Memory and elephants are playing in the band

Ever since the official introduction of the Intel Teraflop chip with 80 cores (octaginta-core?), I've been wondering about how you keep all those cores fed with data. It's not such a trick to imagine what you can do with 80 cores or dream about new kinds of applications that could be brought to the desktop or mobile platforms powered by 80 cores. However, to actually compute, cores need something to operate on, and the speed of memory hasn't exactly been keeping pace with previous clock speed increases.

With their announcement of embedded dynamic random access memory, or eDRAM, it appears that IBM has something that can start to address this problem. The claims are that eDRAM will have memory cycle latency of less than 2 nanoseconds, about 10 times faster than current memory technology.

(Cue Applause sign.)

Still, if I play with the numbers, I'm not sure if we still won't have something that turns out to be a white elephant. There's no mention about when eDRAM will be ready for use (first in corporate servers and later in consumer products). If we suppose that the memory will be ready when the Teraflop chip will be available, then we'll have a 10X increase in memory speed against a 20X increase in the number of cores. That's... let's see... divide by 3... add the 4... carry the 7... that's still lagging in speed by a factor of 2.

The IBM eDRAM is a great leap in the right direction. I'm just hoping that increases in memory access speeds will be able to progress faster than they have in the past. Otherwise a large percentage of cores in future processors won't even be left with peanuts to chew on.
分类:
如需更全面地了解编译器优化,请参阅优化注意事项