3 ms·

Most of the performance advantage of Apple silicon is that it puts CPU, GPU and all memory (that is typically VRAM and DRAM in traditional architectures) millim

Most of the performance advantage of Apple silicon is that it puts CPU, GPU and all memory (that is typically VRAM and DRAM in traditional architectures) millimeters from primary DRAM. When competing purveyors start do this with their own clever packaging, they will start to approach Apple’s mips/watt. Emulation on M1 is outstanding. The spatial locality of main memory nullifies emulation penalty. Apple has come close to what Transmeta tried to do.

wmf 4y ago
Nope. There's little or no advantage to 2 mm vs. 20 mm.
- tdsanchez 4y ago
  When multiplied over billions of clock cycles it makes a tremendous amount of difference. There's a reason why PCB board specs say that all DRAM traces need to be the same length - the length of the trace, and it's capacitance are the main speed limiter for any signal. This is true whether the trace is 20 mm long or 2. "the critical point to DDR circuitry is in the timing. This means that the clock lines will determine the timing of the input and output of the memory chip data lines. Electrical signals can not be sped up, so the only way to control the timing is to add length to the traces in order to delay the arrival of the clock signal. This is accomplished by adding serpentine routing to the trace routing." https://resources.pcb.cadence.com/blog/2019-ddr-routing-techniques-in-your-pcb-design https://resources.pcb.cadence.com/blog/2019-ddr-routing-tech... "So the physical RAM modules are still separate entities, but they are sitting on the same green substrate as the processor. “Big whoop,” I hear you saying. “What’s the big deal?” Well, first of all, this means faster access to memory, which inevitably improves performance. In addition, Apple is tweaking how memory is used within the system." https://www.howtogeek.com/701804/how-unified-memory-speeds-up-apples-m1-arm-macs/ https://www.howtogeek.com/701804/how-unified-memory-speeds-u...

wmf 4y ago

Nope. There's little or no advantage to 2 mm vs. 20 mm.

tdsanchez 4y ago
When multiplied over billions of clock cycles it makes a tremendous amount of difference. There's a reason why PCB board specs say that all DRAM traces need to be the same length - the length of the trace, and it's capacitance are the main speed limiter for any signal. This is true whether the trace is 20 mm long or 2. "the critical point to DDR circuitry is in the timing. This means that the clock lines will determine the timing of the input and output of the memory chip data lines. Electrical signals can not be sped up, so the only way to control the timing is to add length to the traces in order to delay the arrival of the clock signal. This is accomplished by adding serpentine routing to the trace routing." https://resources.pcb.cadence.com/blog/2019-ddr-routing-techniques-in-your-pcb-design https://resources.pcb.cadence.com/blog/2019-ddr-routing-tech... "So the physical RAM modules are still separate entities, but they are sitting on the same green substrate as the processor. “Big whoop,” I hear you saying. “What’s the big deal?” Well, first of all, this means faster access to memory, which inevitably improves performance. In addition, Apple is tweaking how memory is used within the system." https://www.howtogeek.com/701804/how-unified-memory-speeds-up-apples-m1-arm-macs/ https://www.howtogeek.com/701804/how-unified-memory-speeds-u...