Hacker News new | ask | show | jobs
by wbl 557 days ago
Hindsight is 20/20. If higher clocks won out because extractable IPC was lower Netburst was the right design.
2 comments

The problem with Netburst was the clock speed wall. They designed it with a view to continued doubling which has been a trend up until then. It turns out once you go past 3GHz improvements are much, much slower.

We hit 3GHz in 2003. If you go back ten years from that, the top speed available was 66MHz, ten years before that and it was 8MHz. Fast forward 20 years and we're just barely scraping 6GHz.

Doing the math - even a conservative estimate for 1983 to 2003 is eight doublings.

That would give us a hypothetical Pentium 4 at 768GHz today. Which, Netburst or no, would wipe the floor with the fastest available multi-core chip.

While also putting out enough heat energy to keep a small town in Northern Canada warm all winter long.

Foresight (at the time) told you NetBurst was a dead end though. AMD made a lot of sales from people avoiding NetBurst, ignoring clock speeds and looking at actual performance numbers.

NetBurst looked good on benchmarks; code with few/no branches that fit in the L2 cache if not the L1 cache. On real world code consisting almost entirely of branches and pointer chasing the NetBurst chips performed horribly compared to lower clocked Athlons.

Intel's promises of 5+ GHz chips got more laughable as the line went on because they were already straining cooling and power supply options at less than 3GHz.

Netburst's branch predictor choices survived. It was the penalties that were a bad idea.