Hello,

There is an interesting looking upcoming CPU from Intel mentioned here: http://en.wikipedia....microprocessors, the Core i7-3820, with 4 cores.

My question is: This CPU is to be released in November 2011. The Intel Q6600 was released somewhere end 2006 or beginning 2007, 5 years earlier. How much better do you think the Core i7-3820 is than the Q6600? The Q6600 is 2.4GHz while this one is 3.6GHz, but for the rest, both have 4 cores and around 8-10MB cache. Is the architecture a lot better? Does Moore's law still apply?

Thanks!

Does Moore's law still apply?

N.B. Moore's law doesn't state that CPU's can double in speed every 2 years, just that the number of transistors that can be fit into them can double every 2 years. You'll have to wait until the transistor counts are known first, but Moore's law is basically their own target/benchmark/self-expectation.

Besides the clock-speed difference, of note is the more than double bus speed, and having a separate L2 cache per core.

Also the SSE/AVX units are WAY more powerful. You can crank out a ridiculous 16flops/cycle out of the current sandy bridge cores. I measured up to 59Gflp/s on a single core of a i5-2500k and 200Gflop/s on all four... Obviously this doesn't transfer to all real world applications.

Also the SSE/AVX units are WAY more powerful. You can crank out a ridiculous 16flops/cycle out of the current sandy bridge cores. I measured up to 59Gflp/s on a single core of a i5-2500k and 200Gflop/s on all four... Obviously this doesn't transfer to all real world applications.

Somewhat related, but the new core is actually have fairly good video performance. I was pleasantly surprised by how well it stood up while I waited for my video cards to come in the mail.

The i5-2500k/i72700k are great for overclocking too. People get them to 4.5 ghz fairly easily. I'm very pleased with mine. The new intel chipsets that support the new processors also have some pretty snazzy features.

New processors from Intel are the Sandy Bridges architecture, which tends to perform something like 10-30% faster than previous architectures at the same clock speed. Look for example at this: http://www.tomshardware.com/reviews/processor-architecture-benchmark,2974.html and see other benchmarks on that site that are less artificial. Sandy Bridges has a smaller pipeline and other changes which are hard to notice in 'just the numbers' that can make a big difference. My bet is that you would see a big difference, depending on what you are doing with the computer.

Now, if only we could get someone to write software to make use of all of this power.

Rather than being stuck with VMs that are architecturally at level of 386.

Now, if only we could get someone to write software to make use of all of this power.

Put it in a games console (and then wait 5 years) ;)

[quote name='Lode' timestamp='1319444519' post='4876243']Does Moore's law still apply?

N.B. Moore's law doesn't state that CPU's can double in speed every 2 years, just that the number of transistors that can be fit into them can double every 2 years. You'll have to wait until the transistor counts are known first, but Moore's law is basically their own target/benchmark/self-expectation.[/quote]
(emphasis added)

That's it, and nothing more. Moore's "Law" is not a law (such as Newton's laws) at all, but a prognosis for the coming decade made by one guy in the 1960s after watching the development of ICs during the last decade. More transistors for the same amount of money = more power (for the same amount of money).
Incidentially, that one guy was one of the founders of Intel, so unsurprisingly the company followed this prognosis as business plan. This was a wise decision, as tere is a conflict of interest in making too fast CPUs (which they easily could have done at that time), after all you want to be able to still sell some next year and the year thereafter.

Moore's law will eventually become problematic, if CPUs remain physical and mostly two-dimensional, because adding more transistors does not make things endlessly better (neither does adding more cores), as lanes between them grow longer, space is finite, and you cannot make things smaller indefinitely without getting other nasty effects.

We are already there with caches, for a whole while. If you look at L1 caches today and 10 years ago, you'll notice that their size has not changed at all. Certainly more L1 cache would help, so why don't modern processors have 256k or 512k of L1 cache? The reason is it just won't fit, no matter whether you can produce it cheaper.

Apart from newer instructions (such as the totally insignificant AVX set, sigh... bigger is not always better), a major reason to more modern processors performing a lot better (given identical core counts and similar clocks) is improved instructions, better pipelining, and better heuristics. Notable examples are the "supershuffle" which takes a single cycle, and the Penryn division engine (I forgot the name... radix-16?) which takes only 1/2 as many cycles as older generations such as the Q6600 would.

Apart from newer instructions (such as the totally insignificant AVX set, sigh... bigger is not always better)...

It actually is... I'd use 1024 bit AVX instructions if there were any. This obviously doesn't help every type of application but the scientific stuff greatly profits from this.

Apart from games, does anything use these instruction sets? Do compilers (for compiling)? Does Flash? Does Linux?

Thanks!