Discussion of Intel "K" "T" and "S" parts

[Hulk]

Seems like the T and S parts run at really low voltages compared to the K's. Let's consider the 4770K, 4770S, and 4770T.

The nominal speed of the K and S is 3.9GHz while the TDP of the K is 84W and the S is 65W.
This would seem to imply that the S is simply a higher yield part that runs at lower voltage to achieve the 65W power envelope.

The T operates at a nominal frequency of 3.7GHz at 45W TDP. This would seem to be either a S part with a 3.7GHz frequency cap or an even higher yield.

Either way from this data is appears for a hyperthreaded 4/8 core Haswell part the best the current 22nm process can do with volume yields is 3.9GHz at 65W and 3.7GHz at 45W.

Is there a way to determine what Vcore would be for 3.9GHz/65W and 3.7GHz/45W?

Seems like Intel is most likely selling the K's and non K's as their bottom of the bin yields and saving the best silicon for the S and T parts as well as mobile?
Or as is usually the case, there is more to it right?

I wonder how fast and on what volts one of those 4770S or T parts would go?

 

[sm625]

Seems like the T and S parts run at really low voltages compared to the K's. Let's consider the 4770K, 4770S, and 4770T.

The nominal speed of the K and S is 3.9GHz while the TDP of the K is 84W and the S is 65W.

This is not exactly correct. The base clock of a 4770K is 3.5GHz, while the base clock of the 4770S is only 3.1GHz. They both turbo up to 3.9, probably because its just a single core turbo and the max load from a single core can fit well withing a 65W envelope.

For the 4770T you can see that the single core max turbo drops to 3.7GHz. That is kind of surprising because, again, a single core doesnt consume all that much power. So why lower the max turbo? From that information I would guess that the binning really isnt that much better for the T, if at all.

 

[Cerb]

Or as is usually the case, there is more to it right?

Yes. The T and S parts have speed limits.

Power consumption is roughly 1:1 with clock speed, so 4GHz should about 33% higher in Wattage than 3GHz. It then roughly follows the square of the voltage. So, voltage needed to run at 3.5GHz may end up causing 25% more power use than what's needed to run at 3GHz (numbers pulled out of my head to make the point).

All these CPUs have various lower-speed states, with lower than max voltage, as well. The T and S models lack the higher speed states of the non-T/S models. The model number roughly corresponds to price point, more than performance.

AMD does not use a special naming scheme, but they do exactly the same thing, with models slightly slower in clocks, compared to other models of similar price, to meet substantially lower TDP ratings.

It's mostly useful for SFF systems, passive machines, embedded systems, AIOs, etc., where a lower TDP can either allow the system to run in the environments it's made for, or where lowered TDP can reduce overall costs, given the CPU is generally the hardest to cool part.

 

[ShintaiDK]

S and T are simply TDP limited chips.

Also you walk too easily over it by speccing turbomodes. S and T models have alot harder time reaching highest turbomodes than regular chips.

4770S=3.1Ghz baseclock, 3.9Ghz max turbo. 4/5/6/7
4770T=2.5Ghz baseclock, 3.7Ghz max turbo. 6/9/11/12
4770K= 3.5Ghz baseclock, 3.9Ghz max turbo. 2/3/4/4

Take a 4 thread load. Max turbo for K is 3.7Ghz. T is 3.1Ghz and S is 3.5Ghz.

 

[Hulk]

S and T are simply TDP limited chips.

Also you walk too easily over it by speccing turbomodes. S and T models have alot harder time reaching highest turbomodes than regular chips.

4770S=3.1Ghz baseclock, 3.9Ghz max turbo. 4/5/6/7
4770T=2.5Ghz baseclock, 3.7Ghz max turbo. 6/9/11/12
4770K= 3.5Ghz baseclock, 3.9Ghz max turbo. 2/3/4/4

Take a 4 thread load. Max turbo for K is 3.7Ghz. T is 3.1Ghz and S is 3.5Ghz.

Does a 4770K at BIOS default behave differently than a 4770? My 4770k will do 3.9GHz with all cores loaded?

 

[witeken]

I think he means that 4 cores can only go 400MHz faster, while 3 cores have a 500MHz boost etc.

nvm

 

[ShintaiDK]

Does a 4770K at BIOS default behave differently than a 4770? My 4770k will do 3.9GHz with all cores loaded?

Because you got a BIOS setting on called multicore enchancement or something enabled. In short, your mobo autooverclocks your boost modes. A stock 4770(K) will never go beyond 3.7Ghz in 4 threads.

http://www.intel.com/support/processors/corei7/sb/CS-032279.htm

 

[Hulk]

Because you got a BIOS setting on called multicore enchancement or something enabled. In short, your mobo autooverclocks your boost modes. A stock 4770(K) will never go beyond 3.7Ghz in 4 threads.

http://www.intel.com/support/processors/corei7/sb/CS-032279.htm

Understood.

So running 4 cores for each 4770 type:
K provides 3.7GHz at 84W
S provides 3.5GHz at 65W
T provides 3.1GHz at 45W

Does this mean that with 4 cores loaded at 3.9GHz a 4770K is outside of the 84W TDP?

 

[ElFenix]

Understood.

So running 4 cores for each 4770 type:
K provides 3.7GHz at 84W
S provides 3.5GHz at 65W
T provides 3.1GHz at 45W

Does this mean that with 4 cores loaded at 3.9GHz a 4770K is outside of the 84W TDP?

it could, as the processor can go over the TDP for short periods if there is available thermal capacity in the system.

edit: that's at default clocks, your board is apparently overclocking your processor.

for the general consumer on this board (someone building a PC, usually with an aftermarket HSF) there is no reason to get an S or T part. the max performance is worse, the power consumption isn't better for fixed tasks (iow the parts all consume the same power when playing a bluray), and total power consumption for variable tasks is the same and may be worse (the faster parts finish sooner and go back to a near 0 power state sooner).

 

[Hulk]

it could, as the processor can go over the TDP for short periods if there is available thermal capacity in the system.

for the general consumer on this board (someone building a PC, usually with an aftermarket HSF) there is no reason to get an S or T part. the max performance is worse, the power consumption isn't better for fixed tasks (iow the parts all consume the same power when playing a bluray), and total power consumption for variable tasks is the same and may be worse (the faster parts finish sooner and go back to a near 0 power state sooner).

But aren't the S and T parts the better binned parts? Seems like 3.5GHz at 65W would be more efficient for any task than 3.7GHz at 84W. That's a 29% increase in power for a 6% increase in clockspeed? Now I realize the "worst" K would do 84W but that also means the worst S would do 65W, meaning that many of each would be even better, so it's apples-to-apples there.

For pure performance I can see why you wouldn't want an S or a T, they aren't unlocked. But for efficiency it seems like when loading up the cores they'd do better.

Now we know where all the great overclockers are going😉

 

[ShintaiDK]

Does this mean that with 4 cores loaded at 3.9GHz a 4770K is outside of the 84W TDP?

Its running out of spec and warranty is voided. However if it is above or below max TDP depends on the individual chip. I am sure you can find chips both above and below since no 2 chips are created equal.

 

[ShintaiDK]

But aren't the S and T parts the better binned parts?

No they are not. The CPUs are simply specially designed for small cases/coolers that can maximum handle 45 or 65W heat. The T model is even considered so exotic in those terms, that its only avaliable in tray for OEMs.

 

[Hulk]

No they are not. The CPUs are simply specially designed for small cases/coolers that can maximum handle 45 or 65W heat. The T model is even considered so exotic in those terms, that its only avaliable in tray for OEMs.

So ALL 4770K's will run 4 cores fully loaded at 3.7GHz within a 65W thermal envelope like the S parts?

 

[Bubbleawsome]

So ALL 4770K's will run 4 cores fully loaded at 3.7GHz within a 65W thermal envelope like the S parts?

By my understanding they would need to throttle to 3.5. Imagine that 3.9 you have as another turbo bin.

It's a bit off-topic, but what happened to the 95w i7s? If the 4770 does 84w I missed something in the past 2 generations. 😛

 

[Cerb]

So ALL 4770K's will run 4 cores fully loaded at 3.7GHz within a 65W thermal envelope like the S parts?

😕 None of the 4th gen Core i7s, of any model number, of any TDP, are rated for 3.7GHz with all 4 cores fully loaded.

Turbo speeds are to maximize performance within the TDP, when the CPU load is low enough that it won't be exceeded by running a bit faster. If you fully load the cores, you're going to get the advertised base frequency almost the whole time it's loaded down. If all the cores are turned on, but the thermal load very low, it will happily increase the speed. But, if the load is sustained, not bursty, you will be stuck at, or very close to, the base speed.

A fully loaded i7-4770S will run at 3.1GHz.
A fully loaded i7-4770K will run at 3.5GHz.

 

[TreVader]

But aren't the S and T parts the better binned parts?

If by "better" you mean more capable of running a certain clock speed with low voltage, then yes they are binned better. It is likely that if they were unlocked, they would outperform K series parts.




Intel doesn't care what clock speeds you get with a K part. The only time intel cares about what clock speeds it gets these days is when they need low voltage or low TDP parts. As long as the average K hits it's advertised boost, intel couldn't care less.

 

[BSim500]

Hulk, the problem is all CPU's are binned differently, and Intel's max TDP is simply the worst case scenario, not necessarily an accurate reading for that chip. You certainly can't base "hard" calculations off of it. I have no experience with Haswell's, but to give you a detailed breakdown of an Ivy Bridge (if it might help), I have a "77w" non-K i5-3570 that with a little mild-undervolting can run at:-

4.2GHz @ 1.124v (-0.00v offset) = 71w max possible OC
4.0GHz @ 1.016v (-0.09v offset) = 55w mild OC
3.8Ghz @ 1.000v (-0.10v offset) = 48w max turbo
3.4GHz @ 0.896v (-0.12v offset) = 39w stock speed
3.0GHz @ 0.832v (-0.12v offset) = 25w
2.6GHz @ 0.768v (-0.12v offset) = 19w
2.2GHz @ 0.728v (-0.12v offset) = 16w
1.6GHz @ 0.680v (-0.12v offset) = idle state

All above (except 1.6GHz) are LOAD voltages under 4-core Prime (MCE enabled) at varying Speedstep States on a Z77 motherboard. Its idle voltage at 1.6GHz is a very low 0.65-0.68v. And this is just a regular (but pretty well binned) i5-3570 (no letter) chip with a theoretical "77w" rating. Temps never exceed 55c Prime or 50c regular gaming / video encoding even with 1x slow spinning 650rpm 120mm silent fan in a small M-ATX case. Average gaming power consumption @ 4.0GHz is typically 40-50w modern games / 25-40w older games.

You can in fact mimic an S chip with a regular chip by "locking" in the 2nd highest Speedstep setting as the highest in Windows Power Options -> Change Plan Settings -> Change Advanced Settings -> Processor Power Management -> Maximum processor state (and change from 100% to 90%). Likewise you can "mimic" a T chip's 2.6GHz by locking it at 80% max.

You may well get a better "low voltage optimised" binned CPU with an S or T chip, but there are "golden" regular chips out there which can run at the same low voltages. T chips at 2.6GHz stock are just too slow unless you have limited cooling capacity. They're not really "more efficient" if they take 50% longer vs a 3.8GHz chip (ie, they take 24mins load vs a regular i5 which takes 16min load then idles for 8mins). S chips are better for general usage (than T's), but really, you can just buy a Z motherboard and undervolt to gain most of the wattage drop without being limited. As you can see above, even with a non-S chip, with a little undervolting, at stock 3.4GHz speed, the theoretical "77w" is virtually halved to 39-40w max load, which is lower than some poorly binned i3's "out of the box" (stock voltage).

Hope this helps. All CPU's are different with different VID's, and there may be some difference with Ivy vs Haswell for power consumption, but you get the general idea. The ultimate key to low power is not limiting multipliers, but aggressive under-volting.

 

[ShintaiDK]

So ALL 4770K's will run 4 cores fully loaded at 3.7GHz within a 65W thermal envelope like the S parts?

65W? If you are comparing to the S chip. The S chip will run 4 threads at max 3.5Ghz. While the 84W K model will run at max 3.7Ghz. Assuming both can turbo to the max.

 

[ShintaiDK]

It's a bit off-topic, but what happened to the 95w i7s? If the 4770 does 84w I missed something in the past 2 generations. 😛

It keeps decreasing. because thats what the world ask for.

SB 95W, IB 77W, HW 84W and Broadwell might be 65W. for top bins. Skylake would increase it again, only to get it lower again with 10nm. That seems to be the trend.

 

[witeken]

If by "better" you mean more capable of running a certain clock speed with low voltage, then yes they are binned better. It is likely that if they were unlocked, they would outperform K series parts.

Intel doesn't care what clock speeds you get with a K part. The only time intel cares about what clock speeds it gets these days is when they need low voltage or low TDP parts. As long as the average K hits it's advertised boost, intel couldn't care less.

Not sure if what you said is right: IAmA CPU Architect and Designer at Intel, AMA.

For us as architects, we have a team dedicated to putting in overclocking features into the designs and tests in place to cherry pick those parts to box and sell as such. So you are getting parts on the good side of the normal when you buy K CPUs.

 

[know of fence]

Expecting down-clocked parts to be better, sounds like wishful thinking, considering that these SKUs cost exactly the same. However I was surprised to learn, that it's possible to get the responsiveness of a 3.9 GHz chip in a 45 W TDP envelope. Intel thus can sell 303 $ core-i7s even in the 45 TDP space of low noise machines.

Turns out that there are a lot of tweak-able variables beyond v,f and T, like "time spent at turbo frequency", "thermal headroom", hmm "granularity", "burst performance".

Where old school OC tries to get the max performance of a given amount of silicon.
Mobile chips maximize efficiency - at a fixed tolerable level of performance.
The S and T chips offer the best possible experience out of a given amount of TDP, sacrificing a bit of efficiency compared to their mobile counterparts.

 

[ElFenix]

But aren't the S and T parts the better binned parts? Seems like 3.5GHz at 65W would be more efficient for any task than 3.7GHz at 84W. That's a 29% increase in power for a 6% increase in clockspeed? Now I realize the "worst" K would do 84W but that also means the worst S would do 65W, meaning that many of each would be even better, so it's apples-to-apples there.

For pure performance I can see why you wouldn't want an S or a T, they aren't unlocked. But for efficiency it seems like when loading up the cores they'd do better.

Now we know where all the great overclockers are going😉

a nonoverclocked K part (or regular part for that matter) probably isn't going to bump up against its rating much in ordinary use, whereas an S or T part probably bumps up against its rating a lot.

spcr is the only place i'm aware of that's done a thorough review of S and T parts vs. regular parts (even including undervolted regular parts). granted, the review is for sandybridge, but i think its still applicable because intel hasn't exactly uncorked the regular parts since then.

http://www.silentpcreview.com/intel-2100t-2400s

 

[Bubbleawsome]

My 870 is near always over TDP when I load it up. CPU-z says near 120w. Turbo with all 4 cores is 3.2Ghz at 1.3v when hot, sometimes it defaults back to 1.26 or even 1.2; but still at turbo. 😕

 

[TreVader]

Not sure if what you said is right: IAmA CPU Architect and Designer at Intel, AMA.

That literally says nothing about binning higher than the T or S chips. There is a huge difference between a chip being able to run stable at very low volts as a requirement (T or S) and a chip that has been unlocked to allow you to get some unspecified level of performance beyond spec which you may not ever achieve or even want to achieve. One useless if it isn't binned perfectly, the other is "cool" if it performs well over clocked, if you ever decide to do that, which most people with K series chips don't.

 

[crashtech]

I figured the "K" chips merely exhibit relatively higher leakage at their max rated clock, which makes them need more voltage, run hotter, and clock higher, generally speaking. I'd guess that if there is a "highest quality" Haswell i7 part, it would be the E3-128(x)v3, which is specced to run slightly faster than the 4771 within the same 84W TDP.