[URL="http://en.wikipedia.org/wiki/Wirth%27s_law"]Wirth/Gates/May's law[/URL] negates Moore's law and then some.

[QUOTE=Batalov;308075][URL="http://en.wikipedia.org/wiki/Wirth%27s_law"]Wirth/Gates/May's law[/URL] negates Moore's law and then some.[/QUOTE]

w00t! :lol:

[QUOTE=Batalov;308075][URL="http://en.wikipedia.org/wiki/Wirth%27s_law"]Wirth/Gates/May's law[/URL] negates Moore's law and then some.[/QUOTE]

yes but that means it quadruples every 3 years you go back so that means that assuming compatibility, I wonder what the increase in performance of modern processors and the increased speed of using 3 year old software would come to. assuming doubling every 18 months for processors you get 4*4 = 16 times as fast.

[QUOTE=Batalov;308075][URL="http://en.wikipedia.org/wiki/Wirth%27s_law"]Wirth/Gates/May's law[/URL] negates Moore's law and then some.[/QUOTE]
You can't negate "Moore's Law".
If x=e[SUP]at[/SUP] and adkust "a" appropriately with time, it is going to be roughly right.

D

[QUOTE=fivemack;307906]I'm not really interested in tiny little boxes; I've got a workshop to keep the equipment in, I'm mounting everything over NFS so gigabit-ethernet is non-optional, and the NUCs look expensive per hyperthread. I'd be buying an i7/4930 for the quad-channel memory interface, and little boxes without ethernet really aren't a substitute.[/QUOTE]I have a rather nice [URL="http://mbed.org/handbook/mbed-NXP-LPC1768"]tiny little box with ethernet[/URL]. It cost me about £40 a couple of years ago and performance wise it's comparable with a top of the range 486 machine. That may not sound much compared with the latest x86-64 machines but it is dirt cheap and very low power. On compute-bound tasks I've measured it to have several times the compute per watt of the x86-64 systems.

The Raspberry Pi is even cheaper but I don't have one to play with. What's interesting about that machine is that the RAM is mounted piggy-back on top of the cpu and the whole thing takes up about a square centimetre of board; together they cost a few USD. I've done some preliminary designs which suggest that an entire Class C internet of machines, complete with associated glue, would fit on a single double-sided board. A kilocpu parallel computer would fit in a desktop box, cost a few thousand USD and plug into a standard domestic power socket.


Paul

[QUOTE=davieddy;308114]You can't negate "Moore's Law".
If x=e[SUP]at[/SUP] and adkust "a" appropriately with time, it is going to be roughly right.

D[/QUOTE]
Semantics! You say "po-tah-to", I say "po-tay-yo". See [URL]http://en.wiktionary.org/wiki/negate[/URL] :

[QUOTE]2. To [URL="http://en.wiktionary.org/wiki/nullify"][COLOR=#0066cc]nullify[/COLOR][/URL] or cause to be [URL="http://en.wiktionary.org/wiki/ineffective"][COLOR=#0066cc]ineffective[/COLOR][/URL].
[I]Progress on the study has been [B]negated[/B] by the lack of funds.[/I]
[/QUOTE]

[QUOTE=fivemack;308055]The fastest processor currently available is the Radeon 7970, which offers a peak of 1 teraflop DP; and indeed it achieves this by incorporating 512 double-precision ALUs on a die with 4.3 billion 28nm transistors, which is about the Moore's-law prediction.

The transistor counts on CPUs have also been increasing pretty much to schedule, though a lot of them end up as cache; Northwood has 55 million 130nm transistors, a 3960X has 2270 million 32nm transistors on a significantly larger die, IVB has 1400 million 22nm transistors on a die about the same size as Northwood.[/QUOTE]

True, but you can't accurately compare a GPU to a CPU because GPUs can't do everything that CPU can.

On the other hand, Haswell's AVX2 and FMA3 will likely reduce the gap between actual performance and that projected by Moore's law. It probably won't offer twice the speed of Sandy (Ivy) Bridge, but I'd imagine that the performance boost would be similar to that of Sandy Bridge over non-AVX chips.

Of course, bandwidth will still be a major issue. From what George has said, AVX2 and FMA3 will only increase LL speeds by about 25% and 5%, respectively. Hopefully DDR4 will reduce the bottleneck as some predict.

[QUOTE]…because GPUs can't do everything that CPU can.[/QUOTE][URL="http://en.wikipedia.org/wiki/Turing_completeness"]:orly owl:[/URL]

[QUOTE=ixfd64;308163]Of course, bandwidth will still be a major issue. From what George has said, AVX2 and FMA3 will only increase LL speeds by about 25% and 5%, respectively. [/QUOTE]

Did I say that? I don't remember AVX2 having any instructions that LL testing can use. It will be useful for TF, except that GPUs have made CPU TF obsolete.

I was talking about this thread: [url]http://www.mersenneforum.org/showthread.php?t=16152[/url]

[QUOTE=ixfd64;308187]I was talking about this thread: [url]http://www.mersenneforum.org/showthread.php?t=16152[/url][/QUOTE]

and I said:

[quote]
A complete shot in the dark: nearly double the TF speed. For LL testing, FMA and the scatter/gather possible improvements, maybe 5%.[/quote]

The reference to 20% AVX was the version 27 AVX code that has already been released