| Galaki said: Is GDDR3 really that slow? Is it the bottleneck slowing down the cpu/gpu? Isn't GDDR5 is like a superhighway with 8 lanes where only 4 lanes ever get filled, or something like that? |
GDDR3 is based on DDR2 tech, so supports the transfer of up to 4 words every 2 clocks (ie 2 words of data per clock), whereas GDDR5 is based on DDR3 and supports the transfer of 8 words per 2 clocks, (so 4 words of data per cycle). The actual word width per chip is the same, it simply allows the data to be fed in (and out) faster due to a large prefetch cache. That is so long as each transfer is from the same memory row, which is maybe what you're referring to. For video cards this is a not an issue as textures and geometry data are stored sequentially in massive arrays, whereas CPU's can be more random in their memory access. This problem affects GDDR3 as well. However, this is something you can largely mitigate through cache and memory controller design, and through programming (ie keeping data access to 64 bit words only, not using x86's older, single/double byte access instructions).
GDDR5 also supports higher clock speeds through improved analogue design, so you'll see GDDR5 run at up to 1.75GHz (so 7 Gbit/s per pin effective clock) versus GDDR3 which tops out at ~1GHz (so 2 Gbit/s per pin effective clock).
