Post any comments, views or bright ideas about CPU crunching here :)

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

I have that brand new Core i7-4770K configuration (4GB 1333MHz Dual Rank, Dual Channel, CL9 RAM; 320GB 7200rpm HDD; no OC yet) I've mentioned in the parent thread before. My experiences with it are very impressive. The installation of Windows 7 x64 was faster than usual. I do a lot of such installations on very different configurations. My last two was i5-3470s (8GB 1600MHz Dual Rank, Dual channel CL9 RAM; 1TB 7200rpm HDD; no OC).
Right now I'm experimenting with Rosetta @ home on it. The aim of my experiments is to figure out how many simultaneous tasks are optimal, and whether the type of RAM used changes the RAC or not? Also to find out if Windows XP x64 could be installed properly on it, as there is no official support for Windows XP x86 and x64.
I had some random system restarts, but a slight increase in the RAM voltage (by 50mV) fixed this issue. I'm planning to change the RAMs to 1600MHz Single Rank, Dual Channel, CL11.
You can check the results of this experimental host here and here. I've never earned over 1000 credits in R@h for a 24h workunit.
I am an Intel "fan", but I've no doubt that AMD CPUs can outperform Intel CPUs in some applications (and vice versa of course). But if you take the energy costs in consideration, the recent Intel 22nm CPUs are better (the question is that how long it takes to save the price difference between Intel and AMD CPUs by the lower energy costs?)

Well let me start it, as it is mainly my "fault" that this thread existed.

Edit Zoltan beat me, so I am second :)

An i5 has another architecture than an i7 with HT. The Xeon's have yet another architecture and thus AMD processors does also. That must be taken into account when compare, but could be hard in reality.

You cannot compare apples with pears, is something that is what we say in Dutch when things are being compared that are not the same (and can't be compared). It is both fruit and healthy, but that's it.
Same as compare women and men, they are both human but their architecture is significantly different...

---

Greetings from TJ

Same as compare women and men, they are both human but their architecture is significantly different...

Should we start a new thread to discuss that? :D

This is off-thread bur how do you get such long WU's? Mine complete in about 10000 seconds for 50-100 credits.
My Rosie are here: http://boinc.bakerlab.org/rosetta/results.php?userid=308421

What is the important difference between single rank and dual rank memory?

I have 12GB in both i7's but have never seen that more than 5.8 was used.

---

Greetings from TJ

This is off-thread bur how do you get such long WU's? Mine complete in about 10000 seconds for 50-100 credits.

You can set the "target CPU runtime" in your Rosetta@home preferences.

What is the important difference between single rank and dual rank memory?

That's what I'm going to find out :)
Single Rank usually comes in the form of single-sided memory module. It has less chips therefore consumes less energy, and it's faster.
A Dual Rank memory is like two single rank memory on the same module (but it can't be dual channel, since the data bus is also common, and there has to be some delay when switching ranks).
Some older motherboards had limited memory rank support, so if you put dual rank modules in all of their slots, the MB didn't recognized all of the memory. There is no such limitation on modern motherboards (as far as I know).
See this video about single vs dual rank, and this video about single vs dual channel.

I have 12GB in both i7's but have never seen that more than 5.8 was used.

That's okay, but what you don't see in task manager is the saturation of the memory bus. I think R@h is a very memory intensive application, and 6 of them (plus a GPUGrid task) could saturate the memory bus to a level where the PC becames unresponsive (and unusable).

That's okay, but what you don't see in task manager is the saturation of the memory bus. I think R@h is a very memory intensive application, and 6 of them (plus a GPUGrid task) could saturate the memory bus to a level where the PC became unresponsive (and unusable).

Ah I understand, but my unresponsive system was only the i7 with the GTX660 in it. Now with the GTX285 in it it is very responsive again. But you are right, if we forget HT, the four cores are all in use indeed.

Another system a quad with vista x86, so not even 4Gb memory to use, runs 24/7 GPUGRID and 2 Einstein on CPU (thus not Rosetta), and is very responsive. I use it for mail, typing letters, even watch live TV via stream, make post stickers, I fact I use it the most.

You can set the "target CPU runtime" in your Rosetta@home preferences.

Indeed I have asked about that in the forum there but I never got a very clear explanation. The idea is to set how many times decoys are cycled. I will experiment with it.

---

Greetings from TJ

Posting here, mainly for full disclosure about myself.

I am not an Intel fanboy, period. I've owned many non-Intel CPUs, starting with an AMD 386sx, moving to an AMD 386DX-40, then to a UMC 486, then at some point to an AMD Athlon XP and directly after that to an AMD Athlon-64 X2.

Sadly for AMD hot-blooded fanboys, after Intel introduced the Nehalem architecture, AMD lost the performance crown! As simple as that, really! For a few years AMD was playing the tune and Intel danced to it, but unfortunately for AMD, things changed. For a few years, AMD set the performance bar and Intel followed far behind. But that finished and AMD had to return to the price game.

The price game is a good thing for us all, of course! With processing power surpassing almost all our everyday needs, what we really want is cheap adequately powerful CPUs. Without AMD, Intel would (as it has done for years in the past) charge gold for baked sand.

So, we love AMD and we need it to keep Intel's prices under control. And AMD has achieved that pretty successfully over the years, forcing Intel to sell a myriad of low to mid level CPUs at low prices, of course higher than AMD's prices, which is a good thing for AMD!

BUT, at least for the time being, the high performance war has been lost for AMD. I really do hope it will manage to come up with an architecture that will surpass Intel and introduce new levels of performance. This will be good for all of us and I will be among the first to buy the AMD chip!

---

An i5 has another architecture than an i7 with HT. The Xeon's have yet another architecture and thus AMD processors does also. That must be taken into account when compare, but could be hard in reality.

You're mostly wrong here. The single most expensive things with logic circuitry are the fabs (the circuit fabrication factories). A fab cannot produce many different types of circuits, certainly not different architectures. Once an architecture change is done, it's done, there's no way back. So, logic circuitry companies, like Intel, AMD, Nvidia, etc, try very hard to do a single thing: ooze the last drop of performance / efficiency / profit from the architecture they are currently producing. Intel's i3s, i5s and i7s, even lower-level and higher-level CPUs, like some Pentium-branded and Xeons, share the same architecture. It's like Nvidia and Kepler and Fermi before that. Wikipedia has detailed information about these things.

You cannot compare apples with pears, is something that is what we say in Dutch when things are being compared that are not the same (and can't be compared). It is both fruit and healthy, but that's it.

AMD and Intel CPUs are different of course, but they seek to do the same thing: run our programs! So, comparisons are not only valid, they are necessary! Comparisons are what drives performance up and prices down!

---

after Intel introduced the Nehalem architecture, AMD lost the performance crown![/i] As simple as that, really! For a few years AMD was playing the tune and Intel danced to it, but unfortunately for AMD, things changed. For a few years, AMD set the performance bar and Intel followed far behind. But that finished and AMD had to return to the price game.

The price game is a good thing for us all, of course! With processing power surpassing almost all our everyday needs, what we really want is cheap adequately powerful CPUs. Without AMD, Intel would (as it has done for years in the past) charge gold for baked sand.

So, we love AMD and we need it to keep Intel's prices under control. And AMD has achieved that pretty successfully over the years, forcing Intel to sell a myriad of low to mid level CPUs at low prices, of course higher than AMD's prices, which is a good thing for AMD!

BUT, at least for the time being, the high performance war has been lost for AMD. I really do hope it will manage to come up with an architecture that will surpass Intel and introduce new levels of performance. This will be good for all of us and I will be among the first to buy the AMD chip!

Initial insults aside I would pretty much agree with your post. I am an AMD fan and for the reasons I stated before: AMD has been the only thing that has kept Intel honest both in performance and pricing. Intel has played dirty pool against competitors throughout its history with it's FUD, anti-competitive practices, rigging benchmarks, giving payouts to PC makers to not use AMD, etc, etc, etc.

Even in the years when AMD processors were far superior to the flawed Intel P4, most of the hardware sites and benchmark companies fell all over themselves to devise "reasons" why the P4 was faster. Intel ruled with FUD, payoffs and lies. It was pretty sickening and one of the reasons I only buy Intel when I have to. Yes, I'm old and have been around the microcomputer industry since it began. I watched it all happen in real time. Flashhawk posted some links of just a couple examples concerning Intel's behavior. This knd of thing hurts us all. Here are the links again:

http://www.agner.org/optimize/blog/read.php?i=49&v=t

http://semiaccurate.com/2011/06/20/nvidia-amd-and-via-quit-bapco-over-sysmark-2012/

You're not old, you are experienced! In the early eighties I was nagging for a Commodore 64, but it was to expensive, but I got an ZX Spectrum from my dad. I was the only one at school with a computer first, even the math teachers where jealous. Later I got an America friend (his dad was from the Army and they lived in the Netherlands for a few years). Very cheap I got then an real IBM. Rock solid but slow as later found. There was shop with brand-less PC's and one had a mathematical coprocessor and did Fourier calculations at least 1000 times faster then the 5 times expensive IBM. These mathematical coprocessors are now integrated, for the young readers :)

---

Greetings from TJ

There was shop with brand-less PC's and one had a mathematical coprocessor and did Fourier calculations at least 1000 times faster then the 5 times expensive IBM. These mathematical coprocessors are now integrated, for the young readers :)

It was probably an 8086 based machine with an 8087 co-processor. I souped a few of those up with NEC V30 chips. Used the NEC V20 for the 8088 sockets. Rocket fast for those days :-)

Thank you to everyone who started using this thread.

A few general considerations when comparing run times on multicore processors:

If you are looking at someone elses data, making comparisons is a guessing game. You don't know what their CPU clocks are set to, what their RAM is, what else they use the system for or if the system is set to downclock/sleep/LAIM...

For your own data it's still important to take note of the CPU clocks; if the CPU is at stock the clocks may well be different depending on the number of tasks being run. If for example a CPU's clocks are 3.4GHz when 2 cores/threads are running it might be the case that when 8 cores/threads are in use the clocks are 3.1GHz. The clock speed difference is almost 10%

Even if you overclock and fix the clocks so they are static you will get runtime variation due to CPU usage competition and restraints on the CPU from other resources (RAM, drive for heavy I/O projects). This can be clearly seen in climate models for example, where the increase in return rapidly diminishes the more cores you use. When you reach 6 or 7 cores there is virtually no benefit of running a 7th or 8th model. For this reason, and others, it's a good idea to mix and match CPU projects rather than just running one. For WCG crunchers this is the norm and the recommendation, though recent client updates are now preventing this!

To emphasize what Zoltan said in his original post, its about getting the most for your money in the long run. So the purchase cost may or may not be offset by the running costs. This is of course something that varies by region/area. A processor that is best for you may not be the best for someone else's circumstances.

Space is another factor to consider. If you have plenty of space, you can buy more systems. If the electric is cheap, then this is a better option. If however you have almost no space, you might only be able to buy one system, so you might want to go for a higher spec and hope that the outlay will be offset in the long-run against running costs, or at least partially.

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

I had some random system restarts, but a slight increase in the RAM voltage (by 50mV) fixed this issue.

It turned out that not the RAM voltage causing these random restarts. The CPU C3/C6 state support was enabled in the BIOS (I forgot to disable them after a BIOS update enabled them). Still crunching 7 threads of Rosetta @ home on this PC. It's CPU is running at 3.7GHz no matter how many threads (2..7) are running at the same time. I'll leave it like this for the weekend (maybe I'll start a 8th thread tomorrow), and I'll change the RAM modules to single rank on Sunday evening.

Regarding the discussion of AMD vs. Intel at Yoyo-ecm in the last thread: while I think Intels are overall the better choice for crunching, depending on your electricity cost and how long you keep your systems, you have to give credit where it's due: the AMDs seem pretty good at ecm. Whatever this may be worth for you.

Taking a quick look at the current top 3 to 5 I see the following

i7-3930K with HT: 17800s for 570 credits
FX8120: 17000s for 570 credits
X6 1035T: 18450s for 593 credits

So while we don't know much about the actual configs and especially clock speeds (except Beyonds X6), it's clear that AMD is very competitive here. And for this it doesn't matter how many people may be running which CPU here in less-than-optimal configs.. what we see among the top hosts is what the hardware can do, given the chance.

@Zoltan: I'd be very interested in an investigation whether Haswell scales better with HT than previous chips. The reason I suspect this: it's got more execution units! They're not terribly relevant to FP crunching, but in mixed workloads they should definitely show. And BOINC is the best real world usage case for mixed workloads.
I know this is not easy to measure, but so far I haven't seen any meaningful data besides the usual benchmarks. I also suspect Haswell will show quite some teeth under heavy server load.. but still no benchies.

@Haswell in general: I think people are looking at this chip in the wrong way, or were simply expecting too much. I think Haswell is great for BOINC, even if you can't use AVX2 yet. Great in the way "a significantly better choice than Ivy, for the same price".

In regular tests Haswell has been a bit shy of 10% faster per clock than Ivy (side note: and people complain that that it consumes a bit more power.. that additional performance has to come from somewhere!). that's not factoring AVX2 in yet, and I'm sure under BOINC we could find projects where the difference is even larger. As there will be some where it's smaller.

But let's just use the average for now. And assume we're running our Haswell at a nice and smooth 4.1 GHz at slightly above 1.00 V (my Ivy can do this). Now the interesting question is: how hard would I have to push an Ivy to get similar performance? I'd say we'll need about 10% higher clocks, since performance scales slightly sub-linear with frequency even under BOINC loads.

Which means we'd have to drive Ivy up to 4.5 GHz to reach the same performance! My chip would need about 0.2 V more for this, requiring ~44% more power from the voltage alone. The added frequency of Ivy would increase its power consumption further by 10% and would probably cancel the higher power draw per clock of Haswell.

So while Haswell can not clock as high as Sandy and Ivy (due to the mobile 22 nm process being optimized for high frequencies at low voltages), consumes more power at a given frequency and thus runs hotter, it makes up for this by giving us either higher performance proportional to the added power consumption, or we could hit the same performance at significantly lower power consumption (by running at lower clocks & voltage than Ivy).

This improvement is surely too small to make you switch from an Ivy, but even Intel wouldn't expect you to do so anyway. However, if you're buying anyway then Haswell is the clear winner for BOINC.

MrS

---

Scanning for our furry friends since Jan 2002

This improvement is surely too small to make you switch from an Ivy, but even Intel wouldn't expect you to do so anyway. However, if you're buying anyway then Haswell is the clear winner for BOINC.

MrS

The Haswell prices have already dropped in the UK, and in some cases they are cheaper than SB/IB alternatives, which makes them a better choice for a new computer, at least compared against like for like SB and IB systems. In some situations however the older 6core/12thread processors such as the i7-3970X will still be the better choice (PCIE lanes).

One catch is that Haswell needs a new motherboard, so it's not an upgrade path for anyone who opted for an i3/i5 SB for example.
As well as the enthusiast 1366 and 2011 sockets we've recently had LGA1156 and LGA1155 (SB/IB) motherboards, so adding an LGA1150 socket makes you wonder how future proofed you can get with an Intel system.

Another catch is still the purchase price:
FX8350 (8core) £150, FX8320 (8core) £120, FX6300 (6core) £90, FX4130 (4core) £74,
i7-4770K (8thread) £270, i5-4670K (4core/4thread) £190, i5-4430 £146.
For crunching the i7-4770K is likely to be better for most projects than the FX8300's but it still costs £120 to £150 more, and that could get you a mid-range GPU.
The mid-range Haswell's are still expensive and while the i5-4430 might be a very competitive cruncher, there will be some projects where the 8core AMD's win hands down.
There really isn't a lower-end Haswell - there are no i3-4xxx processors, so there is nothing in the £75 to £120 range, at least for now.

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

Two things what I think. A MOBO for Intel, is almost twice as costly than and AMD MOBO, same specs with a lot of room for GPU, USB 3.0.

Indeed AMD uses more Watt then Intel, but not in my case. I have two Bloomfields, using 130W maximum, the AMD would use 125W. Okay not a win.

The plan was a new rig in the autumn, mainly for crunching, but also general use, with Office, mail, TV via browser etc. So that would be an six core Intel with good quality parts. All the Pentiums, Celerons would go out of the door to make room, as they are not used anymore. Sometimes for testing some crap software of experimenting with Linux (Ubuntu).

But now I have 2 GTX660 and they don't run smooth in the things they are in (however the T7400 not bad). So for around 400 euri I could get all the parts I need, the GPU's are already here.

But it is possible that I can't think good anymore this time and two packets of chips for dinner :)

If someone can make a nice system with Intel for 400 euri, I would like to now. Keep in mind the case must be large (roomy), seems to be the most expensive part.

---

Greetings from TJ

If someone can make a nice system with Intel for 400 euri, I would like to now. Keep in mind the case must be large (roomy), seems to be the most expensive part.

There are some good inexpensive cases with excellent airflow. I've used the Antec 300 for years. Nice quality with 3 speed 140mm top fan and 3 speed 120mm rear fan. Add a 120mm side fan and 1 or two 120mm front fans and you're set. There are also some minor variations of the basic 300 case now. Other low cost options are the NZXT Source 120 and the Corsair Carbide 200R, both of which will hold enough fans to made a wind tunnel jealous. Have used dozens of the Antec 300 and am in the process of testing the other two. For all 3 the quality is decent. A super expensive case won't get you much more in practical terms. Definitely prefer cases with the PS on the bottom.

The Bloomfield is a rather old high-end CPU by todays standards. It was a significant step forward back then, but even current AMDs beat its energy efficiency (and I mean in general, not in special AMD-friendly projects).

@SK: sure, the initial purchase price favors AMD. How much this matters actually depends on your local electricity price. I know I'd have to opt for energy efficiency in almost every choice I make around my rig, but in the US electricity is a lot cheaper.

Regarding the mainboards: I don't see much of a difference there. Sure, there are very cheap boards for AMD, but these are.. cheap. The last ones I bought for work don't last very long under sustained load. I'd say about 100€ gets you a fine board for both parties, a bit less if you don't OC.

And finally getting back to the AMD vs Intel topic. Comparing the Richland A10-6700 and A10-6800K something interesting pops up: the 6800K is typically 5% faster in applications but needs about 60% more power to get there. AMD is totally blowing their power budget here by going "full throttle". That's obviously not what we should run for 24/7 crunching.

Using the A10-6700 as basis an scaling up to 4 modules would probably give us around 115 W TDP (same uncore, doubled modules) - which is not far off what we're getting with the slightly higher clocked FX8350. So.. what to do? Scale further! Reduce voltage, maybe reduce frequency to ~3.5 GHz as well and energy efficiency won't be as bad any more. Sub 100 W should be possible without extreme tuning here. We'd loose a bit more performance compared to Intel, but if we choose some project combination where the modules work well (i.e. at least some integer code in there), through and performance/watt should be decent.

Ideally one would schedule things like this: one core of a module crunches some pure integer task (ABC, Collatz, probably ecm, maybe mind modelling) and the other core is assigned a regular FP task. This should use the module architecture in the most efficient way. The only problem is that the integer projects are.. not the most attractive ;)
However, polling a GPU should also be an integer-dominated task! But such BOINC scheduling would require some serious programming effort.

MrS

---

Scanning for our furry friends since Jan 2002

I have browsed a bit and searched only in the Netherlands (where a lot of parts are more expensive than in the US). An i7-4770K (four core), 84W, €317. An i7-3930K (six core), 130W, €530. Nice Xeons with 40W and 80W, but with prices of four digets, so that is no option for me.

Skgiven has a nice idea of getting a cheap i3 or i5 and make a rig for GPU crunching only. But I know myself then I will do more with the system and then it won´t work. Not that I am depended on 1 system. I have taken a lot of old stuff home from work...Won´t do that anymore, unless there is a T7400 :)

---

Greetings from TJ