I've been pondering whether to build a new development and crunching machine. My old one suffered a hardware failure and is barely running right now.

The question is, what to get?

The obvious answer is the Core i7 4770K. It's reasonably priced, overclockable, supports the latest AVX2 instructions, and is very fast.

But is it the best for us?

They're not nearly as popular, but Intel also makes -E series chips, which are souped up versions of Sandy Bridge and Ivy Bridge. They have bigger cache and higher memory bandwidth (quad channel vs. dual channel). Since LLR and GFN are limited by memory bandwidth, both the larger cache and quad channel should improve performance.

The last two SR5 mega primes were found by a 6 core Sandy Bridge-E 3930K CPU, which crunches those tasks in about 2.5 hours. That CPU, however, is just under $600 and is almost double the price of a 4770K CPU.

However, there's quad core versions of the -E chips, such as the Ivy Bridge-E 4820K which is a more reasonable $323 and is actually $10 less than the 4770K. Sure, it lacks the AVX2 instructions, but it's not yet clear to me that AVX2 helps all that much. The 4820K has 10 MB of cache vs the 4770K's 8 MB, and uses quad channel memory instead of dual channel. Also, the 4820K's stock clock frequency is 3.7 GHz vs. the 4770K's 3.5 GHz.

What do you think about quad core Haswell vs. quad core IB-E?
____________
My lucky number is 75898⁵²⁴²⁸⁸+1

On paper, 4820K looks the best with the quad channel. OTOH, once you factor in the system cost, the $600 hexacore will give you the best bang for the buck.

You are sitting on one of the largest benchmark databases -- what does it say?

PS:- I'll just leave this here for a quick reference: http://ark.intel.com/compare/77780,77781,75123

On paper, 4820K looks the best with the quad channel. OTOH, once you factor in the system cost, the $600 hexacore will give you the best bang for the buck.

You are sitting on one of the largest benchmark databases -- what does it say?

PS:- I'll just leave this here for a quick reference: http://ark.intel.com/compare/77780,77781,75123

I have a 2600K and a 3770K. They are o/c'd identically, and the other standard specs (e.g. cache size) and environment (e.g. memory clock, 2-channel, operating system) are the same. For the smaller tests (ppse/pps/sgs) the 3770K is faster by a small amount. For all larger tests, the 2600K is faster, and the ratio grows as the tasks get larger. I believe this is due to the change in the cache replacement algorithm between SB and IB.

Correspondingly, the 3770K runs cooler, although that case has better airflow as well.

I can't comment on haswell; I don't have one.

Just my 2-bit observation :-)

--Gary

Here's what I mean about the database not being too useful for figuring out what's better. One thing I forgot to mention is that some people run their i7 with HT on and some run with HT off. Here's the average CPU time for completed and validated SR5 tasks, by CPU:

Can you look in the database for non-K i5 systems that can't be overclocked or hyperthreaded?

I'm trying to hold out for a Broadwell processor with Iris Pro graphics, and thus the 128MB on-package cache. Come to think of it, there are laptop CPUs with Iris Pro graphics now, although they're expensive and they might fry themselves running GWNums-based code 24/7.
____________

Can you look in the database for non-K i5 systems that can't be overclocked or hyperthreaded?

I'm trying to hold out for a Broadwell processor with Iris Pro graphics, and thus the 128MB on-package cache. Come to think of it, there are laptop CPUs with Iris Pro graphics now, although they're expensive and they might fry themselves running GWNums-based code 24/7.

Technically, there's a small amount of overclocking possible with the 2500 and 3570: 4 bins. That doesn't count the possibility of the mobo allowing all cores to turbo - some do. Given that, I find the performance of Haswell impressive. But it does look like the memory bandwidth on IB-E probably beats it.

On the other hand, also remember that the 4670K, overclockable without HT, is $230 at Amazon; less at a Micro Center store if you have one nearby. And LGA2011 mobos are at least $50 more than LGA115x.
____________

Here's what I mean about the database not being too useful for figuring out what's better. One thing I forgot to mention is that some people run their i7 with HT on and some run with HT off. Here's the average CPU time for completed and validated SR5 tasks, by CPU:.

If I get some time I'll try to see if I can find a pattern in the noise.

One must understand that the database is not intended to be an MIS system and generating statistics is not its purpose. The information necessary for this kind of analysis is not stored permanently -- anywhere -- and therefore our sample size is rather small by definition.
____________
My lucky number is 75898⁵²⁴²⁸⁸+1

One thing you guys are forgetting, i5-4670 is AVX 2.0 compatible. Soon George will release gwnum 28.4 with this ability.

Do we have any idea just how much AVX 2.0 will boost the performance of gwnum? Will it be as good as the difference AVX made?

One thing you guys are forgetting, i5-4670 is AVX 2.0 compatible. Soon George will release gwnum 28.4 with this ability.

Do we have any idea just how much AVX 2.0 will boost the performance of gwnum? Will it be as good as the difference AVX made?

Bottom line, definitely don't spend $300 to replace your 3770K with a 4770K. At least not just for AVX2. Haswell might be faster on its own, but AVX2 isn't going to make a large difference. It might not even be noticeable.

BUT since pc's are not bought today and thrown away tomorrow what you buy today must also work for the next 5 years, right?! So since PG is going the avx path, which presumably will get optimized giving even better performance, getting a cpu now with that capability may make more sense for the units you will want to crunch in 3 or 4 years from now. Unless of course you are one of those people that upgrades the cpu every few years. But that opens a whole new can of worms, which motherboard do you buy that can handle both kinds, the cheaper one now and the more expensive ones later on.

As far as Haswell goes, what matters is not AVX2, but FMA. (That's Fused (floating-point) Multiply-Add.) The good news is that FMA is very helpful in primality testing. The bad news is that, in order to add the FMA path, Intel decided to drop one of its floating-point adders. And apparently floating-point adders are also important in primality testing. So some adds may have to be replaced with FMA (multiplying by 1) for Haswell. Given that, I'm surprised Haswell is working as well as it is now. I would guess, from the 28.3 thread, that Haswell might improve by 8% or so with the next LLR.

One more thing: PPS Sieve could be made to take advantage of AVX2. In theory, going from SSE2 to AVX2 just means going from working with two 64-bit numbers at a time to four. However, PPS Sieve also uses the FPU stack to store six 62-bit numbers (in 80-bit FPU registers). Six is divisible by two, but not by four. I have considered rewriting the PPS Sieve app to take advantage of AVX2, but it would be a complete rewrite of the main algorithm. And since GPUs are so much better at PPS Sieve, I don't really want to bother.
____________

The effects of the quad-channel memory become more important for IB-E or IV-E.

I run SR5 with 4960x@stock(IV-E).
When all 6 cores are running, the turbo-boost technology sets the frequency of the 4960x to 3.7GHz.
And I select 1333cl9 for memory clock, because my memory module does not have 1600 in XMP.

The result is :

4960x(3.7GHz,1333cl9, Core 6, HT off)
dual-channel 3.5hours 85W (http://www.primegrid.com/result.php?resultid=523564454)
quad-channel 2.3hours 95W (http://www.primegrid.com/result.php?resultid=523629619)

The watt of the CPU is measured by HWINFO64. The lower watt means a memory starvation.

For comparison, I disable 2 cores of 4960x in BIOS to make 4-core system because I don't have 4-core CPU.
The size of L3 cache is diffrent from that of a genuine 4-core sytem.

...

The bandwidth of dual-channel memory is exhausted by 4 core.
The productivity for 6 core with dual-channel memory is about same as that for 4 core.
Moreover, quad-channel memory for 4-core system is less important.

Larger FFTs will stress the memory more. So quad channel will be future proof.

As AVX2 isn't a significant boost for LLR and GFN, would it be reasonable to hope that AVX-512 might be? It seems they are doubling the floating point register size again with that instruction set. There are other additions as well.

Unfortunately the info I can find is that this advance won't show up until 2016, at least in consumer-grade CPUs (i.e. i7-ish; Xeon maybe earlier).

--Gary

As AVX2 isn't a significant boost for LLR and GFN, would it be reasonable to hope that AVX-512 might be? It seems they are doubling the floating point register size again with that instruction set. There are other additions as well.

Unfortunately the info I can find is that this advance won't show up until 2016, at least in consumer-grade CPUs (i.e. i7-ish; Xeon maybe earlier).

--Gary