Current ULV Processors - How fast are they?

As we all know, most mainstream computers have gone the way of ULV Core i processors. While a couple years ago your standard Best Buy mainstream computer would have a standard voltage Core i5, now pretty much everything 15 or 17 inch mainstream has a i5-4200U or a i7-4500U.

For reference, I have a E6410 with a i7-640M, pretty much the top model dual-core i7 of the day. If I compare benchmarks like Cinebench with a i7-4500U, the Haswell ULV comes out on top but not by a significant margin. This shouldn't be a surprise as Arrandale launched in Q1 2010, or 5 years ago. However, if I move forward just one generation to the i7-2620M, it still pushes past the Haswell ULV on benchmarks, and Sandy Bridge is almost 4 years old.

My usage pattern is quite intensive and has grown in the past few years, but the core components haven't really changed. I run 2-3 VMs the majority of the time (Windows and Linux). Graphics wise, I do have some CUDA accelerated programs, but they perform on Intel iGPUs fairly well too. CPU wise, I do run statistical models and renders, but compute time spent is not that critical as it is usually overnight. When I used the E6410 (until 2013), the CPU and NVS3100m were still holding up fine but the 8GB memory limit is what was holding me back. It would be great to be able to use something like a T440s (12GB RAM) or E7240 (16GB RAM) for portability and battery life.

So, how far has ULV performance really come and what standard voltage CPU would our most recent ULV cpus (Haswell and Broadwell) be relatively close to? What are these CPUs really capable of/Would it be reasonable to expect to be able to complete my workload on a ULV processor? While the percentage increase in benchmarks between the ULV i7 and the old 640M are not impressive the ULV CPU may also not be able to sustain long periods of load at high frequencies.

 

The ULV CPU's have the same performance as the standard voltage ones till a certain limit.. For example, there is no difference in between the i5-4200u, i5-4200M except for the fact that the i5-4200u is TDP Limited which means it can't hold turbo for that long.. so that's the main problem with them.. So if you are going to buy laptops now, the only option is to get the full voltage i7 quad core...

 

As you note, you're comparing regular cpu's with low voltage versions. Don't.

Consider upgrading when the current platform(s) are available with a cpu line comparable to what you once bought at. Then the performance will be real and measurable.

Even the U processor Broadwell chips are about 30% faster (raw cpu) than i5-4250U, that is a fair and substantial comparison and upgrade to the new platform.

See:
http://www.anandtech.com/show/8941/gigabyte-gbbxi7h5500-broadwell-brix-review/3

ULV cpu's may have similar performance to a limit - but that limit is largely influenced by manufacturer's and how thin and sexy they want to make their chassis look - while ignoring any actual performance gains a new platform may offer.

My advice? Simply ignore ULV parts for systems you want the fastest performance in. Hopefully manufacturers will get the message too.

 

Many current dual core ULV chips are weaker than some Core 2 Duo CPU's, and it is especially true if their ability to turbo has been restricted. Anyone looking at ULV chips as a performance upgrade should stay far away.

 

^^^^^^

This. Times a zillion.

Going from a full-blown Sandy Bridge CPU to anything ULV with an idea of "upgrading" performance-wise - regardless of how "new and improved" Intel wants us to believe their recent offerings are - is setting oneself up for a failure.

 

I don't think it is Intel's fault if someone thinks a 'ULV' processor is the replacement/equivalent to a normal processor from yesteryear.

The failure is totally placed on the individual, not the company.

Buying a new VW in 2015 is not 'upgrading' performance in any way when coming from an circa 1980's Porsche 944 Turbo S. And the fault doesn't lie with the VW (even in GTI trim), rather, the driver for missing the important distinctions between the two model lines.

 

You do make a good point here. Way too many people don't do proper research before purchasing a PC. Or anything else, for that fact...

 

Intel dual core CPU performance hasn't improved substantially over the past 5 years but the power consumption has reduced. There are some charts in my E7440 review that include some of the notebooks you are familiar with.

Quad core might be worth investigating if you are running multiple VMs but the thermal requirements of these CPUs mean they are more commonly found in larger notebooks. I presume that you are using an SSD. If not, that could be a bottleneck.

John

 

^mhm.

Thing is that if you don't run the processor at the limit all the time - which is rare in anything other than server runs, or for example if you have some sort of simulation going constantly - then an ULV processor basically is what it is, a normal processor with better power management. Arguably, the benefits of having lower heat dispersion enables you to run at higher speeds for longer stretches of time as well, thanks to the normally very bad cooling solutions on laptops. Or, that with a normal to high load, you can suddenly run the same tasks for longer before the processor would throttle anyway. Practically, that you are able to run at sub-boost frequency with a mostly (intel special) stable watt-drain, and on comfortable temps. Or it rates the same as a "full watt" processor, if you ran them at the same core speeds. I've actually seen a few examples of the ULV processors having the ability in certain setups to run faster than the full processors, because the cooling in the setup is so horrendously bad. While the synthetic runs tend to be faster in one quick run on the better processor, before it flats out and gives you inconsistent, laggy results. Specially annoying if you rely on stable frequencies over a certain limit to avoid lag in videos, that sort of thing, where the operations complete in a short amount of time.

But you don't see that on the synthetic benchmarks run in ideal conditions. Still, important to note the difference between the celeron cores that are stripped of the extended instruction sets and don't have boost (that also are sold as "ulv" sometimes), and between the i5 and i7 ulv versions (that have comparable watt drain) that are completely normal processors as the rest, just with lower tdp rating and lower minimum/sub boost speeds.

 

Is it correct for someone to take the processor base frequency into account when considering throttling issues, or can the processor frequency drop even lower than that?

For example, the i7 4650u has a base frequency of 1.7 GHz and a max turbo frequency of 3.3 GHz. According to its TDP, this allows the CPU to work in full load with a 1.7 GHz frequency,
using 15 W on average. I think (and correct me if I am wrong) that the cpu will sometimes spike to 3.3 GHz, and sometimes fall under 1.7 GHz, but the end result will be that you maintained
an average frequency of 1.7 GHz to complete your task.

We assume that heat generation is not a problem (cooling is sufficient), so no throttling due to overheating.

 

While true, Intel's naming scheme doesn't make it easy for people to determine performance, you really have to dig deep in order to get a good picture of the CPU performance.

As for improvements, the recent architectures have seen improvements in terms of heat generated, power consumption, but not necessarily in overall performance for some. I do find it unfortunate though.

Finally, nipsen does have a point, the ULV core i are perfectly fine for some applications and workloads. However, they are terrible for some other.

Regarding how high a chip can turbo, it also depends on the chip itself, some will be able to turbo higher just like you can OC some chips to higher levels. It is true however that the 15 W TDP is quite restricting, so it won't hit max turbo 100% of the time.

The other sad thing is that Intel has a crop of full voltage dual cores too, manufacturers are however slapping the ULV parts in every notebook, even the ones that would have the required cooling capacity for the 35 W TDP parts.

 

John, while I might agree you somewhat with your statement above (bold mine), I would say that over the past five years, Intel Platform performance has increased substantially.

A cpu isn't run in a vacuum. The latest/current platforms not only decrease the power consumption by an order of magnitude or more, they also offer performance gains that can be seen/felt in just browsing modern websites. And on certain workloads, more than an order of magnitude of performance is also gained (e.g. Intel Quick Sync), but granted, that is not what most users do with their computers, day in and day out.

If computer manufacturers demanded higher performing mobile chips, Intel could deliver (at a higher power cost, of course). This is market driven, not a reflection of what performance capabilities Intel can offer with their chips.

This is exactly the same thing with SSD's. We can have an SSD that can fully saturate the SATA3 or M.2 interfaces fully and in any workload - but we would not be happy with the power required to do so. Especially on mobile systems that need to work away from wall power for as long as possible.

Balance is the key and for the last few years, Intel and manufacturers have wisely made the decision that slightly better performance but with even better battery life is the best trade off.

I can certainly see the advances made with each platform generation. While year to year they may be small jumps (especially compared to what we were once used to a decade ago), there is no mistaking the huge jump the last half decade has made for performance (mobile or not).

 

Still... same instruction set, same capability. Only differences have been the size of the level 2 cache and the clock speeds. (..modus hasn't been all that different since the pentium, has it.)

 

nipsen, that's like saying car engines are the same too. Hint; they're not. And if all they changed was a bolt and that added power or increased efficiency, then they too are improved. What improved them is not in question here. That they did improve or not is.

 

Sure.

Even if we would probably be talking about a car engine that has extremely low production cost, where designs can be updated on the fly to specification with no extra production cost, where the engine has only standard parts - and also reuse almost the entire production line from year to year. But where each car manufacturer only receive the finalized engine stuck to a chassis in welded iron in style from the 1920s. And where the engine and car manufacturers shake hands and and agree it's only natural to require you to buy a new car every year if you want to fit a new spare lid on the fuel-tank, or if you think you want a new windshield wiper. While also praising the fact that we were so lucky to get the minuscule, but allegedly crucial yearly updates to the engine stuck on the 1920s cold-hammered iron chassis. And if you complain, well, the car manufacturer can give you a new colour and new seats. Obviously the warranty is void if you want to switch out the headlights, or if you cover the trunk with a new carpet (since the exhaust has proven to ignite non-standard carpets made out of folded chainlink fence).

 

I believe that notebook manufacturers have a limited amount of freedom in setting the operating limits on the CPU package. I've just run wPrime while watching the CPU speed and power using HWiNFO. The peak CPU package power was 21.1W but the CPU settled down to 18W at 3GHz running 4 threads. Adjusting the maximum processor state revealed that the CPU package used about 13.3W at 2.6GHz, 119.W at 2.4GHz and 9.6W at 2GHz.

So why does Intel give itself so much wriggle room by declaring the base frequency as 2.1GHz? It's because of the GPU which is also part of the CPU package. Concurrently fully loading the CPU and GPU will result in the CPU speed reducing from its potential maximum speed. I suspect, but haven't tried to measure, that the GPU won't reach its maximum speed if the CPU is also using significant power.

You can gain useful insight into the speed and power relationships by looking at this set of Intel numbers. They show how a modest increase in speed results in a substantial increase in the power requirement.

Unusually, I agree with you about this. If people look at the charts I linked to in my previous post then they will see that factors such as having an SSD make a big difference in the combined benchmarks which attempt to be a better indicator of real life usage than raw CPU performance. Also, the incorporation of the Northbridge into the CPU package (Sandry Bridge onwards) provided a substantial increase in the memory bandwidth which helped the system performance and, particularly, the Intel integrated GPU.

John

 

Thanx for the reply.

I saw from some stress tests that when both the CPU and GPU are under load, at high temperatures the processor clock frequencies are greatly limited for the benefit of the GPU.
For example in the following review, an Intel Core i5-4250U 1.3/2.6 GHz cpu runs at 800 MHz for the CPU and 600-950 MHz for the GPU under extreme load:
http://www.notebookcheck.net/Review-Apple-MacBook-Air-11-Mid-2013-i5-1-3-GHz-128-GB.96570.0.html

We may argue here that the throttling occurs because the temperature in the MAC exceeds 90°C, but even if the temperature is not an issue, like in the following review, extreme loading of both the CPU and the GPU result in very reduced frequencies: http://www.notebookcheck.net/Asus-AsusPro-Advanced-BU401LA-CZ020G-Ultrabook-Review.134034.0.html

So, this is the case where temperature is not an issue, but frequencies are nevertheless reduced after a while so that the TDP is not exceeded infinitely (assuming an infinite benchmark).

To conclude, even though the TDP of the GPU is integrated / taken into account in the overall TDP of the CPU, there are some benchmarks that limit the frequency of the CPU even below its base frequency.

 

Looking at that review I see that something is controlling the overall CPU power management to achieve a stable temperature of a relatively cool 65C. My Dell E7400 seems happy to sustain a temperature of around 80C. Such thermal management rules are coded into the BIOS. Intel's own critical temperature is 100C but notebook manufacturers may be concerned about the long term thermal stresses or having very hot components too close to the shell of the computer and hence apply their own, lower, limits.

Nonetheless, with that Asus notebook having two fans I would expect it to be able to sustain full power and run relately cool. It's possible that Asus copied the thermal and power management rules from another notebook with a less good cooling system. I recall another redent discussion in this forum about a notebook with unnecessary throttling. Here it is. Coincidentally another Asus, albeit with a more powerful CPU.

John

 

I'm not sure, I think the throttling occurs because it wants to limit the power dissipation to the 15W TDP requirement. If we read further down in the review we see the following fact:

And let's also take into account the following fact, found earlier in the review:

Assuming that temperature is not an issue, I suppose the behavior of the CPU is like this: in extreme load the CPU immediately over-exceeds its TDP, and after a few seconds is mini-throttled to a performance that still exceeds the TDP but not by so much as earlier (probably running somewhere between base and turbo frequency). After a while (which is not specified in terms of minutes or seconds in the review), the CPU is further throttled to exactly meet the TDP requirement.

If temperature or/and power supply pose an issue, such over-throttling can come much sooner, like in the case of the MAC:

It is a little confusing (at least for me) because it makes you think that if you run a program for an hour, completely utilizing 2 cores, you will be running it at a frequency of 800MHz after a couple of seconds or minutes, but I don't think that will happen, because realistic programs will not have the high-complexity workload of those benchmarks.

edit://

Well I guess every manufacturer does different stuff and the performance is very dependent to the manufacturer as you said. This is from a Thinkpad T440s review with the i7 4600u 2.1/3.3GHz CPU:

And in an older Thinkpad T440s review with the i5 4200u 1.6/2.6GHz CPU we have that:

 

1. It is usually permissible to significantly exceed the design power for a short period. That's what I've seen in reality. It's easy to check by first running HWiNFO and opening the sensors tab and then running a program to load the processor. HWiNFO's display includes a column for the maximum value (which may not be the absolute maximum due to the sampling interval (1s?).

2. However, when wanting to reduce the power it is normal practice to do it dynamically one step at at time until the power ceiling and/or thermal ceilings are satisfied. Power changes can be monitored and adjusted in very short timesteps while thermal adjustments usually involve monitoring over longer periods.

I cannot think of a good reason for a CPU left under load for long enough to slow down to, or below, its nominal base frequency and I haven't knowingly seen it happen.

John

 

The following image is for a 3720QM, but the base principle is the same for Haswell ULVs. This was taken from XTU, if you look at the specs there is a turbo power max at 45 W, the standard TDP and a burst turbo power max at 56 W. The current ULVs have more power thresholds, but the principle is the same, they can be set to exceed their TDP for a certain amount of time.



Regarding throttling, it is possible for the CPU to hit as low as 800 MHz if it hits Intel's max prescribed temperature, usually 100 or 105 Celsius.

 

It should be quite easy to get the CPU to throttle back to below its base frequency. Concurrently run a GPU and a CPU burn program.

 

Not necessarily, usually, the IGP portion is the one that's gonna eat it and downclock, not the CPU portion. That's at least the behavior I've observed in the rigs I have had access to.

 

...I'm going by memory now, but I think that one of the things they were selling early on about haswell was that they were turning off cores, while reporting them as active ("active idle", or something). Which I suppose in theory would allow the existing networked thermal management to work better (along with the high-level scheduler being automatically more efficient, since threads won't be closed or rely on different cores), to allow the processor to run without having to actually turn off alus that are in use.

Which is probably what happens when the forced throttling takes place, during overheat, on haswell too. That individual alus are switched off for a short amount of time, while the bus and active processor elements maintain the base speed. My guess is that you would likely be able to measure increased latency on the threads on the affected cores from the program layer when that happens, even if the clock speed is the same as before.