After reading "Microprocessor Report: Intel, Fix the PC!" I did notice very little has changed over the yeads. Just adding cache, making branch prediction a little better, or improving the FPU. Even gains from more than 4 cores are minimal. I also heard of specialized cores ( or parts of a core like "quick sync", but they don't seemn to work well yet.
Is this a problem of one hole and too many diggers and this can never be fixed with complex instructions?
Or things can be done to improve performance like 10x ?
I didn't read anything, not even rumors. I hope something is done to get things like natural language (HAL 9000 style), real-time ray-tracing, realistic physics, etc.
The best I read was getting power down 10x, but with no speed gains.
I have no idea.
and Merry Xmas...
I think that speeds have increased dramatically even in the last 5 or 6 years. I mean, come on, Intel is already down into 22nm technology and is going to start production using 14nm technology next year. Core 2 Duos (2006) were using 65nm technology. It has come a long way in such a short amount of time. Check the benchmarks from the link below. Compare the PassMark scores from the Low-Mid end CPUs (~2006) to High End CPUs of today. There is a dramatic increase in scores.
Thanks for the posts.
But I think you missed my point.
Yes, performance has increased by about 2x over 6 years. This is about 10% a year.
1.1 * 1.1 * 1.1 * 1.1 * 1.1 * 1.1 ~= 1.94
Yes, I heard of Moores Law and that has been consistent of doubling the number of transistors every 18 monts-2 years for decades. This should continue for the next 10-20 years.
The problem is transistor count has little to do with performance. If performance was similar to transistor count, then we should see performance like this for a 18 month cycle:
2 * 2 * 2* 2 = 16x or 8x assuming a 2 year cycle over 6 years.
At the rate we're going, I don't expect to see the kind of performance I described within my lifetime.
Here is one piece of (old) news that looks encouraging. Intel may forget the old awkward (awkward and messy even back in the 80s) x86 completely and use AVX as the main engine - just keep a small corner of the die x86 for backward compatibility.
AVX2 included integers and 3-address instruction sets so you can do full branch programming within AVX. AVX is modern, elegant, efficient, and clean - seems like a much better direction. http://vr-zone.com/articles/mainstream-desktop-cpus-fut...