TIL: Why are M1 Macs so fast? π€
Quick, bite-sized thread that answers the above question. Based on my notes from reading: debugger.medium.com/why-is-apples-β¦
Thread π§΅π
Quick, bite-sized thread that answers the above question. Based on my notes from reading: debugger.medium.com/why-is-apples-β¦
Thread π§΅π
as per the benchmarks, M1 is beating almost every available processor in the market.
Stats (from Apple):
- 3.5x faster CPU perf (vs i7 Mac Air)
- 3x CPU perf / watt
- 2x faster GPU
- 15x faster ML (vs i3 Mac)
(will link benchmarks at the end)
Stats (from Apple):
- 3.5x faster CPU perf (vs i7 Mac Air)
- 3x CPU perf / watt
- 2x faster GPU
- 15x faster ML (vs i3 Mac)
(will link benchmarks at the end)
I'm assuming familiarity with basic concepts of computer architecture.
Otherwise, you can read about them in my previous thread:
Otherwise, you can read about them in my previous thread:
https://twitter.com/vermayash8/status/1376173730947244036
M1 isn't just a processor, it's a System on Chip (SoC),
It includes:
π« processor cores
π I/O
πΉ GPU
π§ Neural Engine
ποΈ memory
Unlocks simplicity, more efficiency and amazing performance.
Traditional processors have cores. Rest are separate components via the motherboard.
It includes:
π« processor cores
π I/O
πΉ GPU
π§ Neural Engine
ποΈ memory
Unlocks simplicity, more efficiency and amazing performance.
Traditional processors have cores. Rest are separate components via the motherboard.
β
Reason 1:
Instead of having only general purpose CPU cores, M1 has specialized chips inbuilt.
Basically, for almost every workload β M1 has a guy who does one thing, but does it really well and quick! π
GPU, Neural Engine, ML accelerators.
Instead of having only general purpose CPU cores, M1 has specialized chips inbuilt.
Basically, for almost every workload β M1 has a guy who does one thing, but does it really well and quick! π
GPU, Neural Engine, ML accelerators.
β
Reason 2:
Unified Memory Arch. (UMA)
- Usually, CPU and GPU have different memories or diff memory segments.
- workload handover means CPU gotta copy it to GPU's space.
UMA:
- Specially designed - highβbandwidth, lowβlatency memory
- shared memory space, no copying
Unified Memory Arch. (UMA)
- Usually, CPU and GPU have different memories or diff memory segments.
- workload handover means CPU gotta copy it to GPU's space.
UMA:
- Specially designed - highβbandwidth, lowβlatency memory
- shared memory space, no copying
For those of you wondering why were these separate in the first place:
- different data consumption patterns - CPUs want less data real quick (low latency). GPUs want more data at once (high bandwidth).
- GPUs produce too much heat.
- different data consumption patterns - CPUs want less data real quick (low latency). GPUs want more data at once (high bandwidth).
- GPUs produce too much heat.
β
Reason 3:
Out of Order Execution (OoOE) abilities of ARM CPU
- Each core executes multiple instructions in parallel.
- note that this isn't multithreading
M1 Firestorm cores process twice as many instructions as x86. (assuming same clock freq. and same size instructions)
Out of Order Execution (OoOE) abilities of ARM CPU
- Each core executes multiple instructions in parallel.
- note that this isn't multithreading
M1 Firestorm cores process twice as many instructions as x86. (assuming same clock freq. and same size instructions)
How OoOE works?
[very high level]
CPU analyzes dependencies between instructions, if output from 1 is not affecting 2, then 2 can be executed ahead of time.π€―
Invisible to the developer / end user.
[very high level]
CPU analyzes dependencies between instructions, if output from 1 is not affecting 2, then 2 can be executed ahead of time.π€―
Invisible to the developer / end user.
Why can't Intel and AMD do these?
- x86 arch. does OoOE, however it's inferior because of limit on number of decoders in a single processor.
- this limit is present because of variable size of instructions in CISC (1-15 bytes), therefore, start and end are difficult to identify
- x86 arch. does OoOE, however it's inferior because of limit on number of decoders in a single processor.
- this limit is present because of variable size of instructions in CISC (1-15 bytes), therefore, start and end are difficult to identify
Moreover,
- SoCs are not just processors, intel does not make other stuff, companies like HP, Dell do.
- SoC world works by buying Intellectual property (designs) e.g. ARM arch and putting them together.
- Intel is unlikely to lend it's IP to 3P companies.
- SoCs are not just processors, intel does not make other stuff, companies like HP, Dell do.
- SoC world works by buying Intellectual property (designs) e.g. ARM arch and putting them together.
- Intel is unlikely to lend it's IP to 3P companies.
Apple, controls everything in their ecosystem end to end. Every piece of hardware, software, sometimes including ML libraries which provide abstractions.
Freedom to innovate!
Steve Jobs πββοΈ
Freedom to innovate!
Steve Jobs πββοΈ
That's all, thanks for reading π
If you found this helpful, please like and retweet for better reach. β€οΈ
Thank you!
If you found this helpful, please like and retweet for better reach. β€οΈ
Thank you!
β’ β’ β’
Missing some Tweet in this thread? You can try to
force a refresh
γ
