I know this is a bit off-topic and probably should not be on the console forum, but can someone quickly remind me why RAM chips have historically had much larger transistor counts than CPUs and GPUs, while their clocks are lower? Quickly so we don't derail this too much? Please?
I'm trying to dredge up some recollection and citations, but the general trend stems from the different way DRAM performs its function, and a very heavy emphasis on density and cost.
DRAM relies on using densely packed arrays composed of many rows of bit cells composed of a capacitor and a controlling transistor. Hundreds of these cells are attached to a common bit line, and a read is accomplished by one of the hundreds of transistors on a single bit line being turned on (many bit lines are read in parallel), and the capacitor either discharging its stored charge or taking some charge from the bit line, depending on the value of the cell. This small perturbation is picked up by a sense amplifier, which will work to raise or lower the bit line's voltage to a readable logical 0 or 1.
The bit line's is physical and electrical bulk needs time to be overcome, but until it is, the array has little use.
Some trends:
- The more cells on the wire, the less space is lost to overhead. However, the more cells on a wire, the longer it needs to be physically, which decreases the amount a capacitor can affect its voltage, and increases the time to sense it.
- A node shrink will ideally halve the length of a line, since the cells it connects will require less space. However, the physical size of the capacitor goes down, and it is similarly weakened. In the interest of maintaining density, the devices are tweaked to try and keep performance the same--where array speeds are measured in hundreds of MHz.
- Dense control transistors are small transistors, but small transistors leak more and small capacitors do not hold a lot of charge these days. The voltage levels differ from logic and the transistors combat leakage at the price of lower performance.
- DRAM is sold for its capacity, and not for much money relative to the mm2 being sold. The circuits and process are highly specialized to have high regularity, extremely high yield, and are less likely to use steps that increase manufacturing difficulty.
- Currently, external DRAM compensates by driving the interface very fast, and reading large chunks from the plodding arrays in parallel.
Logic tries to form a link between its output and a non-stored voltage source that pulls the output up or down. Hundreds of transistors do not share a wire, connections are more complicated, capacitance is generally undesired, and transistors can vary in size. Electron mobility is favored, since the transistor needs to drive its output. This means their processes and structures are not close to being as dense as a dedicated DRAM, and they tend to leak worse the denser they get. The kind of capacitors DRAM would like do not fit well in a logic process, so some kind of specialized process or additional steps are needed to even try. CPUs and GPUs tend to sell for more money, and their processes, transistor structure, and metal layers can afford to be more complex. Their acceptable yields are unacceptable for DRAM.
