havent seen them anywhere.
From what I hear, Intel measures TDP of only the CPU cores and not the integrated GPU...
Doubling frequency is not an easy task if your design has been thought for that from early in the design process. We don't know what Apple did, but given the apparent efficiency on Geekbench, they went for high perf/MHz instead of high frequency (or A7 frequency is already higher than the 1.3GHz found).Well, everyone says that, but no one has an alternative. I am just using Geekbench as a substitute for performance. But my main point is: can't apple (or any of the ARM design guys, on their 64bit ARM designed processors)
1) Double the clock rate of the A7 (or its soon to be born brothers from qualcomm etc)
No, high performance doesn't imply x86 for sure. But it took Intel many years and iterations to reach the level they are at. I doubt Apple will be able have a competing chips before 2 or 3 years (assuming one iteration per year), and even then Intel will have moved on.2) Double the GPU area (which is more scalable anyways)
3) Move to a new process node (which is going to happen anyways)
and bring out a processor which is comparable in performance to the core i7 series Intel processors but much lower in power consumption?
After all, the best of breed intel core i-7 processors have south of 1.5bn transistors and the A7 has 1bn transistors... ARM architecture can't be so sh*t compared to x86 that they can't have at least comparable performance!
Which are the real world measurements you're referring to? Please don't list synthetic/scripted browser benchmarks, benchmarks depending on some single vendor optimized execution engine, or compiler-biased ones (e.g. GCC vs. ICC).The problem, as already shown in other threads, is that geekbench is not a good way of showing cross platform performance. In short, the real world performance is lower on the A7 than geekbench shows vs for example x86. The A7 "magic" is already dead.
Which are the real world measurements you're referring to? Please don't list synthetic/scripted browser benchmarks, benchmarks depending on some single vendor optimized execution engine, or compiler-biased ones (e.g. GCC vs. ICC).
What about getting proofs instead of assumptions? Can you show us code that hints at clang cheating at some benchmark?Compiler biased? So anythign compiled with clang is Apple biased I assume.
Doubling frequency is not an easy task if your design has been thought for that from early in the design process. We don't know what Apple did, but given the apparent efficiency on Geekbench, they went for high perf/MHz instead of high frequency (or A7 frequency is already higher than the 1.3GHz found).
No, high performance doesn't imply x86 for sure. But it took Intel many years and iterations to reach the level they are at. I doubt Apple will be able have a competing chips before 2 or 3 years (assuming one iteration per year), and even then Intel will have moved on.
All of the above is pure speculation
That's probably true because AMD will be using an existing CPU design, instead of designing their own. Designing a CPU from the ground up is probably as expensive for anyone as it is for Intel, when you're targeting performance.1) AMD said that one of the reasons they are signing the architecture licensing deal with ARM is because of the ease of designing with ARM. From what I remember reading, it cuts down the time to design significantly and the number of people by ~1/5th ish
ARM 32-bit instruction set has been accumulating stuff for 20 years too, though certainly not anything as bad as x87 :biggrin:2) Legacy is a burden as well as a benefit. Intel has benefited from the status of x86 in legacy systems today and will continue to benefit. But that also brings the burden of carrying on with bits of inefficient legacy blocks in their designs. Whereas ARM had the advantage of designing a 64bit architecture from the ground up, only having to maintain backward compatibility with their 32bit arch, not for stuff from 20 years ago (which is the case with x86)
From a high performance chip design point of view, I am sorry but they areSo, don't for a moment assume that Intel and the ARM design camp are on a level playing field. They are not!
From a high performance chip design point of view, I am sorry but they are
!
Yes, that is difficult, very difficult. In fact high frequency and low frequency mean different micro-architectures.Designing a chip from the ground up will be difficult. But modifying the A7 so that it can operate at higher frequencies - is that so difficult???
There could be a) simple reuse of existing Cortex macros, b) a modified design (also saves work), c) a redesign from ground up.1) AMD said that one of the reasons they are signing the architecture licensing deal with ARM is because of the ease of designing with ARM. From what I remember reading, it cuts down the time to design significantly and the number of people by ~1/5th ish
There could be a) simple reuse of existing Cortex macros, b) a modified design (also saves work), c) a redesign from ground up.
A LinkedIn profile suggests, that AMD is working on some modifications. The owner mentioned the ARM scheduler.
But my main point is: can't apple (or any of the ARM design guys, on their 64bit ARM designed processors)
1) Double the clock rate of the A7 (or its soon to be born brothers from qualcomm etc)
2) Double the GPU area (which is more scalable anyways)
3) Move to a new process node (which is going to happen anyways)
Yes, that is difficult, very difficult. In fact high frequency and low frequency mean different micro-architectures.
This comes from the fact that at a given frequency your electrons can travel through let's say 30 layers of logic per cycle; if you start increasing frequency, given that electron speed is constant, you'll have less layers of logic, hence you'll be able to do less work (that's a simplification, but it's close enough to reality).
So if for instance, Apple had time to multiply two numbers because this requires 30 layers of logic, if they increase freq, which implies less layers, then they'll have to split multiplication into two cycles.
Hope I made this clear enough...