I reposted in the cpu forum.
I remember two years ago when I had a Pentium 4 running at 3.6Ghz and ram running at 400ish.
Technology appears to evolve every month and some of us right now have VRAM running faster than CPUs and maybe in a few months even the normal RAM.
Modern CPUs can process more work in less cycles, contrary to memory that keeps going up and up.
So, is there any new technology in sight to revert this Mhz freaks(if even possible) and keeping bandwidth high with lower freqs.? Or the only way to bandwidth is pure Mhz?
This question may be stupid, but, whatever... :?
What RAM is running at Ghz? There are doublepumped busses and such, but the RAM is running from 400-900 Mhz, and it is very expensive to boot.
You can keep Mhz low and increase memory bandwidth, hence 128 bit, 256 bit, dual channel, etc. memory. Double data rate was another innovation, allowing double the bandwidth on the same pipe. However, these technologies add latency, or at least they don't improve it. Looking at DDR vs DDR2 RAM, you'll see DDR2-1000 memory (500 Mhz bus), with latencies of 5-5-5-15, while DDR-400 (200Mhz bus) had latencies of 2 or 2.5. This is roughly the same number of nanoseconds, so the individual memory cells did not double in performance, but the external communication rates increased. Internal cell latency is steadily improving with improved materials science, and there are all of these breakthrough technologies that are supposed to dramatically improve memory performance, like ZRAM or RDRAM in days of yore, but often by the time they reach market they are only a 10-20% improvement over the existing designs. That was one of the problems with RDRAM, it released designs that were groundbreaking, but by the time it was licensed and in mass production, DRAM had caught up to it then surpassed it.
IBM and Intel (and a ton of others) have shown variations of embedded memory running at CPU clock speeds, solving latency problems and dramatically improving memory access times, but memory is a fairly large silicon structure, so making memory at the same process level as CPUs is rather expensive. Articles about IBM's eDRAM indicate that 128MB is roughly the same size as a CPU, and a commodity CPU optimistically costs ~$50, so that's $.40/MB. Compare to DDR2 at $.08/MB and you'd be hard pressed to find someone who'd pay more for the memory subsystem than the rest of the system. The exception is in GPU's, where VRAM often costs as much as the GPU itself, but that low latency is needed for performance.
Of course, increasing internal memory speeds to CPU performance doesn't help on a very slow bus. If total memory latencies were 1 bus clock, which is impossible, there would still be a 14-24 clock (based on CPU multiplier) delay between the end of the memory request and the start of the transfer. Better than 150+ today, but still not perfect. IBM and Intel have discussed and demonstrated designs with embedded memory so that memory speeds are close to cache speeds, but then you'd be buying a processor with a set amount of memory in it, paying a ton of money for that memory, and upgrading memory would mean getting a new processor.