Hacker News new | past | comments | ask | show | jobs | submit login
AMD's Upcoming Ryzen 9 3900 Listed with 12 Zen 2 Cores at 65W (tomshardware.com)
109 points by rbanffy 20 days ago | hide | past | web | favorite | 55 comments

According to [1] it appears that AMD rates CPU TDP for maximum load (overestimating), while Intel rates them for load without turbo boost (underestimating), so AMD's 65W is really impressive.

[1]: https://www.youtube.com/watch?v=6u4ew6IT4Vo

You almost got it correct.

AMD rates it as non-turbo usage, ex: my $200 3600 is 65W TDP, has a package limit of 88W, and is designed to reach 65W without turboing; if I manually set the package limit to 65W, all cores will sit at the base clock of 3.6GHz forever. 88W hard package limit will be enforced under overclocking unless overridden by user.

The comparison is the $250 9600k, same single-threaded performance, 3600 is roughly 30% faster in multi-threaded; 95W TDP, soft package limit of 150W, and will exceed that under some situations under stock, and will absolutely exceed it overclocked.

So, Intel, in this case, costs 25% more, has 30% less performance, has 46% more TDP, and 70% more actual maximum usage... and to add insult to injury, AMD's stock cooler is beefier than Intel's: I believe AMD's stock cooler can sink 88W, I do not believe the Intel one can manage 150W+.

Intel doesn't even ship coolers for higher TDP chips. I have a i7 7700 (65W TDP) and the stock cooler and it's just too loud under high loads.

AMD's included coolers are one of my favorite little "bonus" things about buying their chips.

As you said, Intel's stock coolers are known for being loud, aren't actually too great at dissipating heat, and aren't even included on higher TDP chips. So you're almost automatically spending an extra $20+ over the chip price on a decent cooler.

On the other hand, AMD ships solid coolers with (pretty much?) every chip. I've run my R5 1600 with the stock cooler for two ish years without any real heat or noise issues.

Amazing, thanks for the details - now the only interesting number missing in the comparison is the profit per CPU...

In the case of AMD, the TDP figure just means it's shipped with a cooler that has capacity for 65W. In reality it just means boost clocks will drop sooner.

In the case of Intel, the claimed TDP is a fairy tale.

Are people still impressed by low TDP numbers? Everyone can manipulate that number to anything you want by changing the base clock. Intel could sell the 9900K with a 25 W TDP by saying the base clock is 1200 MHz, AMD could do the same. (I mean the 95 W TDP for the 9900K is already a meme.) It's just gonna be exactly the same CPU as the 3900X, just a slightly worse binned version.

I think there is some merit to these chips having better power consumption. Generally the CCX's that draw more current under no cpu load are capable of higher clocks albeit at higher power draws at said clock.[i] That is to say is that the higher performance chips from the silicon lottery are the least efficent.

So as you said, its "just a slightly worse binned version" but worse in terms of max clock not performance/watt.

Although there isnt a 'hard' definition of TDP AMD has typically rated there chips more accurately[ii] than intel in the past [0], but that said AMD's TDP is still lower than absolute max power draw.

[i]- Memories from working in validation, although this was not focus of my work, this is what I recall on the topic.

[ii]- from amd's data sheet "The maximum power a processor draws for a thermally significant period while running commercially useful software. The constraining conditions for TDP are specified in the notes in the thermal and power tables."

NOTE! This source is old!! [0]-https://www.anandtech.com/show/2807/2

>That is to say is that the higher performance chips from the silicon lottery are the least efficent.

That is wrong generally. Not sure how you got that idea either. The chips that overclock the highest are typically the most efficient ones, too. A 9900K that can run at 5.3 GHz at 1.35v is going to use less power than a 9900K that needs 1.5v to run at 5.3 GHz. At any frequency. Because the "golden" chip simply leaks less power, and needs less voltage at any given frequency. Lower TDP parts aren't special in any way, they just have lower clocks. If you limit the clocks on higher TDP parts, you get the same (or better) power consumption.

The thing is they are the least efficent at idle. If you are seeing higher baseline current you can interpret this as the chip sort of having a lower internal resistance which is a double edge sword.

If that difference exists, I'd be very surprised if it was even measurable with normal methods. Since on modern x86 CPUs most of the chip is literally just dead / not doing anything at idle. But feel free to surprise me in case someone actually tried that.

I want to start by agreeing with you. It is true what you said about overclocking, chips that run cooler will be able to clock higher, and it is true that chips will disable sections of or even entire cores if they are un-used. So I don't know if you can measure this outside of validation setting. I think I have explained myself poorly so I will try again from a more technical perspective.

To drive the transistors in a chip you need to supply or sink enough charge so that the MOSFET will change state. The rate at which we can supply this charge places a fundamental limit on the switching speed of this transistor. Because until we overcome this charge requirement the transistor will not turn on/off.

Now the speed we can move this charge in and out of the input of the chip is governed by resistance. Higher resistance between VCC and the gate means it takes longer to over come the capacitance and change the state of the MOSFET.

Lets say for some given arbitrary Transistor its input capacitance is some 50pF , and lets also say our resistance to vcc is 10 ohm, then tao or the time constant is nothing but RC or 500 pico-seconds. In other words the maximum switching speed that we can see on the output is 1/tao or some 2Ghz. Now logic circuits are more than one transistor deep, and we can't reasonably run the circuit at exactly tao (Also at t=tao the gate is not exactly vcc it is ~.63vcc as this is an exponential function) and expect any level of signal integrity. But this tao is sort of a hard limit to switching speed as it is the time it takes the transistor to swap state.

In some sense low resistance is a boon. Low resistance means low tao, as tao is just RC. If tao is small than the transistor changes state more quickly. As a result for a given clock, the transistor has longer to overcome any inductance/capacitance present and is able to emit a cleaner signal.

But what lower resistance means is that you have higher power consumption as: p=iv & v=ir -> P=V^2/R So we see low resistance-> higher current. and higher current -> more power.

So we see it is a balancing act. We can have higher speeds or we can have lower power but not both. So if we are binning for a certain speed we are also binning for a certain power consumption characteristic. By selecting chips with speeds greater than XGhz we are also selecting chips with internal resistance < R.

Now what you said about over clocking is true. But because of an added twist.

Say we are binning a set of R that satisfy clk > XGhz. So naturally some subset of R will still be comfortably capable of XGhz while using less power than the average member of R. Now the chip has real limits to its capability to perform while hot. So this subset performs well as it is just low enough resistance to run a XGhz while limiting power use. On the high edge of the set R, we see chips that barely made XGhz and cant clock better. On the low edge of set R we see chips that are runing so hot as to barely maintain XGhz. So there is a sweet spot.

> But what lower resistance means is that you have higher power consumption as: p=iv & v=ir -> P=V^2/R So we see low resistance-> higher current. and higher current -> more power.

I think you're forgetting time in this equation. If two otherwise identical MOSFETs have different gate resistances, the lower-resistance MOSFET will charge faster and draw switching current for less time.

Take this all the way back to a simple linear circuit: a 1-ohm, 1-farad RC circuit will have a time constant of 1s, and if supplied with 1V the current will be I(t) = exp(-t). The power lost to heat in the resistor is then I^2R = exp(-2t), and integrated from 0 to infinity the net energy loss is 0.5J.

If instead the circuit has a resistance of 0.5 ohm, the resulting current is I(t) = 2exp(-2t). Power loss is I^2R = (0.5)*(4 exp(-4t)) = 2 exp(-4t), and integrated from 0 to infinity gives a net energy loss of again 0.5J.

In the meantime, the lower-resistance MOSFET is likely to save power from other effects. Most notably, faster switching implies the transistor spends less time in its linear region where it itself acts as a resistor between source and drain.

You bring up an excellent point, I had neglected to show the effect of time in my explanation as you said. And it is true that this has real impact as to the magnitude of what I was talking about. But so long as the switching time != infinity, the power consumed is not equal. Using your example,

1 Ohm Case : P_res = integral exp(-2t) from 0 to tao -> .4323J 1/2 Ohm Case : P_res = integral 2*exp(-4t) from 0 to tao -> .4908J

Of course the effect diminishes greatly (to zero as you showed) as the switching time increases. But we are dealing with a great number of transistors switching very many times so the effect is still noticeable given timings are tight enough.

The thought on power savings as you mention is interesting but i am unsure off the top of my head of it.

No, the power consumed is equal. The capacitance doesn't stop charging at time tao as you imply, it continues to charge until it reaches the final driving voltage, which is why integrating to infinity is a (very) much closer approximation

What I am meaning is the input to the transistor on a chip is highly volatile, so the aproxmation of its power consumption using infinite or long periods of time of a single value is misleading.

If a transistor is only reciving 5v than it is fair to say the power consumed is equal.

But realisticly the input could change every clock cycle and we would see the cases vary.

The reason I only integrate to tao was to show if the input switches before t=infinity then a diffrence between cases clearly manifest. But as you say, when it does not switch for time t>>tao the diffrence is small.

I don't doubt that this works in theory. I just don't think this effect is relevant in practice (with the same manufacturing process). Especially because with lower resistance you can usually decrease the voltage too.

If the performance is alright for my needs, why wouldn't I prefer a processor with a lower TDP? I agree that it would be nice to have consumption numbers for some standardized workload, but in the absence of that, TDP seems useful to me.

> consumption numbers for some standardized workload

It would be nice if any of the synthetic benchmarks included "10 hours of power-on, with 7 hours of web browsing, 2 hours of CAD, and a bunch of just-plain-idle time". Of course the choice of OS and mobo would matter a lot here too, but...

...I think it's entirely valid that idle power is probably a large part of power consumption and it's wholly ignored by current benchmark techniques. TDP is great for sizing heatsinks, but maybe I want to estimate my power bill.

SilentPCReview used to do a lot of this work, since heat means fans means noise. They'd have a "recommended system" every few months, which was the current performance-per-watt champion combo of CPU+mobo. I'm sitting in front of my last build from their 2008-era recommendation right now. But they're mothballed and I don't think anyone else has taken up that particular torch to run with it.

>If the performance is alright for my needs, why wouldn't I prefer a processor with a lower TDP?

Because 1. a processor with a lower TDP doesn't necessarily actually use less power, at least not if your cooler can dissipate much more than the TDP, which it typically can.

And 2. because you can simply limit the clock speeds of any higher TDP CPU to reach any TDP you want. If you want a 35 W 3900X or 9900K, just set the clocks to whatever you need to reach 35 W max power usage, and you're done.

Of course the lower TDP parts are typically cheaper, so it makes more sense to buy those. But that's their differentiating factor, the price, not the TDP.

TDP is very useful if you want a quiet PC without water cooling. Under clocking may be required based on case airflow, but you want someone to be making passive CPU coolers for your socket.

Not sure what you're getting at, this is the same socket as all the other Ryzen 3000 CPUs...

Not to mention that every slightly larger air cooler is a "passive cooler". Just don't connect the fan. Or set a fan curve that disables the fan as long as the CPU temperature is sub-62 C. Or 95 C, if you really don't want them to turn on.

Yes, in this specific case it not a big deal, and they make some surprisingly effective passive coolers today. However, I have run into this issue in the past.

As to running a larger air cooler without a fan. That’s heavily dependent on case airflow. High speed case fans really defeat the purpose of a passive cooler.

You don't really need any case airflow at all, it just depends on how much performance you're willing to sacrifice. Case in point, you can run cinebench on an 8700K without any cooler at all: https://www.youtube.com/watch?v=yA0oo12rbiM

For something a bit more practical, if you're willing to get creative you can passively cool CPUs while getting fairly high performance too: https://www.youtube.com/watch?v=N-z9PidYH4E

Those are just NH-D15s, completely standard coolers with mounting brackets for pretty much every consumer socket out there.

That second video uses fans see 5:40. Lots of closely spaced thin fins actually produce less cooling without good airflow. There are real engineering reasons passive coolers have significant air gaps between the plates.

PS: In terms of sacrificing performance, let’s agree to avoid the absurd. Or as the guidelines put it: Please respond to the strongest plausible interpretation of what someone says, not a weaker one that's easier to criticize. Assume good faith.

It uses fans, however IIRC the fan curves are such that they only kick in if you render videos or do something else that 100% uses the CPU, and even then they are fairly quiet. And again, if you simply limit the clock speeds and voltages, you could disable the fans completely. (Or you could just disable the fans completely in the first place and let the automatic regulation take over.)

>Lots of closely spaced thin fins actually produce less cooling without good airflow.

That is true, however Noctua coolers tend to have a fairly big gap between fins, because they are designed for very low RPM use.

>In terms of sacrificing performance, let’s agree to avoid the absurd.

Not sure why you think it's absurd - the video just shows that there is no real minimum amount of cooling required anymore, at least with Intel CPUs. Obviously you're going to have to sacrifice some performance if you go passive only.

Although I'm not even sure anymore what we're even talking about. All consumer CPUs use the same sockets. All these coolers are available for every modern consumer or "pro-sumer" platform.

> All consumer CPUs use the same sockets.

Motherboard’s use a small set of different mounting brackets for CPU coolers. But the physical CPU socket depends on several things including the physical size of the chip and thus the amount of surface area you want in contact with the the cooler.

In case you where unaware actually running a CPU at 100C will drastically lower it’s lifespan. You encounter similar issue if there are significant temperature differences across the chip. Which is why packaging includes a metal plate over the CPU even though it reduces cooling. However, this is a real tradeoff which means the contact area must be reasonably close to design spec.

Yes, bracket, not socket. What are we talking about again?

Selecting and then using passive CPU coolers. I am saying both the bracket and socket are important when choosing a cooling.

For example, when introduced there was no aftermarket passive cooling available for the AM4 socket.

Really of the 5 considerations “Does it fit the mounting bracket?” is probably the least important. Fitting the motherboard and case are mandatory. Fitting the socket and TDP have a little wiggle room. However, with mounting brackets you can generally get something to work as long as you keep firm contact and it does not wiggle around it’s fine.

Aren't you just testing the processor's internal overheating protection in that case? It might not shut down the system, but throttle itself into a snails pace.

In my years long experience of looking into and building quiet PC's, going for a nice spacious case with a high-end semi passive power supply and putting in some smooth low rpm air-cooling in a nice open airflow path beats trying to go completely fan-less or god-forbid a water-cooled solution.

TDP would be useful if we standardized how it is measured, and at a whole systems level on standardized representative workloads.

Just build in a Lian-Li o11 dynamic, Corsair HX psu. I honestly don't hear it over other ambient noises around (AC, road outside, etc)... I am using an AIO 360 for top-exhaust, but it's a far cry from a custom open-loop, not much harder than a regular HSF. Been very happy with it so far.

(rambling ahead)

Using an r5-3600 as a placeholder waiting for a 3950X. Was running an i7-4790k since that dropped, and wasn't unhappy with it, but there are things where it gets sluggish with the 4 cores, etc. I have definitely noticed the performance difference. Also jumped to Linux (Pop!_OS) from 2 years of hackintosh. I need to update to kernel 5.3 and update video drivers before swapping the older rx570, with an rx5700xt aftermarket.

Actually being able to order certain parts has been a bit of a pain though.

I would suggest undervolting if you want it quieter

Most people won't undervolt, so the heatsink manufacturers won't pander to those who do. You're also limited to the manufacturer's bounds for undervolting before you're in unstable, warranty-breaking territory.

So you're saying TDP is a measure of the efficiency of a CPU? That is if 2 CPUs are listed at the same TDP and one has a higher clock speed, then the latter would be more efficient.

Intel measure the TDP based on the base clock, yes. However AMD seem to measure it based on what the CPU is likely to draw with boost clocks over a sustained period of time, so their TDP numbers are quite accurate.

AMD used to do that until Zen 2. With new generation AMD joined intel in lying to consumers :(

The 3900X still uses ~140 W at full load with a 105 W TDP. And still just like Intel, if you limit the clock speeds you can make it use 105 W, 65 W, or 35 W at full load, just like with Intel chips. AMD calculates the TDP a bit differently, but nothing really changes fundamentally. All CPUs in a generation are essentially the same chip, TDP is completely irrelevant.

140W is considerably closer to the stated 105W TDP value of the 3900X than a "95W" 9900k which can easily pull down 180W. Whilst I agree that TDP shouldn't be taken as gospel, at least in AMDs case, it can be a useful approximation the amount of heat generated (and thus power draw).

If you then compare the actual performance per watt, a 12 core CPU at 3.1 base / 4.2 boost pulling down approximately 65W is very impressive. Something similar from Intel (in terms of number of cores at even just the 3.1 base frequency) would likely have a TDP of at least 120W.

>140W is considerably closer

Closer, yes, but more useful? Both are off by a mile - not to mention that TDP isn't supposed to tell you a chip's maximum power consumption in the first place. It's supposed to tell you how good your cooler has to be to sustain base clocks / single core turbo, or in AMD's case, to sustain base clocks and some? boost clocks for a limited time.

>If you then compare the actual performance per watt

Performance per watt can certainly be interesting for many people. Performance per watt just isn't connected to TDP. The 3900X (105 W TDP) has higher performance per watt than the 9900K (95 W TDP). Probably at any frequency - although potentially the 3900X can't clock down as much. The 9900K can run Cinebench without a cooler after all (at least the 8700K can, so I'd suspect the 9900K can do that too), while the AMD processors can't. Also, the 3900 (65 W TDP) has the same performance per watt as the 3900X (105 W TDP), if you let them run at the same clock speeds.

Ryzen SKUs with a 105W TDP have a power limit (PPT) of 142W at stock. 65W TDP SKUs have a power limit of 88W. I.e., while these CPUs are allowed to use more power than their TDP, it's limited to 35% above the TDP. This means that the advertised TDP is not just an arbitrary number they slap on the package.

It's not an arbitrary number on Intel's CPUs either, it's just still not very useful. Or at least it's not useful as an indicator of some sort of technological progress. If you limit a 3900X to 3900 clock speeds and voltages, it's going to behave exactly the same.

Can you tell me the formula which will approximately give me the max TDP used by a recent Intel CPU based on its rated TDP?

For users who have a fixed TDP budget, there's a substantial difference.

The AMD power calculation is not "bit different", it's Intel who is cheating big time.

On a system with constrained power values, those who buy Intel will suffer a substantially higher hit than those who buy AMD.

> For users who have a fixed TDP budget, there's a substantial difference.

I really, really don't see how. Both TDPs have very little connection to how much power the chips actually pulls at full load. 105 to 140 isn't useful in any way. If it was 107, sure, but 140 is 1/3rd higher than "spec" (I write "spec" because TDP isn't supposed to tell you what the highest power draw possible is if your cooling can dissipate more than the TDP in the first place, people just somehow seem to use it that way).

If you have a fixed TDP cooler, all you have to do is put that cooler on CPUs and see which ones will be faster. No looking at the specs required. If you have a fixed power consumption budget, but a better cooler, all you have to do is downclock until the CPU stays within that budget and then compare performance. Again, spec wasn't useful.

Not the same chip, quality differs, some units would require higher voltage for the same frequency. Some CPUs are more power efficient because they can take lover voltage.

Yeah and those chips are typically the ones being sold as higher TDP chips.

I'm impressed that it's only slightly higher than 1/3rd of the 2920X's TDP for the same number of similar cores. That doesn't feel like a normal jump whether it's massaged by marketing or not.

Why can't we just get a range?

I care about TDP so I can estimate total power draw for my system when designing a custom build. I need to know about how big the PSU and UPS need to be, and getting a rough idea of electricity costs may influence my choices of components.

I can't trust the TDP figures and instead need to look up actual tests to know what it really draws, which is unnecessarily complicated.

Power consumption changes drastically depending on how many volts the motherboard feeds into the CPU. Different motherboards use different default values and different algorithms. And on top of that users can go in and tweak OC settings.

The only honest TDP range they can give you would be: TDP between 10 and 300W.

I have a 9700K (95W TDP). At idle, the entire system consumes 30W at the wall. When running Prime95, the entire system consumes 250W, and that's undervolted. The 95W number is utterly useless. If I had bought a 95W capable cooler, the system would throttle under load.

Intel will for sure look like a hawk on the numbers that AMD claims.

I don't think they will gain any advantage with lying.

No, the power consumed is equal. The capacitance doesn't stop charging at time tao as you imply, it continues to charge until it reaches the final driving voltage, which is why integrating to infinity is a (very) much closer approximation

The real question, will there be enough to meet demand... The only 3900X availability currently, are price gouging.

Still eagerly waiting for the r9-3950X though.

If memory serves, things usually drop towards MSRP within a month or two for AMD. Hopefully that'll happen with these.

Now imagine something like this with an integrated Navi GPU. That would be killer.

Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | Legal | Apply to YC | Contact