It's actually the other way around: link 1 and link 2.

MrS

---

Scanning for our furry friends since Jan 2002

It's actually the other way around: link 1 and link 2.

MrS

Wow!
Then this is a completely different Haswell CPU. I didn't thought that Intel will actually make 2 different chips (just disable the 'unnecessary' parts in the lesser chips)

Ideed, it's almost quadratic rather than rectangular like with GT1 or GT2 grapgics. Doubling the EUs from 20 to 40 increases the iGPU die size significantly. And especially with the numbers Intels sells it would be far too wasteful to only work with deactivated units in this granularity.

MrS

---

Scanning for our furry friends since Jan 2002

It also makes one wonder if all this heat buildup is only due to the poor transfer material or if Intel is being less than honest about the power draw of Haswell. Low power ratings sell processors...

They use a poor (cheap) TIM because of the low TDP there's no need for better at 3.6GHz (considering everyday usage - no, crunching is not an everyday use). It's a shame however, that they don't use a better TIM (e.g. soldering) in the K series, which definitely made for overclockers (and costs more while some features are disabled). So the way to heavily overclock a K series CPU leads through IHS removal and TIM replacement - voiding warranty. In my opinion this is intentional from Intel's part, they don't want to give warranty for heavy overclockers.

So the Intel desktop strategy is: minimal performance "upgrade", increase power consumption, charge twice as much as the competition for substandard construction. Is that the gist of it? We've seen this before from the big "I" (Only the Paranoid Survive).

So the Intel desktop strategy is: minimal performance "upgrade", increase power consumption, charge twice as much as the competition for substandard construction. Is that the gist of it? We've seen this before from the big "I"

Tides are turning. We all have to bow to the fact that the desktop segment became less important (less profitable etc.) than the mobile segment, so the desktop strategy is subject to the mobile strategy. The desktop computing is the past, it has only a couple of years left.
It depends on how you define "performance". In the terms of computing efficiency, the 4xxx series is better than the 3xxx. However, there's no point in upgrading from the 3xxx series for those who already have that. The 4xxx series has a larger iGPU, so using the same manufacturing technology as the 3xxx series, it's quite logical that its power consumption will be higher.

(Only the Paranoid Survive).

And the lucky too:)
BTW the success of mobile computing made Intel (and AMD) paranoid, as their products are not as power efficient as the competition's.

If I understand correct, that if I would like a new CPU and not use the iGPU and be power effective, I could better buy a 3xxxx series?

---

Greetings from TJ

If I understand correct, that if I would like a new CPU and not use the iGPU and be power effective, I could better buy a 3xxxx series?

No, if you don't use the iGPU, it won't consume much energy.
If you intend to buy a new CPU (or GPU) for crunching, don't buy ones built on older technology.

So the Intel desktop strategy is: minimal performance "upgrade", increase power consumption, charge twice as much as the competition for substandard construction. Is that the gist of it? We've seen this before from the big "I" (Only the Paranoid Survive).

Nope. The strategy is "don't do more than you have to in order to stay profitable".

minimal performance "upgrade"

Increasing IPC by almost 10% is actually a lot, if the starting point of the optimization was as refined as Ivy Bridge! For an evolution of the architecture that's quite good, since the low-hanging fruit have pretty much already been picked. There are certainly diminishing returns if you try to push CPU performance further.

You could argue that you'd rather want more cores or something truely revolutionary.. but Intel would argue that more cores wouldn't benefit most users that much, especially not with dual channel memory, would blow the mainstream power budget and would happily sell you socket 2011 if you really want them.

For massively parallel FP code you could push much more aggressively for higher performance.. but AVX2 already adresses that. And pushing much further would result in something like including Larrabee cores.

increase power consumption

Nothing's wrong with that as long as performance also increases. Then you can achieve the same performance at lower clocks & voltages and improve load power efficiency. I don't really get what people expected from Haswell, at the same node as Ivy.. higher performance and lower power consumption? You can have either, but not both.

charge twice as much as the competition

... for a faster and more energy efficient CPU, sure. Compare prices for CPUs of roughly equivalent performance and it's going to be Piledriver vs. i5 or Richland vs. i3. You can make the point that AMD offers good value here (depending on what's important to you), but Intel charges nowhere near double the price for these models.

for substandard construction

What does the construcion matter if the product is still excellent? Sure, for the K models using the TIM can not be excused.. but for stock CPUs it hardly matters.

The 4xxx series has a larger iGPU, so using the same manufacturing technology as the 3xxx series, it's quite logical that its power consumption will be higher.

No, it's actually the CPU part that draws more power. The iGPU is completely power gated if it's not being used and practically doesn't draw any power at all.

MrS

---

Scanning for our furry friends since Jan 2002

The 4xxx series has a larger iGPU, so using the same manufacturing technology as the 3xxx series, it's quite logical that its power consumption will be higher.

No, it's actually the CPU part that draws more power.

I still have much to learn about Haswell. :)
For those, who want to compare the power consumpion of their 3xxx series (or other): according to HwMonitor, my 4770K@3.7GHz (1.085V) consumes 52.57W~56.62W (IA Cores: 43.71W~47.74W, Uncore: 9.55W~10.22W) while crunching 4 rosetta's and 2 GPUGrid tasks. (The iGPU is turned off)

Could you do a quick test and change CPU voltage? Maybe go straight for 1.00 V at default clock, I'd expect it to work. The reason I'm asking: from experience with my Sandy and Ivy I suspect the reported power draw does not take the actual voltage into account (so I don't know which voltage). I'm 100% sure about this for the reported power draw of my HD4000.

So in essence.. I'm not sure these reported power draw numbers are any good.

MrS

---

Scanning for our furry friends since Jan 2002

I have built my AMD rig (sooner than planned and in an old case) it has the Sabertooth mobo and FX8350 black edition with stock CPU cooler.
I like this set up as it is drawing only 230Watt when running 8 fightmalaria WU's.
That did never happened with my Bloomfield. The CPU temperature stays at 61°C with 8 cores full working. They are all at 100%, that is also better then I saw on the Bloomfield. When crunching GPUGRID with the GTX770 the CPU temperature does not change but power draw increases to 320Watt with non CPU tasks and 384W with 4 CPU tasks.
When the system is idle it is using 120W. This is awesome low compared to my Bloomfields.

The only negative is the noise of the CPU stock cooler, that is like a jet engine.
I am not saying I am an AMD fan now, but there will be a second one in the future.

---

Greetings from TJ

I just replied in some other thread.. 120 W idle is much better now than 190 W. Yet.. it's still far too much for a modern system, you should be seeing around 70 W.

MrS

---

Scanning for our furry friends since Jan 2002

I just replied in some other thread.. 120 W idle is much better now than 190 W. Yet.. it's still far too much for a modern system, you should be seeing around 70 W.

MrS

Yes I saw your post, 190W was my mistake. When idle its a bit varying between 101-120W. I will check power management.

But the PSU isn't modern that one is from 2009 a M600 80Plus Bronze.

---

Greetings from TJ

I've been reading some interesting performance observations regarding the choice of CPU for the Titan/GTX780.
Basically, it's been suggested that there might be some gain from using the Intel PCIE3 on-die controllers rather than an AMD chipset.

Can we give this due consideration by continuing this thought-provoking discussion here?

Ref: http://www.gpugrid.net/forum_thread.php?id=3440&nowrap=true#32274

FWIW, my TITAN does them in 4300-4600 with no threads reserved. I will try reserving a thread to see if there is any difference. This is not in the DP-enhanced mode. So it will automatically OC up as must as temps allow.

After more testing, I see that my TITAN takes 4520 seconds with or without a reserved thread. No difference at all.

That's goos news.
Zarck's Titan still needs 4700 secs. Then maybe the AMD architecture is to blame for that. The AMD FX CPU don't have integrated PCIe controller, it uses a Hypertransport link to the North Bridge. The AMD 990FX NB has "only" 2x PCIe 2.0 x16 support, while the Intel i7-3770 and 4770 has (only one) integrated PCIe 3.0 x16. The PCIe 2.0 x16 is quite enough for the GK104 (up to the GTX 680 and 770), however it could be hindering the performance of the GK110 based cards (GTX780 and Titan), because they have 50% and 75% (respectively) more CUDA cores than a GK104 based card.

---

FAQ's

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

I expect the next generation to enter Amd PCI Express 3.0.



@+
*_*

---

Some units with my Titan is not more than 66% load, why?

https://www.dropbox.com/s/0qrhfzoxkb5446g/GpuGridTitan.jpg

@+
*_*

---

You are running 8 Rosetta WU's - Try running 7 Rosetta WU's; set Boinc to use 99% of your CPU. Post your findings (including CPU usage from Task Manager, and run relative GPUGrid run times).

---

FAQ's

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

While other GPUGRID units running at over 84% without changing the settings?

https://www.dropbox.com/s/stx56zuv50atzqj/GpuGridTitan2.jpg

If I run Folding I do not have this problem.

https://www.dropbox.com/s/t5fs17aq678uahq/Folding.jpg

@-
*_*

---

I cant see the names of the respective GPUGrid WU's but there have been several different WU types, and different work utilizes the GPU to different extents.

The Rosetta WU's might also tax the CPU more at certain stages of the run, or your chipset might be limiting performance somewhat. For any given GPUGrid WU, if you change the CPU usage from 100% to 99% you will either notice a change in GPU utilization or you won't. If GPU usage rises then its clear that the CPU usage is an issue. If you don't see any improvement then the CPU usage is probably not hindering GPU computation to any great extent (though you would really need to test this against several WU's to get an accurate performance measurement). This varies depending on the GPUGrid WU type and what the CPU is doing and possibly it's reliance on RAM...

Folding is a very different application to ACEMD. They have different GPU usages and use the CPU to different extents (ditto for apps at other projects such as MW, Albert, Einstein...). This thread is about CPU comparisons, so if you want to compare Folding to ACEMD start another thread in GPUGRID CAFE or Number Crunching.

---

FAQ's

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

@Zoltans comment: there's also a latency penalty involved with having to go through the chipset northbridge on AMD versus PCIe controllers integrated into the CPU for Intel. Usually this doesn't matter much because the latency is still high on Intels compared to whatever CPU and GPu are doing locally, so GP-GPU apps have to by written to avoid communication and especially time-critical communication. And there's PCIe bandwidth, of course.

@Zarck: I agree with SK, this really doesn't fit this topic. One small comment, though: the GPU utilization at GPU-Grid varies with WU type and more specifically the size of the simulated system. It's normal to see lower utilization on the short queue, for example. So a certain range is totally expected, especially for high-end cards.

MrS

---

Scanning for our furry friends since Jan 2002