The Raspberry Pi 4 CPU Explained
Q: Why do you use ARM processors and not an x86 processor like Intel? Why not partner with AMD? Ryzen on Raspberry!
A: It's an interesting idea, primarily because it would allow us to run "regular" Windows. But that's less useful than it once was, and the power/price/performance/area advantages of the ARM architecture are pretty convincing.
Q: Could you expand more on “the power/price/performance/area advantages” of ARM processors? What is the performance increase? I thought x86 was industry standard because it was the most powerful. I am a gamer, so Intel and AMD is all I know.
A: So, you can get higher total performance designs from Intel and AMD, but ARM designs win out if you normalise for performance per Watt (useful if you're a thermally constrained platform like us) or per square millimetre (and thus per dollar - useful if you're a cost constrained platform).
The argument is that there's less instruction set cruft to worry about in ARM land, so ARM cores can be leaner and meaner. Not sure I totally buy that (AArch32 is plenty crufty by now), but nobody's ever really fielded a decent low-power x86 implementation, so maybe there's something to it.
Q: The current power usage is a lot higher when compared to previous generations, probably because the way faster out of order CPU cores. Did you guys experiment with heat sinks and metal casings (used as a heat sink)?
A: The intention is that "regular" users will be okay without an additional thermal solution, but I think that FLIRC-type metal cases will be useful for power users.
MORE: Raspberry Pi 4 Firmware Update Available: Less Heat, Same Speeds
Q: With the movement from 40 to 28 nanometer production what advantages have you been able to capitalize on? Would costs increase considerably if moving to a lower power node such as TSMCs 14nm?
A: 28nm brings quite substantial improvements in energy. It's also the current "value node" (lowest cost per transistor), and I think will remain so for the foreseeable future. We would currently expect unit costs to increase if we moved to a smaller (16nm, 12nm, 7nm) process. This won't be the case forever, but I could see us staying on 28nm for another five years at this point.
MORE: Overclocking the Raspberry Pi 4 - Tom's Hardware
Q: Was a FinFET process considered and if so what killed it for this iteration (development cost, marginal costs, available IP, etc.)?
A: A little of all of those, but primarily projected unit costs.
Q: How much is the Pi ecosystem locked in on Broadcom? Will an Amlogic or Realtek chip in the future be possible for example?
A: I think we're happy that the Broadcom silicon offers the right feature set, and of course total backward compatibility with earlier models.
Q: Do you have any thoughts on RISCV and if it can potentially be used on your products?
A: It's exciting, and I think there are probably near-term applications in the microcontroller space, but we're quite a long way from having a properly mature software ecosystem, or licensable cores in the A72 class, so I don't expect to see a RISC V Raspberry Pi in the near future.
