They went from 4 to 5 decoders. Rather than 1-2 complex decoder and a few that only work on simple instructions, every decoder can now decode everything, I-cache handles fusion.
This all happens with a 75% reduction in decoder size due to dropping more complex 32-bit encoding. As the decoder is always on to feed the speculative execution, this is undoubtedly a substantial decrease in power consumption.
x86 decoder is 8-10% of the total die area. If they could reduce that to 2%, that would represent 6-8% smaller core size and at least 6-8% reduction in power consumption (likely more as other areas will stay power-gated much more of the time).
This is a more general point. X2 in the Qualcomm 8cx gen 3 beats Zen 1 desktop chips and is rather close to Zen 2 U-series chips in single and multicore geekbench scores. It uses 7-10w peak while those U-series use 30-50w peak which is why despite so much battery being used for stuff like wifi and the screen, thin and light laptops with the 8cx have generally gotten 2x the battery life.
I wonder how many of these ARMv8/9 chips will have to beat x86 in all these metrics before people will finally be convinced that ISA really does matter (or both AMD and Intel are utterly incompetent). Maybe it will take Intel finally releasing a RISC-V chip to change minds.
I also believe that RISC-V is the future. It would be great to have implementations that can compete with ARM at least CPU wise first. Then we will need GPU's and eventually co-processors which are optimized for IO, media, machine learning, ... to create the full package. Also a new memory standard would ve great because I don't see on-chip memory being economic like Apple is doing it right now.
What prevents me from getting an ARM laptop with higher end SoCs is Linux support. I can install Ubuntu or Fedora immediately when I buy pretty much every x86 laptop and expect everything to work, but I am not as confident about ARM laptops (PineBook family excepted, but they have some limits). Having manufacturers come forward and offer fully supported mainlined Linux models would work wonders.
I wouldn't be surprised if Apple's ARM MacBooks get there first. A lot of hardware works already, and it seems like there has been a lot of progress on the GPU too.
I don't find that very appealing. MacOS is a good Unix and I only start Linux containers on it when I'm testing something specifically for Linux (which I usually consider a bug - the software I write should run well on anything Unixy enough).
could someone explain how the Arm product pipeline works?
This is my (probably totally wrong) understanding:
Arm makes chip specification and defines assembly language --> companies (Samsung, Apple) license both of these and design their own chips following those specs --> these designs are sent to fabs and produced.
But what I don't get is, what "wiggle room" do companies like Apple and Samsung have to make their chips special? Obviously the M1 is a different beast from a Samsung chip, but they are both "Arm". So what is Apple able to do that makes the M1 the M1, while still also being "Armv8", the same "Armv8" as a Samsung chip? Does "Armv8" only mean "it must accept this instruction set, the rest is up to you"? Or is there requirements at the silicon level as well? Is this new Corex-A715 simply a "reference" chip for the next gen of the Arm instruction set?
ARM produces both architecture specifications and the designs for chips that implement those specifications.
Some companies license the architecture specification and design their own implementations, Apple being the leading example. They then pay TSMC to fabricate the chips for them.
Other companies license the implementations and pay to have them fabricated. The latter category varies in the level of customization they apply to the implementations they license, it's generally much less customized than the what those in the first category achieve.
And Apple itself has a special architecture licensing relationship as one of the founding parties behind ARM Inc. Their $3MM investment ended up being a great deal for them. They can do whatever they want with the platform.
A lot of the customization is in the peripherals. ARM (AFAIK) really only specifies the processor, modern devices have lots of other on-board signals.
The instruction set really only defines how to run mathematical operations on the data on its bus, what to write to registers, and what instructions should load and store memory to various pages on the memory map.
Licensors add peripherals to do things that the instruction set doesn't define, like how to toggle an output pin from low to high. You use the instruction set to set a bit in a particular location in memory, and that turns on an LED attached to that pin - or they make a handy peripheral that lets you write to a word, because setting one bit in a 32-bit memory value is a read-modify-write 3-step operation and many of their devices have thousands of bytes of storage and RAM in an address space with billions of memory locations. They add I2C, SPI, CAN, Ethernet, USB, EEPROM, Flash, SRAM, video, etc. etc. etc. to make useful systems on chip - the processor is only one small part of that equation.
>Why don't phone makers optimize for power efficiency?
They do. I don't know what misunderstandings could lead you to believe that phone makers don't already spend significant effort and resources on power efficiency.
And memory is also a large user of power budget. This is exactly right… larger and brighter screens, better wifi, BT, and cellular systems, more memory, and higher system speeds (because you know that 20% more efficient processor is X% faster offsetting that efficiency) all eat up that power budget.
That's because, if the chip uses 20% less power, if the latest Soc performs the same as last year but consumes 1w less power no one will buy it. So Qualcomm has to increase its power limits to show graphs like 50% FASTER WITH SIMILAR POWER CONSUMPTION
Personal anecdote:
the lower end chips like SD695 and SD732 have amazing battery lives. I personally use a SD732G smartphone with a 6000MAH battery and it gives me amazing 3-4 days of battery on one full charge
Phone makers spend a lot of time worrying about power. But because your screen and cell radio take up a minimum amount of power there's a level of diminishing returns in optimizing the application cores beyond a certain point.
It's a cycle. You get cool chips, then hot chips, then cool chips. For example, the Snapdragon 800 was a hot chip, then 801, 805 chips ran cool then the 810 had terrible heat problems. We're in a "hot" cycle right now with 888 and 8 Gen 1 phones typically having significant throttling problems.
Because while the CPU gets more efficient, other parts go the other direction? Higher resolution screens, more sensors, radios with harder to implement requirements, and so on.
Most likely in my mind because phone makers would like you to purchase a new phone every 2 years, and a new battery only begins to degrade around that time as well.
Eh you seem to like conspiracy theories. Not to say there aren't conspiracies out there, but for some people this is the first and only explanation for everything. As you can see in the other comments there are technical reasons for what the phone makers are doing.
I mean, I enjoy using Apple products, but they quite literally paid out a very large lawsuit which found them culpable for planned obsolescence with iPhone 5 batteries and etc.
Each minor version of ARMv8 has a lot of optional extensions. What ARMv9 does is make some of those features mandatory. ARMv9 also includes new optional features such as Scalable Vector Extensions 2 (SVE2) and Confidential Compute Architecture (CCA) (Intel SGX / AMD SEV equivalent).
This all happens with a 75% reduction in decoder size due to dropping more complex 32-bit encoding. As the decoder is always on to feed the speculative execution, this is undoubtedly a substantial decrease in power consumption.
x86 decoder is 8-10% of the total die area. If they could reduce that to 2%, that would represent 6-8% smaller core size and at least 6-8% reduction in power consumption (likely more as other areas will stay power-gated much more of the time).
This is a more general point. X2 in the Qualcomm 8cx gen 3 beats Zen 1 desktop chips and is rather close to Zen 2 U-series chips in single and multicore geekbench scores. It uses 7-10w peak while those U-series use 30-50w peak which is why despite so much battery being used for stuff like wifi and the screen, thin and light laptops with the 8cx have generally gotten 2x the battery life.
I wonder how many of these ARMv8/9 chips will have to beat x86 in all these metrics before people will finally be convinced that ISA really does matter (or both AMD and Intel are utterly incompetent). Maybe it will take Intel finally releasing a RISC-V chip to change minds.