That is, the reason you don't use battery backed DRAM for all of your photos is not because you don't want to, but because 8TB of the stuff would be very expensive. And so most of us have RAM leading to SSD leading to spinning platters.
So the reason to have a CPU cache at all is (insert interesting explanations of caches here). But the reason to have more than one CPU cache is because of the relative cost of the first cache, right ?
If cost was no object, wouldn't you just have a huge primary cache ?
Speed is a big issue, as well as address space. Caches can't be huge and really fast, partly because a fast cache is crazy power hungry, and a big cache has large distances to travel. So your on-die caches are hierarchical in part so that you can have a fast cache and a big cache.
The second problem, address space- larger address space requires more addressing bits, which makes the CPU bigger, and bigger means slower. This constrains cache as well as memory. The final level, disk, has that as an advantage over DRAM because it is not memory-mapped. It's not frequently a problem, but for example Nehalem, despite being a "64-bit" architecture, actually only supports 44-bits of address, setting a hard limit at 16 tebibytes of RAM (not considering MMIO etc)
Not really, SRAM(what most caches use) isn't just more costly is also consumes a lot more power. You'd be hitting thermal limits much sooner than with traditional DRAM.
Is that right? Traditionally DRAM has used much more power because it requires a periodic refresh whereas the power consumption of SRAM is purely leakage. Now, maybe leakage power has grown so much in recent years that this is no longer true but if so I'd find that very surprising. Do you have any numbers?
In SRAM you get Leakage + Switching current. For cases like a cache where you're going to be constantly churning bits this can drive up power quite a bit.
I don't claim to be an expert so there are probably cases where it's lower. However when I was looking into fast RAM for storing data from an FPGA for a simple Logic Analyzer most of the SRAM was almost 2-4x what DRAM was for power consumption at the same storage size(with SRAM being a blazing fast cycle access).
A hardware project I worked on some years ago included a chip that contained a small SRAM and a lithium battery, all soldered right onto the board, as a form of nonvolatile storage. Some SRAM can indeed be very low-power, but caches are probably a different animal.
That is, the reason you don't use battery backed DRAM for all of your photos is not because you don't want to, but because 8TB of the stuff would be very expensive. And so most of us have RAM leading to SSD leading to spinning platters.
So the reason to have a CPU cache at all is (insert interesting explanations of caches here). But the reason to have more than one CPU cache is because of the relative cost of the first cache, right ?
If cost was no object, wouldn't you just have a huge primary cache ?