TR Forums

Does anyone know of any sites that are doing VRM temperature testing with X299 boards? Or are any sites monitoring the temps on the EPS12V connector cables?

I still remember the days when Prescott would murder 4-pin AUX motherboards if they were cheapies or the board was overclocked too far. For anyone that hadn't see it, Tech Report's news blurb on Der8auer's overeating VRMs and EPS12V cable melting concerns

I was shopping X299 boards and I've now seen boards using three different power configs. A single EPS12V, an EPS12V+AUX, and even dual EPS12V connectors. if I buy a board it will use dual EPS12V connectors, but the range of options is indicative of the range of processors these boards are expected to take. 12, 14, 16, and 18 core models haven't even launched yet and we know people are going to clock those as high as they can get them stable.

Fixing the VRM cooling is straightforward enough, either remove the chunk of metal insulating them or get a better board design. But der8auer did also mention the temps on the EPS12V cables got hot, and from past experience I can attest it doesn't take much to slag even an EPS12V connector. I expect this will become an issue since 7900X overclocking all but requires watercooling, which means direct airflow may be nonexistent. And that's not getting into 12, 14, 16, and 18 core chips that haven't launched yet that are going to go into these motherboards.

Edit: It appears Threadripper's X399 platform shares the same potential problem. I've now found photos showing all three power plug configurations, same as the X299 platform.

Haven't seen any in depth power testing on other tech sites yet, mostly general reviews of the X299 boards so far. I also noticed the different power connections between the board makers and wondering how these will compare. ASUS, ASRock and Gigabyte displayed previews of their X399's at Computex or other sites but did not see any from MSI. I'm looking to build another HEDT later in the year but do not necessarily need a high core count just more PCIe lanes. So an X399 setup may be the better value if price and power draw are more reasonable then the Intel platform even if a bit lower in performance.

Seems that while Sky-X chips may be stable at >4.5Ghz overclocks, a combination of CPU/VRM throttling may actually be killing performance. Need to use HWMonitor & benchmarks to actually see it.

This is getting very technical (at least for me), but Der8auer has posted a follow up video addressing many points and clarifying his results. https://youtu.be/89Tgazt8v5Y For example, he's verified this behavior with SuperFlower, Seasonic, and Corsair PSUs. It also is not a board specific issue.

More importantly, he has spent several days working with Tom Logan from OC3D who has now also been able to reproduce the issue. https://youtu.be/3h2p5383ubw Prime95 is required, OCCT doesn't cut it. If the 10-core chip can draw 400w loads from the VRMs with Prime95, I shudder to think what the 12, 14, 16, and 18 core chips will draw with other programs. Or Threadripper for that matter.

It is reassuring to know the VRMs can self-throttle from high temps on their own, but I think it's also disturbing that chunks of finless metal are supposed to somehow dissipate the heat from 300-400w VRM loads. Average users won't hit 400w unless they disable all the safety protections in the BIOS and change the CPU to a 140% current cap, but again we know kids are going to do this for that "magic" 5Ghz number. It also still leaves me concerned about the long term viability of these motherboards for moderate overclocks on sustained 24/7 loads without watercooling the VRMs.

Exotic subzero cooling overlocks will be something to see, once the CPU throttling is removed as the limiting factor who knows how much the VRMs will be putting out, 500w will be just the start given they are rated for well over 400w. I'll have to start watching OC events again just to enjoy the fireworks.

TL/DR:

Direct fan cooling is required for VRMs, case fans behind or directly above the board won't do anything. And a single EPS12V cable is going to get very very warm feeding 300-400w loads to the VRMs, regardless of X299 or X399 platforms... I bet the 12-18 cores chips may start slagging boards that only have a single EPS12V connector.

Kougar wrote:

It is reassuring to know the VRMs can self-throttle from high temps on their own, but I think it's also disturbing that chunks of finless metal are supposed to somehow dissipate the heat from 300-400w VRM loads.

They're 90%+ efficient, so we're only talking about 30-40 watts of heat to dissipate (assuming all that 100% of the power goes straight to the CPU and no other components in the system...) and they'll happily run at 100C for years without issue.

They aren't CPUs we're talking about here - power circuitry is much hardier and designed for more extreme environments. Granted, idiotic motherboard design leaves them with quite...inefficient...heatsinks, but they're still far better than running bare assuming the TIM isn't pure junk.

Waco wrote:

They're 90%+ efficient, so we're only talking about 30-40 watts of heat to dissipate (assuming all that 100% of the power goes straight to the CPU and no other components in the system...) and they'll happily run at 100C for years without issue.

Are they really, though? ASUS was playing coy and wouldn't tell Tom what the VRM was specc'd to handle nor its safe thermal limit. All they would say was 400w was well within the design spec, but the 105c temps he was getting was in fact a problem.

My understanding is that efficiency and current ratings vary depending on operating temperature, so they can't really give you a simple number to fit all cases. I think efficiency is usually very high though, and there's a relatively small amount of waste heat for the amount of current that's being provided. In general a VRM will be able to provide more current at lower temperatures, but more current means higher heat dissipation, and that requires better cooling. Exactly how much cooling you'll need depends on the configuration of the VRM. A cheap 3 or 4 phase power setup will need better cooling than a 12 or 16 phase setup, because each phase of the VRM is responsible for a greater portion of the total power draw, but you can still run into trouble with tiny heatsinks especially when overclocking.

Edit: And there's probably a bunch of stuff I'm leaving out, because I'm not really a hardware guy.

Kougar wrote:

Waco wrote:

They're 90%+ efficient, so we're only talking about 30-40 watts of heat to dissipate (assuming all that 100% of the power goes straight to the CPU and no other components in the system...) and they'll happily run at 100C for years without issue.

Are they really, though? ASUS was playing coy and wouldn't tell Tom what the VRM was specc'd to handle nor its safe thermal limit. All they would say was 400w was well within the design spec, but the 105c temps he was getting was in fact a problem.

They are rated that high, yes.

Thermals are a problem, obviously, because style matters far more than function on modern boards. For short bursts, before the thermal limits are reached, the 16+ phase designs on modern boards could nearly power a damn arc welder.

105 C won't damage any driver FET. 110 C is up there, but is again unlikely to cause damage. These boards throttle early because they typically have a handful of sensors and assume the worst. If you install real heatsinks on them that don't look like guns, dragons, or boobies, or whatever and provide a bit of airflow...they'll run for years at 100% load even with ludicrous overclocks.

Case in point is server motherboards. They rarely have stupid power configurations that can feed many hundreds of watts per socket. They rarely support any kind of overclocking...yet with only 3-4 phases they power 100+ watt CPUs at 24/7 100% load with a handful of tiny heatsinks and decent airflow.

That[s good to know the VRMs are up to that, so we aren't worried about the solder running hot either? The single EPS12V is still a problem. The VRMs are still going to be drawing 300-400w mostly through the EPS12V connector. I know there's a reason >300W GPUs have 8+8 pin arrangements, yet the 7900X can pull 400w through the VRMs and boards are feeding most of that through a single 8-pin connector? Nobody else here sees a problem with this?

I should correct myself that what I thought were single EPS12V Threadripper boards are in fact 8+4pin boards, ASRock hid the AUX connector on the opposite side of the socket/memory slots on their designs. So I haven't found any single 8-pin X399 board yet.

An 8-pin EPS connector has 4 hot pins, vs a GPU 8-pins 3 hot. All things being equal in a perfect world, using recent pins, not 8 year old terminals, a standard mini-fit jr pin and good 18AWG wire can provide 9A*12V, so 108Watts. That's 436 Watts per 8-pin EPS. The 300W seems to be based on older 7A terminals from quite a few years back, 7*12*4 is 336Watts. There are even 11A terminals.

Would you want to do that 24/7? Nope. Besides, there are other interconnected parts involved too.

I thought the original youtube video was pretty good. Ive always hated the heatsinks on PCH and VRMs. Most are for show and some chips that also could benefit from it dont get any.

What I most wonder about this whole debacle is who decided that many-core chips like these need to run all-core clocks this high. If you've got a lightly threaded workload that isn't running as fast as you want, you don't have many choices other than turning up the clocks. SKL-X turbo seems good enough that that's no excuse for stupid high all-core clocks. If you've got a workload that can actually use 10 cores, and it isn't running fast enough, it would probably be much better served by going to 12C (~1.2x power cost) than by boosting clocks 20% (>1.4x power cost). Obviously Intel wants to keep their profit margins by not selling anyone more silicon than they have to, but optimizing for perf/mm² at the expense of perf/W isn't how you end up commanding the kind of prices Intel likes to command.

The logic that does work in Intel's favor when they do this is that of having a winning halo product, but they've already got 10C to AMD's 8C in the current lineups, and they've got 18C on the way to out-halo AMD's 16C. They've got good enough turbo that they surely can't be just trying to maintain the single-threaded performance crown at the same time (as again, they've already got a solid win). If AMD's going to beat them on technical grounds, it's at perf/W, not halo stuff. When they do this, they're just handing AMD a more solid perf/W lead for free.

This doesn't excuse choking on thermals at <70% of the power density that KBL-S can handle, and I'd think vendors of top-end mobos would want to make sure they're never limiting the CPU regardless of insane overclocks, but seemingly none of this would be the kind of in-your-face issue it now is if Intel hadn't lead with "what if we tried more power?"

blahsaysblah wrote:

An 8-pin EPS connector has 4 hot pins, vs a GPU 8-pins 3 hot.

High-end GPUs that draw this kind of power have 6 hot pins, so you're missing half of them. It also means GPUs have more hot pins than the EPS12V connector, and let us not forget that the GPU draws up to 75w directly from the PCIe slot in addition to those six PCIe power pins.

Again, I have a serious problem with this much current being drawn from just 4 hot EPS12V pins, especially as my understanding is the EPS12V connectors also happens to power the quad-channel memory banks. Maybe I'm wrong on that detail, it's been awhile since I've seen really in depth mainboard reviews that touched on this.

Something I mentioned on the TR news article is that the 7920X is a 12-core HCC chip. Almost half of the die area is dark silicon, so combined with a much larger die area and probably real solder the HCC chips should have more headroom before thermal throttling kicks in, which means more power draw than even what the 7900X can do.

synthtel2 wrote:

What I most wonder about this whole debacle is who decided that many-core chips like these need to run all-core clocks this high. If you've got a lightly threaded workload that isn't running as fast as you want, you don't have many choices other than turning up the clocks. SKL-X turbo seems good enough that that's no excuse for stupid high all-core clocks. If you've got a workload that can actually use 10 cores, and it isn't running fast enough, it would probably be much better served by going to 12C (~1.2x power cost) than by boosting clocks 20% (>1.4x power cost). Obviously Intel wants to keep their profit margins by not selling anyone more silicon than they have to, but optimizing for perf/mm² at the expense of perf/W isn't how you end up commanding the kind of prices Intel likes to command.

I'm going to go out on a limb and guess that you've never had a one/two core application jump to a "slow"/base-clocked core. On Windows 10 it happens quite frequently. This is particularly fun when it's combined with core parking/power-saving modes.

The best way to combat this as a power user is to simply buy a CPU with less cores. More realistically, the best path to performance consistency is to force all cores to a given MHz, and damn the (power use) consequences.

(Yes, I'm aware of affinity-setting. I'm also aware of the reasons that you don't want to go down that rabbit hole.)

That's not the only edge case that can hurt it by a long shot. Nonetheless, there are fewer cases all the time where an S part would be suitable and an LCC part wouldn't due to performance, and that would be true even if SKL-X's all-core turbos were a few bins lower.

Vhalidictes wrote:

I'm going to go out on a limb and guess that you've never had a one/two core application jump to a "slow"/base-clocked core. On Windows 10 it happens quite frequently. This is particularly fun when it's combined with core parking/power-saving modes.

Sky-X chips with Speed Shift are supposed to finally solve that. Two cores get favored status and clock to 4Ghz for precisely that scenario. Now that Intel and Microsoft are baking the driver into Win 10 the only reason it won't work is if venders have it disabled by default in the BIOS. Apparently quite a few still are.

Kougar wrote:

blahsaysblah wrote:

An 8-pin EPS connector has 4 hot pins, vs a GPU 8-pins 3 hot.

High-end GPUs that draw this kind of power have 6 hot pins, so you're missing half of them. It also means GPUs have more hot pins than the EPS12V connector, and let us not forget that the GPU draws up to 75w directly from the PCIe slot in addition to those six PCIe power pins.

Again, I have a serious problem with this much current being drawn from just 4 hot EPS12V pins, especially as my understanding is the EPS12V connectors also happens to power the quad-channel memory banks. Maybe I'm wrong on that detail, it's been awhile since I've seen really in depth mainboard reviews that touched on this.

Something I mentioned on the TR news article is that the 7920X is a 12-core HCC chip. Almost half of the die area is dark silicon, so combined with a much larger die area and probably real solder the HCC chips should have more headroom before thermal throttling kicks in, which means more power draw than even what the 7900X can do.

I totally agree that some light on how the stuff actually works, which power were and how much, would be pretty cool.
I thought memory is off the 3V and that's why its separate VRM is always near it and ATX connector. I am like 95% certain because a faulty Skylake BIOS/RAM almost fried my PSU and it was the little rails part that started smokingsmelling of burning. I personally had no clue it was overloading the 3.3V rails. It should have never been able to touchoverload the 12V without clearly evident destruction somewhere.

The wiring between the PSU and either the EPS12V or the 6/8 pin GPU plugs is 18-gauge. The US National Electric Code limits ampacity in 18-gauge cords to 7 amps per if 3 or more conductors, or 10 amps if but 2.

To stay in spec, a 4-hot/8-pin EPS12V can legally carry 28 amps at 12 volts, or 300 watts per EPS12V. A 6-hot/8-pin GPU can carry 42 amps, or 504 watts per 8-pin GPU. The cords from the PSU are NOT the problem. They're well within spec.

The only source of 3.3v and 5V to a modern motherboard comes from the ATX 24-pin (absent older mobos with a Molex plug up against a slot).

If the mobo mfgs are skimping on the sockets, then it's all on them. The PSU is NOT the problem.

Captain Ned wrote:

The wiring between the PSU and either the EPS12V or the 6/8 pin GPU plugs is 18-gauge. The US National Electric Code limits ampacity in 18-gauge cords to 7 amps per if 3 or more conductors, or 10 amps if but 2.

To stay in spec, a 4-hot/8-pin EPS12V can legally carry 28 amps at 12 volts, or 300 watts per EPS12V. A 6-hot/8-pin GPU can carry 42 amps, or 504 watts per 8-pin GPU. The cords from the PSU are NOT the problem. They're well within spec.

The only source of 3.3v and 5V to a modern motherboard comes from the ATX 24-pin (absent older mobos with a Molex plug up against a slot).

If the mobo mfgs are skimping on the sockets, then it's all on them. The PSU is NOT the problem.

I dont think that standard was designed to protect the inside of computers. When they say cords, they are talking about actual bundled wires like the standard Romex 14/2 wire in your walls for your 15 Amp circuits. The NEC is to ensure people/infrastructure safety.
That chart didnt even say what type of wire, material, stranded and was showing numbers for STOW sheating of all things.

blahsaysblah wrote:

That chart didnt even say what type of wire, material, stranded and was showing numbers for STOW sheating of all things.

In my experience PSU wiring is THHN, which can handle those ratings. Unsheathed cables would do even better since they're not in a thick ROMEX cladding and thus more likely to shed heat.

The shortcoming is on the mobo side, not the PSU side. IOW, the mobo is asking for current the PSU can safely provide, but the mobo connectors aren't specced for what they're asking for. After all, the PSU is only going to pump out the volts/amps requested/accepted by the destination. I don't have a dog in this hunt since I'm still running an i7-4790K, but everything I've read on this bears the hallmarks of a product that should have spent another 6 months in the incubator but was birthed early for competitive reasons. One would think that some would remember the P4 experience.

Oh i absolutely agree. Good(verified by jonnygru,..) PSUs right now are easy to come by and they are outstanding at their job. Hence the 7-10 year warranties.

edit2: what i mean, is that it seems like PSU have hit some kind of physics limit for a few years now. They're making excellent PSUs. The goal post needs to move, but im happy with Gold's 90%. Titanium moving that to start at 10% load, where do you go from there? For general PC. They seem to have hit some walls when almost everyone is making good platinum/titanium PSUs at decent prices. Basically, a little afraid of the 7-10 year warranties. Stagnation?

edit: i thought i read it here, that why mini-ITX Ryzen is hard to come by,.. is Ryzen shipped really late, and QA/design was overloaded for that brand new platform and than there is Threadripper/EPYC somewhere in background and add in Intels new platform and chips, which was also moved up. 2017 has multiple salvoes on all fronts after steady monopoly years of Intel.

Captain Ned wrote:

To stay in spec, a 4-hot/8-pin EPS12V can legally carry 28 amps at 12 volts, or 300 watts per EPS12V.

Der8auer has posted a new, short vid on the redesigned ASUS Apex VRM cooling. He was able to attain 340 watts without the VRMs throttling making it the first board to do so. The cooler actually has fins, they added a backplate, and even a VRM fan. https://youtu.be/hejMP5hrhjc

I still say if a 7900X can do 340 on CLC cooling then the 7920x with almost 45% of the chip area being dark silicon +2 extra cores should do one better. I'm looking forward to JonnyGuru testing EPS12V cable temps when he gets back from his vacation. Even in an 8+4 configuration Threadripper is going to be interesting as well.

Kougar wrote:

Captain Ned wrote:

To stay in spec, a 4-hot/8-pin EPS12V can legally carry 28 amps at 12 volts, or 300 watts per EPS12V.

Der8auer has posted a new, short vid on the redesigned ASUS Apex VRM cooling. He was able to attain 340 watts without the VRMs throttling making it the first board to do so. The cooler actually has fins, they added a backplate, and even a VRM fan. https://youtu.be/hejMP5hrhjc

I still say if a 7900X can do 340 on CLC cooling then the 7920x with almost 45% of the chip area being dark silicon +2 extra cores should do one better. I'm looking forward to JonnyGuru testing EPS12V cable temps when he gets back from his vacation. Even in an 8+4 configuration Threadripper is going to be interesting as well.

That's interesting to see. From der8auer's perspective it looks like this Asus motherboard has solved the motherboard-side issues he was complaining about earlier.

It's also amusing how he was complaining about how his 7900X was limiting his testing since it was "only" running at 4.9GHz under Prime95 torture test conditions.
He has different problems than the rest of us.

Would've vastly preferred if they could've gotten decent cooling for the VRMs without the use of an additional fan. It's just one more thing to fail.

just brew it! wrote:

Would've vastly preferred if they could've gotten decent cooling for the VRMs without the use of an additional fan. It's just one more thing to fail.

The VRMs were in-spec at 340 watts of consumption* even without the fan, it's just that with the fan the temperature dropped noticeably.
Frankly you could probably get similar results with good internal case airflow.

* And despite some of the irrational hatred flying around about Skylake X since apparently having a 10-core speed of 4.9GHz is now a sign of "failure", no 340 watts is not "normal" for even a heavily overclocked part.

Whelp, JonnyGURU tested a 450w load and only saw a peak of 65c on the EPS12V cable.

just brew it! wrote:

Would've vastly preferred if they could've gotten decent cooling for the VRMs without the use of an additional fan. It's just one more thing to fail.

The only difference between a chunk of metal and a heatsink is that one has fins, and I'm just happy to actually see something that resembles a heatsink for a change.

The fan was optional, so you can just angle a silent 120mm fan at the socket/VRM area and use that instead for cooling. I've been doing that anyway on every desktop I've built since I started watercooling.

Just FYI, latest is 2017, but i cant find online link for that, i was reading this is 2013 version: PSU Design Guide for Desktop Platform Form Factors
If you go to page 27:

4.2.2.4 +12V Power Connector
Connector: Molex* 0039012040 or equivalent.
Contact: Molex 44476 - 1112 (HCS) or equivalent (Mating motherboard connector is Molex 39-29-9042 or equivalent).

NOTES:
1. 18 AWG wire.

If you lookup the Molex 44476-1112 (HCS, high current series), it lists 11 amps.

blahsaysblah wrote:

If you lookup the Molex 44476-1112 (HCS, high current series), it lists 11 amps.

If you look up the Ampacity Tables, 18 AWG in a bundle like a braided PSU core is rated for 7 amps / 84 watts per conductor. If you're old-school like me with the sloppy cables, it's 10 amps per conductor. It doesn't take a rocket scientist to think that the end-stream connector might be able to take an amp or two over wire spec given which monkey had carnal relations with the PSU, hence the 11 amps. Here's the deal, that's 11 amps PER PIN, not across the entire connector.

I should make you lot work in my dirt-floor breaker panel where I live. Talk about a Sorting Hat. Have fun with the 240V/30A double breaker circuit.

I know I've said it many times before, but I have not the slightest qualm about working in there live. Mainly because my headlamp buggered off and the cobwebs are of mystical proportions.

I bought 18AWG AWM style 1007 VW-1 wire and 9A pins because that was so way over what each hot wire would use in PCI-E and EFS ports. 50W/wire so maybe 4A.

I swear i saw much kinder ampacity chart for short runs of hookup wire but i looked again and was wrong or at least cant find it anymore.

No i would not monkey around at all with power. Safety first.

LIVE? ARE YOU INSANE? What if you sneeze and while raising your hand you graze some hots while your other hand happens to be touching the breaker panel box?
I really dont like electricity at all. It's too dangerous. Just have to get careless once and youre dead, not you lose a finger or something. You die dead. The dead kind of dead.

Captain Ned wrote:

I should make you lot work in my dirt-floor breaker panel where I live. Talk about a Sorting Hat. Have fun with the 240V/30A double breaker circuit.

I know I've said it many times before, but I have not the slightest qualm about working in there live. Mainly because my headlamp buggered off and the cobwebs are of mystical proportions.

Working on a 240V circuit live is incredibly dangerous without proper training and a second set of eyes/hands from another properly trained person watching you work.