Core i5 or i7?

[petrw1]

Originally Posted by sdbardwick

My earlier post has a bit of what you are looking for.
Link to spreadsheet

So if I try to extrapolate: 2240 is a new FFT size but allow me to assume the exponent in question is 40,000,001.

Each i5-750 core at 0.034 ms will take 15.74 days.
In 63 days this CPU will complete 16 exponents.

Each 1090T core at 0.44 ms will take 20.37 days.
In 61 days this CPU will complete 18 exponents.

Round 1 goes to AMD with a small disclaimer that it is OC'd to 3.2 vs 3.0 for Intel. Assuming a linear OC to speed increase Intel at 3.2 would process 17 exponents. Still an edge to AMD.

[sdbardwick]

Minor note: 1090T is not OC'd (3.2GHz is stock); 750 OC'd from 2.66 to 3 GHz.
But your conclusion is correct.

Intel has ability to clock higher than AMD; a heavily OC'd 750 can probably beat a 1090T.
Personally, I won't do high overclocks (on either Intel or AMD) as you start getting 'silent' errors. Silent errors are what I call errors that are not detected by Prime95 during the LL run (error codes are all 0 ) but later residues don't match. The boxes generating silent errors can pass days of Prime95 torture tests.

[Brain]

Originally Posted by sdbardwick

Silent errors are what I call errors that are not detected by Prime95 during the LL run (error codes are all 0 ) but later residues don't match. The boxes generating silent errors can pass days of Prime95 torture tests.

I'm scared. Is this commonly agreed? Solution? (Run always a DC on one core..?)

By the way, I suggest a PrimeScore feature in Prime95. This would be a positive integer describing a (CPU) performance index that is easily comparable and generated after benchmarking. This could also make Prime95 a test reference in hardware online magazines.

Formula (naive) suggestion, given no helper threads: PrimeScore = Theoretical CPU throughput = Sum(FFT size*(1/Iteration time))*Number of possible workers

[sdbardwick]

Let me do a bit more research before jumping to conclusions (or into a panic). A fair portion of those errors occurred fairly long ago using old versions of Prime95 (all the way back to v19). Also, IIRC, all but one or two were from OC'd processors. I'll pull up my bad LL results and check.

[sdbardwick]

Not going to be very precise, because this is based on me manually counting numbers as they scroll by on the screen (computer I'm using right now is not mine and has limited resources).

Quick check of my LL tests reveals that since 04 Apr 2008 I have had (about)
38 Bad results with all 0s in the error code field.
Of those 38, 24 are from OC'd boxes.
Of the remaining 14, 6 were from non-OC'd boxes with NVidia integrated video (I've had problems with NVidia drivers in the past).
Of the remaining 8, 2 were from a box where the RAM went bad later that month.
The last 6 cannot be tied to either bad RAM, OC, or faulty drivers (some results are missing a computer name, some I don't remember details about the boxes).

Prior to 4 Apr 2008 (or without a completion date), I had about 72 bad results with all 0s in the error code field.
70 were from OC'd processors. Those were mostly P4 processors, and the ones giving bad results were mostly Prescott based IIRC.

PrimeNet throws a CGI time out so I can't give a total number of tests, but I completed about 600 LL and DC in the past 365 days.
Edit: In the past 365 days, I've had 8 silent errors, 6 from 1 non OC'd box that has been since rebuilt (might have had bad RAM), the other 2 from non-OC'd boxes that had incorrect memory timings and/or bad RAM.

[nucleon]

I have I7-920 and i7-930 here at home. Both I like.

I tend to agree with this statement:

Originally Posted by Brain

Why don't you wait for Intel's Sandy Bridge architecture next January?

January is not that far away. If you can wait until then, I think that's the better option.

-- Craig

[petrw1]

Originally Posted by nucleon

I have I7-920 and i7-930 here at home. Both I like.

I tend to agree with this statement:



January is not that far away. If you can wait until then, I think that's the better option.

-- Craig

In my experience "available" and "affordable" are about one or two years apart.

[petrw1]

Originally Posted by nucleon

I have I7-920 and i7-930 here at home. Both I like.

I tend to agree with this statement:



January is not that far away. If you can wait until then, I think that's the better option.

-- Craig

With January only two months away I thought we would have heard something more official from Intel by now??? Or at least so good leaks.

[henryzz]

Originally Posted by petrw1

In my experience "available" and "affordable" are about one or two years apart.

That is quite true but when new cpus are introduced the price of current cpus might go down.

[R.D. Silverman]

Originally Posted by nucleon

I have I7-920 and i7-930 here at home. Both I like.

I tend to agree with this statement:



January is not that far away. If you can wait until then, I think that's the better option.

-- Craig

Based upon what I have read Sandy Bridge will not be any faster/more
powerful than the current i7's.

I have a thought.......

Rather than concentrate on new architectures, smaller dies, etc., I think
it would be marvelous if they just gave us a new chip with the same
architecture as the current i7, but with 1 Gbyte of L2 cache per core.

We'd really see some performance improvements!

Give us cache, lots of cache in the sunny skies above....

[R.D. Silverman]

Originally Posted by petrw1

With January only two months away I thought we would have heard something more official from Intel by now??? Or at least so good leaks.

I have seen some "official" discussions. I am less than impressed.