Kick-off entry and a short primer on Manycore
Wednesday Jun 13, 2007
Kicking off the shiny new Sun blog. I liked this theme for its green color (Go Green!! Go White). The title (Manycore Mania) represents, in my view, the architectural driver of the near-future silicon/EDA industry. Just how may cores can/will we go? Will we see a Moore's law corollary for core explosion? So why manycore from single-core architectures? one word: POWER. With increasing clock frequencies, it was no longer viable to obtain perf. improvements without precipitously pushing the power dissipation higher. Things started to run into a wall when increasing power led to hotter chips, unmanageable leakage, thermal breakdown, decreasing reliability, failure rates, etc. This reminds me of what I heard in HPCA-11 a few years ago: "Pipelining was the last great innovation in computer architecture" (I don't recall who said it but I promise to find out soon). Beyond a certain limit, deeper pipelines, issue widths, superscalarity, higher clock freq., etc., were not yielding desired RoI. So the industry switched to multicore and we obtained overall power reduction by using slower clocks and simpler cores and sacrificing per-core performance but gaining in overall throughput due to parallel execution. From thereon, we have evolved to manycore.
|
|
|
|
