Does temperature affects performance?
- Thread starter MistaPi
- Start date
Most modern processors will operate at full speed into the high 80's or more. The i7 series will not clock-throttle until 100*C; I think Fermi is even higher, and the Radeon 5xxx series are probably in the low 100's if I recall correctly. I know that older C2D-based CPU's typically hit the ceiling around 85*C (my Q9550 and E8400 both have this as their TJ point.)
Last edited by a moderator:
Electron drift speed is generally indifferent to temperature, it's more a function of the material it's passing through. However, electron mobility is influenced by temperature, but the change in temperatures that you're talking about being able to achieve is still so small (relatively speaking) that it won't be measurable in terms of how your applications will perform.
Last edited by a moderator:
Chips can be clocked to considerable higher clock speed on lower temperature.
And there are no "separate power(temperature) and clock speed walls", they are the same wall, you can just come to it from different angles.
Some chip may be completely stable at 2.5 GHz at 60 celcius(normal operation), requires downlocking to 2 GHz at 110 celcius, but may be able to be overclocked to 3 GHz at -10 celcius.
There have been companies developing refridgerator cooling systems, for "factory overclocked systems", for example there was company called Kryotech which sold some 700 MHz Athlons "factory overclocked" to 1 GHz, with a -40 celsius cooling.
By this do you mean you can overclock until you reach the max operating temp ?
If so ive seen cpu's fail at a certain clockspeed even though the temp is still very low
No, not really.
The "clock wall" in a given SOC will be the minimum signal propagation time of the 'longest' circuit. The "power wall" in a given SOC will be the maximum current that can be carried by the individual traces.
Using more power to drive amplitude to ensure solid clock "ticks" is one thing, but there will come two very separate points where no more power can be applied, and yet more clock is possible -- or where the signal routing is such that no more clock speed can be attained, and yet power delivery could still be increased.
Temperature in SOC's is a measure of inefficiency; super-cooling the device can (in a very tiny way) increase electron mobility, but it doesn't really make the overall computation capability faster. As proof -- you could take two identical chips, super-cool one of them to basically 0K and leave the other at room temperature, and so long as they're clocked the same, their computational performance should be essentially identical.
Supercooling allows you to circumvent the eventual thermal issues that result from all the electron drift -- more drift comes from more power. You don't supercool to accelerate electrons, you supercool to ensure the medium doesn't burn up from all the energy dissipated from drift.
Well, yes and no. At 45nm, the difference in rise and fall time can be pretty big, IIRC +/- 10% when going from say 60 to 80.
But it doesn't matter. The rated CPU clock is (most often) characterized at 125C.
Mintmaster
Veteran
Just so you know, drift velocity is proportional to mobility at a given voltage, so you can't have one stay the same and the other change.
MistaPi, the real reason that there's no real impact is that the signal speed - which is orders of magnitude faster than the electron velocity - is independent of temperature. If you're curious about the difference, consider that it takes a second or two to suck water up a straw, but a tenth of a millisecond for your mouth's sucking force to travel down the straw and show movement in the cup.
No, exactly the opposite.
There is no such thing as "max operating temp" , other than the temperature where the chip starts to melt.
If you run your chip with low clock speed, or high voltage, the chip can be stable even at very high temperatures. (like 120 celsius)
Erm, signal speed in CMOS has everything to do with electron mobility. A transistor doesn't switch the minute an electron is nudged on its gate. A sufficient number of electrons has to either be pushed onto the gate (discharge) or attracted away from the gate (charge).
When the charge accumulated on the gate is enough to create a potential at the threshold voltage of the source-channel junction, the transistor moves into conducting range (well linear region at first, but MOSFETs will quickly move into saturation).
How fast the previous transistor can charge up the gate of the transistor its driving as well as the wire connecting them is mainly a function of how much current that driving transistor can supply -- why do you think performance characteristics of process technologies are all measured in Ion (current while on) and Ioff (current while off)?
Current supplied is a function of temperature (which affects electron mobility), voltage, material (base mobility, dielectric constant), and the dimensions of the transistor. There are also secondary effects that are starting to make a big difference at the smaller (45nm, 28nm, etc.) geometries such as diffusion density, temperature inversion, etc.
Interestingly, starting at around 65nm, the carrier mobility is actually highest at around ~0C. Lower temperatures actually reduce electron mobility due to inhibiting drift.
Well, unless you have an asynchronous chip.
Power and clock "walls" wouldn't be the same there, either.
Mintmaster
Veteran
Okay, but I was distinguishing between signal speed and switch speed. MistaPi was asking about how fast electrons move, presumably thinking about them moving across the chip, and I was just telling him that by itself that's not an important metric.
I should have been more clear, though.
Sure it does. A minute is a long time
Point taken. Switch speed does indeed depend on resistance, which in turn depends on mobility. I didn't know that the max was at 0C, so thanks for the data.
But it is an important metric. You're right in that a signal isn't considered changed by when an electron travels and arrives, but the rate at which electrons (which carry the electric field that is used as the "signal") can move is very much at the heart of how fast transistors in circuits switch on or switch off.
In your straw analogy, if you can move the liquid faster, say use water instead of milkshake, you won't need as thick of a straw (larger transistor) in order to empty the cup. When that cup gets empty is when the destination transistor switches "on".
Bah, fine, the picosecond it's nudged
Fmax corner started being at 0C for really small (45nm) geometries. Secondary effects of short channel length that are affected by temperature start overtaking first-order effects like scattering.
spacemonkey
Newcomer
I still don't get it - is the transistor's switching speed affected by temperature?
Max. stable speed - yes
If for instance your CPU is running stable @3800MHz and 50C it might not want to do so over 60C.
This is also why people are utilizing LN2 or even LHe cooling to beat frequency records. Lower temperature = higher max. frequency.
Of course this is process dependant and there are limits of how cold or hot given process can run.
Good example would be AMD original Phenom manufactured using AMD's 65nm process which hated low temps and getting it to work under 0C was achievement, going to -50C was about max. on very few samples.
On another hand Phenom II 45nm die tolerates cold very well and can operate correctly even frozen to -230C (LHe cooling).
Same goes the other way around. Intel chips using 45nm process on average clocks better in room temp. than AMD. The temperature/frequency curve is different for those two 45nm processes. Part of it comes from using bulk silicon (Intel) or SOI (AMD) process, but I'm not knowledgeable enough to explain details.
[EDIT] Just to clarify - transistor will switch slower in higher temp. This is because capacitance of transistor junction. Switching (charging) that junction will take longer with lesser electron current and the current flow depends on temperature. So higher temp results in longer time to switch the transistor.
I'm sorry but my English vocabulary is too poor to explain it better, but at least I've tried.
Last edited by a moderator:
Similar threads
