This article states the following power consumption:
> The 10900K has a 125-watt TDP, for example, while AMD's Ryzen 9 3900X's is just 105-watts.
My understanding from other articles (like [1]) is however that Intel had to _massively_ increase the amount of power the CPU consumes when turboing under load:
> Not only that, despite the 125 W TDP listed on the box, Intel states that the turbo power recommendation is 250 W – the motherboard manufacturers we’ve spoken to have prepared for 320-350 W from their own testing, in order to maintain that top turbo for as long as possible.
Somehow it feels like back in the Pentium 4 days again.
Intel's TDP is just at base clock which Intel keeps at 3.7GHz. Pretty it's close to a hoax.
Power increases linearly with frequency and squarely with voltage. Higher freq. needs higher voltage. This applies to pretty much any CPU/GPU. Also AVX loads are significantly more power hungry.
230W would be a low estimate for a fully clocked one.
It would be interesting if someone would take an Intel CPU and an AMD CPU and calculate the electricity bill difference someone would pay in a year considering 40 hours / week run time.
Power increases linearly with frequency and squarely with voltage.
Going by the common use of "squarely," the power usage might also be interpreted as being linearly with voltage. Rather, I understood you meant that power increases proportional to the voltage squared.
(For the lurkers: Friction increases proportional to velocity squared, as does kinetic energy. These are all quite useful things to know, and should be part of the takeaways from your basic science education.)
True that, but what bothers me is that people may take it 'as 4x times increase' (quad=4, similar to twice, thrice, and so on). This is pretty much what I used to use.
That's not right - it comes from the formula of power dissipation in an asic: P=CfV^2 (power equals capactiance times frequency times voltage squared).
Clearly we don't turn all the mosfets on every cycle but it's the rough idea.
The real draw under turbo is likely much higher. My i9-9900k for instance has a rated TDP of 95W but it can only boost to about 4.5Ghz all-core when limited to that TDP.
I run it at a 5.2ghz all-core overlock and it pulls nearly 200W under load. (yes, beefy watercooling). From memory it was around 150W @ 5ghz but I haven't played around with the OC settings for a while after I got it tuned right.
But... the overclocking is the official spec. It happens dynamically all the time, it’s not a knob you throw. If TDP is too facile, another more reflective metric may be needed, but for now it’s clear TDP for Intel is marketing not grounded in reality — and not directly comparable between products in their lineup let alone across brands.
The complaint was that when being restricted to TDP power draw processor won't run above about 4.5GHz. Having overclocking mentioned after that as drawing more power made me think, that by overclocking they meant going beyond Turbo Boost, as 4.5GHz is approximately the all-core Turbo Boost frequency for the processor.
If that is correct, then there is no surprise, that going beyond max boost requires drawing power beyond declared TDP.
What sort of water cooling setup do you have? From all that I've read, the typical AIO water cooling setup is only barely better than a high-end air/fan cooling setup.
>>...barely better than a high-end air/fan cooling setup.
A large air cooler will be as effective, possibly more effective, than an AIO water cooler. What matters is radiator surface air x airflow. AOIs tend to have small surface area radiators and quiet fans because that is their target market. Water cooling does have some advanteages, but for coninuous heavy loads all that really matters is radiator capacity. The large air-cooled radiator is going to have more veins with more surface area covered by a more powerful fan than the typical AOI radiator.
100% this, the ultimate performance of a cooling solution is how fast you can move heat into the working fluid and how fast the radiator can move it out of the working fluid.
AIOs confer no particularly special advantage in either. The fact that the working fluid is water means nothing, heatpipes actually use water internally too. The difference is that heatpipes are actually evaporating the water so they can actually more more heat energy (due to enthalpy of evaporation) than AIOs.
AIOs actually tend to be louder for a given amount of cooling due to cheaper, louder fans and crappy noisy pumps. Custom loops don't really suffer that but they also cost 10x as much. They mitigate a lot of the downsides like pump noise by throwing expensive, high quality components at it. That's fine as far as it goes, but AIOs themselves are not an automatic win.
A large air cooler (eg Noctua NH-D15) is going to perform very similarly to a 240/280mm with identical fans (Noctua) at the same RPMs/noise levels. The difference is that most AIOs ship with very cheap fans that are just designed to spin fast and be loud, but that cools quite effectively. You can punch up the fans on a big air cooler and push a lot more heat through them. They just are optimized for silent performance out of the box.
Furthermore, most situations are currently limited by the IHS. AMD once pushed 500W through a 120mm radiator on their 295x2 card (OC'd), and temps would stay under 60C while doing it. The radiator size is not as big an impact as most people expect, moving heat out of the loop is not the bottleneck, it's how fast you can move heat into the loop that matters. GPUs do that very efficiently (bare die, dual chip on the 295x2). 5 GHz 14nm chips push an incredible amount of wattage, 7nm chips are so tiny that they result in very high thermal density, making both of them fairly prone to IHS thermal limitations. Colloquially: the heat just can't move out of the die into the IHS fast enough. The surface of the IHS is actually fairly cool, the liquid inside an AIO is barely above room temperature, but the heat just is not moving into the loop very well.
There is a resulting thing where people splurge on high-end gear, say a 280mm, and see high temps and tell themselves "wow, good thing I bought a 280mm, a 120mm would have been way hotter!". And no, not really, the amount of radiator fins isn't the limitation, it's how fast you can move heat into the loop. If you put a probe inside the reservoir to actually see the temperature of the fluid, it would barely be different between a 120mm and a 280mm.
And in fact the temperatures would probably be barely different than if you bought a NH-D15, or a Scythe Fuma, or other large multi-tower cooler. At the end of the day the working fluid doesn't matter, it's all fins and airflow.
I always thought the value of an AIO was in packaging. Sure, I can buy a huge case that has vast amounts of open space and configurable venting. Or I can jam everything into a smaller case, install an AIO, and ensure I’ve got reasonable airflow for the heat exchanger.
Yeah that's generally the main benefit. AIOs let you build in smaller cases without too much hassle. (Although when you go down to really tiny small form factor cases often you have to use super low profile air coolers again because there's no room for the AIO radiator)
> the ultimate performance of a cooling solution is how fast you can move heat into the working fluid and how fast the radiator can move it out of the working fluid.
Really, it’s all to do with how fast the heat can be permanently ejected from the system, almost always as heated air.
Analogised to a storage array, metal heat sinks and water reservoirs are a high performance write cache. Brilliant for bursty loads but once the cache is filled they don’t improve sustained performance.
There was a Linus Tech Tips video where they water cooled via just running water from a tap through the system and then dumping it back in the sink - no radiator or fans involved. The performance was of course fantastic, but the amount of water you use running a tap constantly is pretty astronomical so it's not actually practical.
The other problem is the mineral fallout in the water. A month or so later they had destroyed their waterblocks.
What you really want is a closed-loop heat exchanger, where the components only touch distilled water, and it gets pumped to a bathtub somewhere with a heat exchanger that pushes the heat out.
> how fast you can move heat into the working fluid and how fast the radiator can move it out of the working fluid.
Nitpick: "working fluid" here implies that cooling has to involve a liquid. But there are solid-state (e.g. peltier) CPU coolers, which of course have no "working fluid." They suck, but they exist!
(I think the suckiness of the existing peltier CPU coolers might be due to their small size, though. If you gave one of them as much solid-state thermal mass to work with as a water-cooling setup has fluid thermal mass, it might be rather efficient, for the same reason a chest freezer is efficient.)
In principle, yeah, but peltiers (a) suck at moving non-trivial amounts of heat, and (b) add a huge amount of heat of their own. Like, ballpark 10x the amount of heat you're moving.
You could move like 5W and you might have to dissipate 50W of heat total. Peltiers can't realistically move 100W and you don't want to cool 1000W.
Helpfully, Linus Tech Tips has walked through this particular experiment for you already ;)
> AOIs tend to have small surface area radiators and quiet fans because that is their target market.
What target market? I know there are some that just have one square 120mm radiator, but I guess I naively assumed that given the fact that they offer 1x, 2x, and 3x length radiators that the average size of a water cooler was somewhere in the middle. I've always understood that the basic advantage of a water cooler is that it lets you move the air-heat interface to a remote object instead of requiring it to be suspended horizontally from the motherboard. Given you get to shove it on the top or front of your case, why wouldn't people choose as big a radiator as possible? What do you think the average water cooler size is?
My guess is that the average AIO water cooler is something like a Corsair H60. The places where you're going to really leverage a water cooler are small and cramped cases where you otherwise can't get good airflow off of the CPU, and those tend to be smaller cases.
(The H60 isn't that much cooler than a Noctua UH-12, either.)
I'd actually be really interested to see statistics from Corsair, NZXT, Cooler Master, or another big AIO OEM about what they sell the most of.
My perspective from participating in PC building communities and consuming content from the big hardware sites and channels is that most people who use an AIO go for either a 240mm or 280mm radiator. I can't remember the last time I saw a 120mm or 140mm radiator recommended for a non small form factor build, as they usually perform equal to or worse than much cheaper air coolers.
If you press your hand to the bottom of a heat sink and warm it up, then turn the water pump on, it feels like water just washed over your hand, the metal cools down as soon as the water passes over it.
Some people seem to think air can be better because at idle, heat pipes can end up a few degrees cooler. Ultimately what you want though is heat dissapation, ideally without super powerful fans and water cooling is far better.
If you move the heat away from a small area quickly, you can easily spread the heat out into a MUCH bigger radiator which will cool more efficiently with air. A nice water cooling loop effectively increases the radiating surface area, with a nice side effect that it moves the heat further away from the rest of the computer.
It doesn't ignore that at all, the whole point is to move the heat from the tiny CPU to a big radiator with a lot of surface area. Air is a terrible conductor of heat compared to water, which pales in comparison to metal, but even metal doesn't move heat as well as flowing water.
They are comparable on cooling capability except for noise levels when the CPU is overclocked. I have a big air cooler and it gets really noisy when CPU is loaded.
I found an AIO (Corsair H115i 280mm rad) wasn't significantly quieter than a large air cooler.
On the other hand the custom loop I now have which has a 420 (triple 140) and a 280 (double 140) rad and includes the CPU and GPU is whisper quiet unless you're running a load that taxes both the CPU and GPU. It's actually eerie when you have it running because I was always used to that fan hum with previous builds.
Noise is a thing. I went back to the classic open loop water cooler on my threadrippers. The RX480 (130 x 56 x 518.5mm) radiator is complete overkill for keeping a 1950x cool - only have 2 of the 4 lower db 120mm fans even plugged in. I'm to the point where I wonder if I could go a single fan or even just a bare radiator with the CPU paired with it now. Only 30C in the loop @180W TDP.
(pump and radiator, longer term, will get tied to a 280w 3970x at some point this summer. Was not expecting the socket update they added with the sTR4x and 39xx series threadrippers)
You'll definitely need it for the 3970X. I run one on a Noctua NH-U14 (running in push-pull with one of their 2000RPM industrial fans and the included 1500RPM fan), and it can get toasty under sustained load. The Noctua typically tests similar to a 240-280 AIO in terms of heat dissipation capacity. It's a longer term goal to do a custom loop with 420mm+ of radiator, but I have other projects before that.
I can maintain turbo under load, but it decays from a full 4.5GHz to 3.9GHz-4.1GHz under a sustained prime95 load. My primary workloads are more bursty, so this is not a big problem for me. I just don't like leaving performance on the table. Also, I cannot turn on PBO on air. It runs up to 95C and then bounces around from 90C-95C, and I'm just not comfortable bumping up against the thermal limit like that long term.
A downside to watercooling in general is that it has a higher thermal capacity relative to a fan and heatsink. This means that a watercooled CPU will stay hotter for longer after it is no longer under load, relative to a heatsink setup. IMO an AIO is a notch down in cooling performance relative to a large heatsink. I still went AIO for noise reasons though.
This is irrelevant entirely, as long as it stays in the desired/designed range.
More importantly the higher thermal capacity allows for easier short turbo boosts which most of the load is. For sustained load, it matters only the hottest part (cores) of the cpu and water cooling is a lot better to transfer heat away.
Yes, it’s not much different than a bowl of water being heated from one side and cooled by a heat sink with a fan on it on the other. It takes time to move the temperature of that water around with some fixed heat source/sink.
That's not quite the right explanation. With the pump in the loop, the heat transfer speed is actually quite high, and ideally the temperature difference between "after CPU" and "after radiator" is small. Without the pump (ie. "bowl"), the temp difference would be quite big until convection kicks in.
However, cooling on any radiator is a function of the temperature difference between air and metal. The heat from the CPU will slowly heat up the water and radiator, which has actually a large heat capacity. That takes times, so the temperatures go up slower. When the heater stops, it will take the radiator also comparatively long to get that heat out of the water again.
A normal air cooler has much smaller heat capacity, so will react much faster, in both directions. These faster cycles are certainly less good for the CPU than the slower cycles of a water loop.
Gamers Nexus is an excellent Youtube channel and they have heavily invested in testing equipment to meaningfully evaluate coolers - some of their initial findings are already interesting. If you really are interested in cooling solutions I’d give their channel a visit.
The closest thing I have found to filling the void when Kyle had the gall to shut down HardOCP :/
There is a lot of variety... Most 360 AIO will do much better than any air cooler with less noise. Many 240-280 will as well... Good/Great air coolers will often do as well as a 120-240mm aio.
This will vary by radiator size/thickness, the fans and the case airflow and any obstructions.
Personally, O11 dynamic, with a 360 aio at the top/exhaust and 3x side, 3x bottom intake... It's really quite pretty much all the time... If I stick my head close, I can hear a little hum, but that my well be the video card for all I can tell. The 4-bay nas on the shelf nearby is more noticeable.
> Somehow it feels like back in the Pentium 4 days again.
What we've got here is actually the mysterious, long-overdued Pentium 5 (which was known for its greater-than 5 GHz clock speed and never released due to the huge heat dissipation issue). History repeats, but this time, it actually got released.
This is the cycle with Intel, though. They innovate, blow out the market, and then stagnate until AMD or someone else comes out with an architecture that's a category improvement on whatever Intel's offering.
This looks like they're turning every nob to 11 on the current architecture to try to get something that they can sell until they can come up with whatever will succeed the Core line.
It looks almost identical to the AthlonXP/64 vs Pentium 4 days when Intel just kept bumping the clocks and power consumption. Until they came out with Conroe.
Jim Keller might be doing just that with Intel right now. I only hope whatever they come up with, AMD is ready and able to keep up. Otherwise we're going to see a rerun of the following period too: Intel tweaking the same architecture for years at a time, and squeezing low single digit performance improvements that come with questionable security sacrifices and deeply unpleasant sticker prices.
While Conroe was definitely solid, last time around in getting there Intel didn't just pursue the technical side of bumping clocks/power, they also engaged in some genuinely illegal anticompetitive practices to get OEMs to not use AMD. So AMD wasn't able to leverage their few years of advantage into marketshare that would enable more sustained R&D, thus when Intel did regain the lead with a new architecture they had lost essentially nothing in the way of strategic position. The $1.25 billion AMD ultimately got out of Intel in a settlement years later didn't erase that advantage.
This time around though Intel isn't in the same position at all to dictate to major market players for a variety of reasons. The current big lucky break catapulting AMD way into the lead is certainly temporary, but it does seem like they'll be able to gain some real marketshare and attention. That will hopefully serve to make this time around end up in a much more stable back and forth.
I remember seeing AMD is growing R&D by 18% next year... what that means in practice, who knows. I am hoping to see them more competitive on the high end of graphics though... > $1100 for a top end consumer video card is crazy.
My understanding is that Nvidia can't produce enough GPU to saturate their pro market, the demand is huge as we speak and only increasing for now (looking at you, ML).
So AMD (or rather, their RTG division specifically) would be remiss not to align their price/perf with Nvidia to maximize profit.
A reason why they'll probably stay 'cheaper' to some extent for some time (perhaps long) is that they don't compete against CUDA, Nvidia's silver bullet. And don't just think “muh fancy library”, think hordes of engineers that Nvidia has the financial muscle to deploy in the field, 'lending' them to customers who need ad hoc driver work, custom implementations, general help, and what-have-you.
CUDA is a war machine, because that's how Nvidia operates its strategy.
It's how they owned the video game market back then, and I'm pretty sure they do the same with CUDA / ML / compute (why wouldn't they).
I thought that OpenCL would eventually be to CUDA what Vulkan is becoming to DirectX (btw: thank you, Linux); but it seems to me that OpenCL is going the way of the dodo? (still very little support in all ML/compute software I encounter, anecdotally).
AMD can sell a ton of GPU at good price, and they should, but they're lagging behind because optimized software libraries support (CUDA, the whole commercial hammer it represents) is where it's at.
Fortunately or not for them, Intel's graphics efforts seem to be... hype, more than a truly organized effort (insiders leaks speak of a terrible, terrible political / manegerial / organization situation at Intel's). It's fortunate for AMD's graphics market share, but it's unfortunate in that Intel would be a major support for OpenCL (last I checked, at least) and Intel's rise on that market could have taken AMD's hardware along with it in a tsunami of competition for CUDA.
Eeeeh, still seeing signs of this in the NUC and Laptop space. For instance you can find a laptop with anything faster than a RTX 2060 despite 4800/4900HS completely thrashing the Intel line of laptop chips.
It would be incredible if they could actually top what AMD has now and is planning in the future. I don't think there will be that same amount of room available to be able to blow away Zen 4 or whatever AMD will have by then.
Jim Keller has said he is technically the manager for thousands of people now and that they are working on even more IPC with even bigger out of order buffers. Even so I don't see the same leap being made in power and single threaded performance.
Then again apple has an incredible amount of performance per watt in their cpus, so I don't know where the limits really are.
AMD's not really any better when it comes to fudging specs. At least Intel CPUs actually limit their sustained power draw to the TDP when running in their factory default configuration, even if enthusiast motherboards usually override it.
AMD CPUs OTOH run significantly higher power than their spec'd TDP right out-of-the box. For example, that "105 W" 3900X actually ships with a power limit of 142 W and it is quite capable of sustaining that under an all-core load.
AMD's turbo numbers are also pure fantasy. If AMD rated turbo frequencies the same way Intel did that 3900X would have a max turbo of 4.2 or 4.3 GHz instead of 4.6 GHz. When Intel says their CPU can hit a given turbo frequency they mean it will actually hit that and stay there so long as power/temperature remains within limits. Meanwhile AMD CPUs may-or-may not ever hit their max turbo frequency, and if they do it's only for a fraction of a second while off full load.
The outrage over Intel CPU's power consumption is pretty silly when you realize that the only reason AMD CPUs don't draw just as much is because their chips would explode if you tried to pump that much power into them. If you care about power consumption just set your power limits as desired and Intel CPUs will deliver perfectly reasonable power efficiency and performance.
>>AMD CPUs OTOH run significantly higher power than their spec'd TDP right out-of-the box. For example, that "105 W" 3900X actually ships with a power limit of 142 W and it is quite capable of sustaining that under an all-core load.
I'm pretty sure you've got that exactly reversed, AMD's power draw tends to stay closer to the rated TDP than Intel's.
>>The outrage over Intel CPU's power consumption is pretty silly when you realize that the only reason AMD CPUs don't draw just as much is because their chips would explode if you tried to pump that much power into them. If you care about power consumption just set your power limits as desired and Intel CPUs will deliver perfectly reasonable power efficiency and performance.
I'm not even sure what this means. AMD CPUs are on a more advanced and physically smaller process, with smaller wires, and a different voltage-frequency-response curve. Of course they would "explode" if you try to pump them with voltages that the process isn't designed to operate with, that the Intel CPU with it's larger process can handle. But the power draw isn't really the "point" of the CPU -- performance is.
Imagine thinking that the Pentium 4 or Bulldozer was better than their contemporaries because they were capable of drawing incredible amounts of power.
> I'm pretty sure you've got that exactly reversed, AMD's power draw tends to stay closer to the rated TDP than Intel's.
I was talking about the CPUs in their factory default configuration. Intel CPUs default to limiting their sustained power draw to exactly their TDP, AMD CPUs meanwhile run significantly higher.
Where Intel CPUs will draw more is when you disable the power limits, but that power buys you extra frequency over what AMD is capable of.
> That's like complaining that speaker wire is deficient because you can't use it to charge an electric car. It's a CPU, not a space heater -- how much power it can draw isn't the point, performance is.
Yes, and Intel CPUs achieve significantly higher frequency, and thus per-thread performance, in exchange for that higher power.
With Intel you have the option of running high power/high frequency or lower power/high efficiency. With AMD you don't get the former option due to limitations of the fabrication process used. Somehow this has come to be considered a win for AMD.
> Intel CPUs default to limiting their sustained power draw to exactly their TDP
No, no they don't. That's a motherboard setting not a CPU one, and basically no desktop motherboard at least follows Intel's "recommendation" out of the box.
Yes, they do. The power limits are set in MSRs within the CPU and the CPU has default values for those MSRs.
> basically no desktop motherboard at least follows Intel's "recommendation" out of the box.
Wrong again. Motherboards using chipsets other than the Z series typically do not change the power limits unless the user explicitly sets them. These days even Z series boards do not change the power limits unless the user enables at least one overclocking feature, e.g. XMP. This is how the motherboard manufacturers technically follow Intel's specs while actually running out-of-spec with the settings most enthusiast users use.
I'm pretty certain that I remember watching videos from GN, LTT, or Bitwit (can't recall which) where they noted that AMD chips were turboing up to a few hundred MHz above the specced numbers.
The most recent generation had turbo numbers that were about 100MHz higher than they should be, and it was difficult or impossible to reach the stated number on most chips, especially on launch or near-launch BIOS versions.
Most came close, but you're right... Of course, Intel's max boost is the same, and unlikely in most consumer systems.
The AMD bios's were updated and do get better boosts now, but really in short bursts. That said, they do incredibly well for most circumstances and depending on your use, a much better option at all price points, except for some gaming at the ~$500 mark.
I'd still take an R9 3900X over an i9 *900K series. Using a 3950X now, no plans to upgrade for 3-5 years.
Those youtubers are measuring "turboing" by looking at the maximum frequency column in HWiNFO. So they consider a CPU to have hit a frequency even if it only touched it for a single millisecond during a minutes long benchmark run. For AMD CPUs the average frequency is typically significantly lower.
Maybe smaller channels, but I find that most of the bigger ones (especially GamersNexus) put a ton of effort into having a good, repeatable process for in-depth looks at performance. Certainly on par with Anandtech or any of the other typically trusted review sites.
Yes, that is the nature of turbo. If the CPUs could sustain operating at those speeds, it would be the base speed, not the turbo. Turbo is for really short burst. After that the system ends up being restrained by the available cooling and will go at lower speed (but usually still higher than base).
There's a difference between "might hit the max turbo frequency for a millisecond under light load if the phase of the moon is just right" and "can sustain max turbo indefinitely so long as adequate power and cooling are provided". AMD CPUs are the former, Intel the latter.
> When Intel says their CPU can hit a given turbo frequency they mean it will actually hit that and stay there so long as power/temperature remains within limits.
Have you actually tried to do that? When Intel says their CPU can hit a Turbo frequency, it's usually pure fantasy because they can't actually do it within the power/temp constraints.
Yes, they'll get to 5-whatever GHz on maybe half the cores if you're lucky, but full all-core load at the top stated Turbo? No chip will stay there, it will throttle down, likely by a lot.
And AVX loads will bring it to, or close to, maximum non-turbo clocks, that's how hot they run.
It could maybe be done with extremely good cooling and undervolting, and completely removing power limits (in which case it will probably consume close to 500W if it doesn't fry something first :D).
> The 10900K has a 125-watt TDP, for example, while AMD's Ryzen 9 3900X's is just 105-watts.
My understanding from other articles (like [1]) is however that Intel had to _massively_ increase the amount of power the CPU consumes when turboing under load:
> Not only that, despite the 125 W TDP listed on the box, Intel states that the turbo power recommendation is 250 W – the motherboard manufacturers we’ve spoken to have prepared for 320-350 W from their own testing, in order to maintain that top turbo for as long as possible.
Somehow it feels like back in the Pentium 4 days again.
[1]: https://www.anandtech.com/show/15758/intels-10th-gen-comet-l...