TR Forums

So the HASWELL launch has come and gone. A number of previews mentioned a delicious, if overpriced i7-4765t model. You know, really, really low TDP for an i7 part.

So where is it?

I always got the feeling that such parts were destined for big system builders, but the "processor chart" never listed this item as OEM only. Not only that, I don't think I've come across many systems offering this chip inside. So what gives? And where can I go to get one if newegg, and amazon and the like don't carry it?


-GB

Forgive me if I've already asked this (I've been on a forum-skimmed diet lately and my memory is pretty poor in that regards )

Geeze google, thanks!

You know, after searching for weeks for this thing, Google finally turned up a different result.

http://www.acmemicro.com/Product/12680/Intel-Core-i7-4765T-Haswell-Desktop-processor-CM8064601466200-LGA1150-Quad-Core-2-GHz-8MB-35W-Tray?gclid=CJyPybHm_7gCFctlOgod4j8AAA

Shop google to the rescue. Sort of. Overpriced even more, and many items listed as used. But at least there's something here.


I still ask the same question to those of you who build more often then me. Where should I look to find "less mainstream popular" parts like the 4765T (I noticed that low TDP desktop processors are definitely harder to find). Do I have to start trolling "enterprise IT part suppliers" like CDW and the like?

I'll assume you're specifically looking for a 35W CPU.

I don't have any data to back this up (which usually sucks a lot), but are you sure the ultra low TDP version consumes so much less power at the same clock speed as the vanilla edition so that it is worth an extra €100?

You could consider buying and downclocking the regular 4770 using the multiplier to similar or equal clocks and voltage.

Actually Mouser.com sells it.

http://www.mouser.com/ProductDetail/Int ... QgodqSoAdw

It's even in-stock.

OK, you guys rock.

As for why I want it, I totally admit that I'm just hypnotized by the thought of a 4-core, HT enabled i7 comes in under 35TDP.

Way overkill for a HTPC/Steambox build I suppose, but I've always ended up going in for a little excess whenever I build a rig.


That being said, if I have a moment of sanity, I suppose really all I need is an i5-4570S 65W TDP part, and I could even clock that down a bit if I wanted, huh? Almost $200 makes for quite a bit of change (that can also be re-invested in other components).

Sigh, damn sanity.

Worse, I'll tell you a little secret - the actual in practice temperatures will be pretty close with the same cooling setup. Don't fear the i5, though honestly I doubt I'd even use that for a "gaming" htpc, I'd use a Core i3-3240, $125 and still runs the games at nearly the same FPS. Or you could be bold and go with the Core i3-2120T - it's very available and $135 but it runs at 35W. It's actually what I have in my HTPC and it works great with my passive HD7850.

I'll tell you guys a little secret too. I'm currently sporting an irrational fear of dual-core on my HTPC.

I've been trucking along on a little Core2Duo for years now, mostly happily. Except that I've noticed I get "the hitches" on WMC every so often, often shortly after a wake up (I can start browsing the chanel guide immediately, but then 10 seconds in, WMC freezes for like 30 seconds -- windows is still plenty responsive though, I can cut out to desktop and kill processes, etc, Also, occasionally get short hitching in blu-ray playback).

Of course I cant positively identify it as a CPU availability problem, and modern dual cores are certainly better than the used to be, but still...

Anyway, that's my confession.

You may tru undervolting a normal i5/i7 and you may get lucky? That may be more economical.

Is there any reason why you specifically would like this CPU? Most of the difference between the high TDP and low TDP parts is literally just a setting. I would think there's not a ton of noticeable binning going on with the i7 line and a mature 22nm process.

My point is, why wait/pay more for this CPU when you could just nerf a cheaper i7 down to your own TDP/frequency/voltage specifications?

Most motherboards allow you to change the max TDP of a CPU. I have my home server (Sandy Bridge Celeron, 2x2.6GHz) sitting at a 35W max TDP instead of the stock 45W. It's a bit cooler/quieter at load now. I haven't looked into hardware undervolting before, though I know it's also common. The TDP rating is the easiest to just set-and-forget though.

Hmmm, never paid any attention to seeing if there was a max TDP setting in the BIOS. That could be a one-stop lazy way to getting what I want.

and I have been leaning towards that i5 the more that I think about it. TDP or undervolt it down and hmmm...

Duct Tape Dude wrote:

Most motherboards allow you to change the max TDP of a CPU. I have my home server (Sandy Bridge Celeron, 2x2.6GHz) sitting at a 35W max TDP instead of the stock 45W. It's a bit cooler/quieter at load now. I haven't looked into hardware undervolting before, though I know it's also common. The TDP rating is the easiest to just set-and-forget though.

That is actually a relatively new BIOS feature that I have not seen before (my last motherboard is a P55 chipset one). Setting max TDP can be simpler. Which motherboard is in your system? Undervolt (tested of course) can maintain the same clock. The auto max TDP may involve the motherboard messing with clocks though.

I have a Biostar H61 board (http://www.newegg.com/Product/Product.a ... 6813138335). I think TDP became a settable parameter starting with Sandy Bridge, since power targets became a thing for Intel after Nehalem.

The CPU is a Celeron G550 (http://ark.intel.com/products/53418), which apparently is 65W (I was mistaken earlier). Runs just fine at a nerfed 35W TDP. I opted for the Celeron because unlike previous generations, the SB Celerons had all power-saving features enabled (in some past gens, SpeedStep was disabled).

Tangential rant:
A few years after I got my first laptop (Core 2 Duo Penryn) is when I discovered TDP itself is an interestingly useless number. It gives an approximation of the designed cooling required, but the actual power consumption can vary wildly. Everyone knows more voltage+frequency = more power, yet there's a wide array of processors that vary in frequency/cache that have the same TDP (for Penryn, there's everything from the 2.2GHz 1MB single core Celeron 900 to the 3.06GHz 6MB T9900 all in a "35W" TDP rating). Obviously the Celeron 900 will run cooler at load, even with binning differences. Penryn was a special class of CPU since it had a 25W mid-power P series (ex: 2.26GHz 3MB P8400), but the highest-end P series was bumped to a 28W rating (2.8GHz 6MB P9700). It ran noticeably hotter. I wish that there was another (marketed) TDP rating that indicated base and maximum power usage. It'd be much easier to see that number scale across different CPUs.

As most of us know, you can't stop physics, and TDP is just a marketing number to show roughly what class a CPU is pre-tuned to. And since it's just a marketing scheme, there is no harm in gaming that scheme by setting your own power targets lower on a CPU with your desired processing features.

Flying Fox wrote:

That is actually a relatively new BIOS feature that I have not seen before (my last motherboard is a P55 chipset one). Setting max TDP can be simpler. Which motherboard is in your system? Undervolt (tested of course) can maintain the same clock. The auto max TDP may involve the motherboard messing with clocks though.

I know Asus Z87 boards allow TDP adjustments in the AI Suite. (anandtech video interview here)

Even my crappy Gigabyte Z68 mobo (Z68AP-D3) in my testing rig allows for setting of the max CPU power consumption.

Setting it to 35 watts with a 2600K results in great performance for lightly threaded workloads (since it doesn't really throttle clock speeds) and reasonable power consumption when fully loaded though the clocks obviously take a pretty hefty hit.

Please note that Intel has locked out power limits below roughly 75-80% of the stock TDP value on the standard TDP parts (not sure about low TDP parts, I believe eXtremely-expensive and -U/-Y chips have no TDP lower limit) to prevent what you're suggesting. This has been a point of consternation on my i7-3720QM, which will not go below 36W (stock is 45W).

You can test this out using throttlestop using TPL (turbo power limit) with clamping mode enabled; if it isn't supported, power draw will be clamped to the lowest value programmed into the CPU. This lower power limit is stored in a processor MSR, so you can read it back: see http://download.intel.com/products/proc ... 253669.pdf page 24 - minimum power limit

I've tried these sorts of things on the Xeon chips (which don't have these power limits - and it would make sense to have this feature for server power capping) and it's amusing to see the max multiplier drop to the minimum, then on-demand throttling take over as the minimum multiplier results in power consumption above the power target.

Are you positive about that applying to all chips (including desktop)? I think I need to retest and check real power consumption.

Given Intel's rather diverse implementations of product segmentation, it certainly strikes me as very likely.

Why bother making a low TDP chip if people can just take any model and make it do the same thing?

That's a pretty significant waste of expensive wafers and equipment.

Perhaps, but even with limited TDP a full-fat leaky 2600K/3770K/4770K is going to use more power than the S or T variants at similar performance levels.

Waco wrote:

Perhaps, but even with limited TDP a full-fat leaky 2600K/3770K/4770K is going to use more power than the S or T variants at similar performance levels.

This is at least only the case when the -S/-T chip is a dual core, as those chips are cooler.

Intel chips really have four modes of operation in this regard:
1) Guaranteed speed - traditional mode where the OS asks the CPU for a specific speed and the chip gives you that speed (3770K = 3.5GHz)
2) Turbo speed, frequency limited - no clock speed guarantee to OS, but we are below the power limit and at the max ratio limit
3) Turbo speed, power limited - no clock speed guarantee to OS, but we are at the power limit and must run below the max ratio limit. Note that the CPU frequency will never drop below the guaranteed speed, even if this results in TDP violations (ie. via power virus)
4) Power limit with clamping enabled - this is nonstandard operation. Enabling the clamping bit allows the CPU to go below the guaranteed frequency to hit power targets. This is nonstandard, I think, because power states no longer have any meaning for the OS (ie. no longer a guaranteed frequency).

Adjusting the max non-turbo ratio is the way the S/T chips do it, as they have very low guaranteed frequencies which essentially cannot exceed the specified TDP even under power virus conditions. The way S/T chips are implemented can be simulated via programming the max non-turbo ratio (it's a CPU MSR) to a specific low value, and using turbo mode to control power consumption. Not sure you need a K-series chip but it may be necessary.
I've done this (setting the max non-turbo ratio to 12x) on my 3720QM before to try going below 36W, but it didn't work.

It's the case much of time time since the K series chips tend to be leaky where the S and T chips are binned for low leakage and low voltage operation.


If I have time tonight or tomorrow I'll play on my test rig again and see what it really does when playing with the power settings.

Aren't the low-power desktop i5s and i7s designed for SBCs, mITX and other ultra-compact implementations where everything's OEM and soldered together? I'm actually kind of surprised that someone is selling it as a bare chip...anyone considering buying one may want to double-check with the retailer that it's compatible with a standard motherboard socket.

Waco wrote:

It's the case much of time time since the K series chips tend to be leaky where the S and T chips are binned for low leakage and low voltage operation.


If I have time tonight or tomorrow I'll play on my test rig again and see what it really does when playing with the power settings.

Leakage is proportional to voltage, and if the S/T chips can't clock high enough to use high voltages, then you would be right. But it doesn't mean that a K-chip can't do the same thing if we limit its frequencies (and by extension, voltages - Nehalem and later chips automatically ramp voltages with frequency).

I'll only believe this when I see it myself... my guess is that a K-series and a S-series chip, which are both rather high bins of the silicon stock, would be capable of very similar things limited to the same power level. In the old Pentium M / Core 1 ULV/LV days, the Standard Voltage chips were locked out of the lowest voltages below 0.95V whereas the ULV/LV chips could go to 0.7V. If Intel is doing voltage lockout with the K chips but not the S/T chips, you would be correct, but I don't see evidence of that anymore: the VID is now requested by the CPU uncore instead of the chipset and a standard voltage chip can request as low as ~0.8V now.

Leakage is proportional to voltage but I'd be surprised if there's a lot of bin overlap in terms of leakage between the K and S/T chips.

Leaky chips tend to clock better (at least, they have in the past) so it would make sense if the K chips, even at the same voltage and frequency, used more power.

I'll tell you guys another secret.

People are wrong about using the stock chips and just by downclocking/undervolting them means you are automagically ok. (or that they even boot or don't crash, or that the board supports it, or that it works outside of winblows)

They can still overheat in real world use when there is no headroom, running 24/7 in higher temp industrial and/or fanless. Been there, done that, that home cpu idling away most of its lifetime doesn't prove anything.

i7-4765T is pretty awesome considering it has all cpu features in 35W, is 4C8T, still turbos to 3ghz for that nasty irq-heavy single thread and has my favorite, AES-NI. mmmmm VPNs

Now if only they would come out with an update to the thin itx board so I could build more, those embedded 1u rack servers are total ripoffs.

Of course if you're building a home idlebox, go cheap. Also go to microcenter, they always have a handful of cpu/board combos deals that are untouchable online.

Bauxite wrote:

I'll tell you guys another secret.

People are wrong about using the stock chips and just by downclocking/undervolting them means you are automagically ok. (or that they even boot or don't crash, or that the board supports it, or that it works outside of winblows)

They can still overheat in real world use when there is no headroom, running 24/7 in higher temp industrial and/or fanless. Been there, done that, that home cpu idling away most of its lifetime doesn't prove anything.

That's why you need to stress test the heck out of a undervolting/underclocking set up the same as you test an overclocking/overvolting setup. It is all about finding the limit, just in the other direction.

Flying Fox wrote:

That's why you need to stress test the heck out of a undervolting/underclocking set up the same as you test an overclocking/overvolting setup. It is all about finding the limit, just in the other direction.

This. Anyone who changes anything related to clocks or volts and doesn't stress test is just asking for trouble.

Just want to follow up. I bought the Core i5. Every so often, I get an in itch to pick up the 35W i7 and return the i5 (especially since it isn't here and unboxed yet), but those moments are getting less frequent.

....thanks.... I think.


Also, thanks for the links, the input, and for just taking this thread and running with it in another direction. Man, I love TR.

Not sure if this helps since you're on Haswell, but Anandtech did an article undervolting an i7-3770K. It's a worthwhile read.