Xeon W-3175X (the "5GHz" 28 core) uses paste rather than solder

[Amazonsucks]

56 minutes ago, Jurrunio said:

uh huh, did you read the title of this post? People buying the 28 core W-3175X wont be caring about 'saving money'. You need to swim in money to consider this CPU in the first place.

Then maybe they did it for thermal cycling issues? Has anyone bothered to get a statement from them about it?

[leadeater]

9 minutes ago, Amazonsucks said:

Then maybe they did it for thermal cycling issues? Has anyone bothered to get a statement from them about it?

If that was a real issue then the enterprise Xeons (not Xeon W) would not be soldered and they are. The argument for not doing it because of small dies and risk of damage is very valid, that's why Xeons were soldered and the now much smaller desktop ones were not. Even when the dies weren't much bigger the larger process nodes were less susceptible to heat damage and less of the functional parts of the die would get damaged anyway.

 

And it's not like thermal paste TIM is the problem anyway, it's the gap between the IHS and the die that is the problem and solder while also being far more thermally conductive also makes sure that gap is extremely small, because they are fused together. A new IHS and mounting design would probably fix that issue almost as well as solder would.

[Mira Yurizaki]

I encourage the armchair thermodynamic scientists here to read https://medium.com/@OpenSeason/soldered-cpu-vs-cheap-paste-59fb96a4fca7

[Arika]

18 minutes ago, M.Yurizaki said:

I encourage the armchair thermodynamic scientists here to read https://medium.com/@OpenSeason/soldered-cpu-vs-cheap-paste-59fb96a4fca7

that was an interesting read, i am by no means a  self proclaimed thermodynamic specialist, i was curious .

 

i never really understood the rabid reaction some people have to companies not using solder and they only ever attribute it to cost, but according to that paper, there is much more to consider  about what to use than just cost

 

with this

Quote

 From here going forward, I doubt any 7nm or small die size solutions will have Indium solder TIM1. At the time of writing this piece, Ryzen 3 has not been launched and the package has not been analyzed. It will be quite interesting to see what unfolds.

it will be interesting to see the forum's reaction if AMD don't use solder (because of the reasons listed in the article) and might prompt them to look into it more

[Amazonsucks]

18 minutes ago, leadeater said:

If that was a real issue then the enterprise Xeons (not Xeon W) would not be soldered and they are. The argument for not doing it because of small dies and risk of damage is very valid, that's why Xeons were soldered and the now much smaller desktop ones were not. Even when the dies weren't much bigger the larger process nodes were less susceptible to heat damage and less of the functional parts of the die would get damaged anyway.

 

And it's not like thermal paste TIM is the problem anyway, it's the gap between the IHS and the die that is the problem and solder while also being far more thermally conductive also makes sure that gap is extremely small, because they are fused together. A new IHS and mounting design would probably fix that issue almost as well as solder would.

I said in an earlier post why datacenter and HPC workloads, which the Xeon Scalable and other soldered CPUs are used in, typically dont involve that kind of thermal cycling.

[Jurrunio]

so in the future CPUs that need thermal repaste under the lid will become a thing?

[leadeater]

1 minute ago, Amazonsucks said:

I said in an earlier post why datacenter and HPC workloads, which the Xeon Scalable and other soldered CPUs are used in, typically dont involve that kind of thermal cycling.

They do, a lot more than a desktop would. Thermal cycling is an issue with rapid changes that happen repeatedly so a computer going from power on to off over night is not a problem. A server on the other hand operating 24/7 with dynamic loads on it that often run around 30%-50% average with continuous load spikes on various different cores as demand and activity changes.

 

A server running 30 VMs will have more thermal cycling than a desktop CPU ever will.

 

Cracking just is not an issue, destroying the die trying to solder it actually is.

[Mira Yurizaki]

46 minutes ago, Arika S said:

it will be interesting to see the forum's reaction if AMD don't use solder (because of the reasons listed in the article) and might prompt them to look into it more

It's like the "LOL INTEL TOOTHPASTE" people don't acknowledge this exists:

 

[Amazonsucks]

10 hours ago, leadeater said:

They do, a lot more than a desktop would. Thermal cycling is an issue with rapid changes that happen repeatedly so a computer going from power on to off over night is not a problem. A server on the other hand operating 24/7 with dynamic loads on it that often run around 30%-50% average with continuous load spikes on various different cores as demand and activity changes.

 

A server running 30 VMs will have more thermal cycling than a desktop CPU ever will.

 

Cracking just is not an issue, destroying the die trying to solder it actually is.

Thermal cycling is excursions from cool to hot, which is different but obviously correlated to workload variability. In a datacenter, efficient utilization is good for ROI, because it means servers aren't sitting idle, taking up space. Same thing with HPC installations. What you'll find in a typical high end server, with big soldered chips, is that there will be one of several cooling solutions implemented:

Forced air over large, usually folded fin solid copper heatsinks, which tend to have a large, constant volume of airflow. The combination of high airflow volume and the thermal mass of the heatsink means that the CPU's tend to stay within a tighter temperature band. Since noise isn't an issue with datacenters, the fans can run at maximum speed constantly, and the workload variability doesn't correlate to huge swings in temperature like it would on a desktop, which would be about 20C off/hibernating/sleeping, then rapidly go to 70 or 80C, then back down to 20C over and over. 

With liquid cooling, there are various different methods: chilled water, warm water and hot water cooling. In any of the cases, the temperature of the CPU package itself is not going to be able to do what it does in a desktop, because of the much faster removal of heat when compared to most desktop CPU cooling solutions, with the exception of high end liquid loops with large flow rates(not common in desktops or laptops obviously). In the case of datacenter and HPC liquid cooled installations, the coldplate and CPU tend to stay much closer to the coolant temperature, once again not having massive swings from say 20C to 80C.

In either air or liquid cooled server/HPC hardware, the variation of temperature is not going to look much like a desktop, even with large swings in actual workload, for the simple reason that the cooling medium to CPU heat load ratio is going to be much higher in a datacenter, specifically for the purpose of reliability. In a desktop, PWM fans and pumps are commonplace. In high end servers and HPC, deterministic performance is preferable. That's also why, until very recently and still in some cases, turbo boost and DVFS are disabled because they create nondeterministic performace across a large cluster with thousands of CPUs. In something like a mainframe, its even more extreme, with CPUs operating at 5GHz for decades with very little variability in temperature.

 

At least based on all the research on the subject of HPC and datacenter CPU utilization, cooling design and reliability engineering that I've read...

But, the other factor to consider is that it's not in soldering the CPU that it gets broken. It's the fact that soldering a tiny chip with a CTE mismatch to indium based solder(all the metallic TIM for CPU die to IHS I am familar with uses indium with layers of gold) can cause issues with pulling on the smaller piece of silicon and cracking the solder interface.

[leadeater]

45 minutes ago, Amazonsucks said:

-snip-

FYI I'm a Systems Engineer at one the largest university's in my country and we managed two large datacenters with another 3 smaller ones. We have both large scale many hundreds of enterprise server host (over a thousand VMs hosted on them) for corporate applications and the general function of the business along with multiple research units with their own HPC clusters (in the thousands of cores, small scale stuff internationally though).

 

If you go read the supplied link by @M.Yurizaki about the solder process and when and why you would use paste over solder you'd understand the issue is at the time of soldering the die and IHS not afterwards during operation. The temperatures are not hot enough nor are the variations wide enough to worry about thermal expansion layers of differing metals and substances. Quite simply the large Xeon dies are big enough and the heat load is distributed enough that solder is the best choice thermally for those CPUs and is why Intel solders them, all Skylake-SP Xeons are soldered and that won't change with the next generation, or the next.

 

When people complain about CPUs with paste and bad thermals they incorrectly blame the paste which is not the issue, it's the height/gap between the die and IHS. You don't have to use liquid metal if one delids, just remove the IHS and reapply paste and don't seal it and thermal performance will greatly improve (there are already videos/reviews showing this). As for the reason paste is used, read the link TL;DR die is too small and cores are too few and too close together.

[Jurrunio]

@leadeater then why make CPU with lids?

[leadeater]

2 minutes ago, Jurrunio said:

@leadeater then why make CPU with lids?

So you don't crack the die when mounting the cooler or if a cooler maker gets the pressure wrong, it's there to protect it. Direct die used to be a thing, is now for extreme people .

[The Benjamins]

12 hours ago, Arika S said:

that was an interesting read, i am by no means a  self proclaimed thermodynamic specialist, i was curious .

 

i never really understood the rabid reaction some people have to companies not using solder and they only ever attribute it to cost, but according to that paper, there is much more to consider  about what to use than just cost

 

with this

it will be interesting to see the forum's reaction if AMD don't use solder (because of the reasons listed in the article) and might prompt them to look into it more

 

11 hours ago, M.Yurizaki said:

It's like the "LOL INTEL TOOTHPASTE" people don't acknowledge this exists:

 

People don't because that is not a issue.

AMD has solder on high end CPUs, and uses paste for cost savings on low end CPUs.

Intel uses paste as cost savings on most CPUs, If intel would of had x299 and i7 and or i5 as soldered it would not of been a issue.

[leadeater]

@Amazonsucks

Here's an example ESXi host temperature graph over time if you're interested, not sure what the sample interval is for checking the temps and the tool (HPE OneView) uses F not C which is really annoying for me. Actual real time temperature fluctuations are rapid, CPU load changes a lot, enough to kick up and down CPU power states.

 

 

Edit:

CPU load on an ESXi host, so you can see just how much variation there actually is.

Spoiler

I don't really expect anyone to be able to real that all that well, that's 48 cores of mess right there lol. Load average over that hour is 20%, it is 2am here though so during the day it's about 40%-50% and there are more cores spiking and more often as there are more users during the day than night.

[leadeater]

1 hour ago, Amazonsucks said:

Since noise isn't an issue with datacenters, the fans can run at maximum speed constantly,

Fans are controlled by the system, servers have a hell of a lot of sensors in them and each fan/zone can be controlled. Running at max RPM isn't a thing due to reliability.

 

HPE DL380 Gen9

 

P.S. Sorry for the multiple replies. not really reading that post fully/intently as it's rather long and doesn't actaully say anything I don't already know. Late and I'm being lazy, just picking out parts that stick out to me. Probably going to bed now though.

[Amazonsucks]

34 minutes ago, leadeater said:

Fans are controlled by the system, servers have a hell of a lot of sensors in them and each fan/zone can be controlled. Running at max RPM isn't a thing due to reliability.

 

HPE DL380 Gen9

 

P.S. Sorry for the multiple replies. not really reading that post fully/intently as it's rather long and doesn't actaully say anything I don't already know. Late and I'm being lazy, just picking out parts that stick out to me. Probably going to bed now though.

Credentialsim is cool and all, but perhaps you should read the article that you told me to just read more carefully. It's not during soldering itself that the CTE mismatch is the issue. It's after, as the solution solidifies and cools, that the pulling effect happens. 

 

From the linked article:

 

"However the same sort of stresses are imparted onto the packages as the Indium cools causing further bending of the substrate."

 

Which is pretty much exactly what I said...

 

And:

 

"Due to the the property of the indium based solder, thermal cycling forms microcracks in the solder layer. The crack formation resembles that of a layered separation leaving voids where thermal energy cannot transfer. The crack pattern is referred to as “Delamination.” Over time these cracks contribute to significant increase in overall thermal resistance of the TIM1 layer."

 

Also, pretty much exactly what I said.

 

So, it is both after the actual soldering happens, during the cooling phase, as well as during the operational life of the package as it undergoes thermal cycling.

 

And as I was saying, small thermal variations, like the one you show in your chart aren't the kind I am referring to. As you can see from your own chart, the delta T of the CPU in your example is ~10F which is insignificant in the context of coefficient of thermal expansion mismatch related delamination or crack formation in the materials being discussed. 

 

If it was a delta T of something like 60C, which is what you'd expect from a powered off CPU to full load, then back down, the actual thermal expansion would be much more significant than the small 10F fluctuation in your chart. 

 

While it looks jagged to someone who doesn't know how to read a graph, its really an insignificant variation in the context of the materials being discussed and their respective delamination from one another as a result of thermal cycling. I mean, the chart actually starts at 50 and goes to 100, and it stays between 70 and 80 for the most part but that chart actually illustrates what I was saying about temperature variations, as it stays within a very narrow temperature band.

 

Plenty of systems have also constant speed fans. And if the fan's speed is constant, it's not going to drop below that maximum. I obviously don't mean the absolute maximum speed that the fan is physically capable of before prematurely failing, but rather the maximum speed it's designed to constantly run at since it's a constant speed fan, as opposed to a variable speed fan, but I thought that went without overtly stating.

 

Not sure what the graph of sensor data is supposed to show about constant vs variable speed fans...

 

"not really reading that post fully/intently as it's rather long and doesn't actaully say anything I don't already know." lol ok

[Carclis]

So the Xeons get the $TIM instead of STIM. It all makes sense now.

[BuckGup]

12 hours ago, M.Yurizaki said:

It's like the "LOL INTEL TOOTHPASTE" people don't acknowledge this exists:

 

One costs $100 one costs $5000. There is a difference 

[leadeater]

7 minutes ago, Amazonsucks said:

Credentialsim is cool and all, but perhaps you should read the article that you told me to just read more carefully. It's not during soldering itself that the CTE mismatch is the issue. It's after, as the solution solidifies and cools, that the pulling effect happens. 

 

From the linked article:

 

"However the same sort of stresses are imparted onto the packages as the Indium cools causing further bending of the substrate."

 

Which is pretty much exactly what I said...

And the process you just mentioned is at production when the the die and IHS is soldered, when else do you think the solution is solidifying? There's only one time that ever happens.

 

Quote

While the largest thermal swing occurs at production time, the life cycle of the processor can also contribute to delamination. Thus, a fresh soldered processor BNIB is at its peak thermal performance prior to running hundreds of heating cycles. Larger dies are less prone to stress based delamination at production time.

 

There's a really big difference in theory and actuality here, the theory is more than correct however the impact is very small. There are thousands and thousands of Xeon 5100 series through to 5600 series all running perfectly fine with zero thermal issues, that and I have never heard of any Xeon having this problem in real life.

 

The only reason Intel doesn't solder the desktop CPUs is the die size, as is near as much directly stated in the link, so much as so that they postulate that 7nm Ryzen will be paste not solder because of the die size problem.

[Mira Yurizaki]

1 hour ago, The Benjamins said:

People don't because that is not a issue.

AMD has solder on high end CPUs, and uses paste for cost savings on low end CPUs.

Intel uses paste as cost savings on most CPUs, If intel would of had x299 and i7 and or i5 as soldered it would not of been a issue.

The Ryzen 5 2400G is also pasted, even though the Ryzen 5 2600 is soldered. They're both the same MSRP (although the 2600 is $10 cheaper  on NewEgg at the time of this writing)

 

Also I'd really like to see some evidence that it's cheaper and by how much. Everyone who's all for solder CPUs makes this assertion, but has no actual data to back it up other than some assumption that thermal paste is always cheaper.

[The Benjamins]

4 minutes ago, M.Yurizaki said:

The Ryzen 5 2400G is also pasted, even though the Ryzen 5 2600 is soldered. They're both the same MSRP (although the 2600 is $10 cheaper  on NewEgg at the time of this writing)

What is your point, both the 2200g and 2400g are low end parts, and they are a different die to the other CPUs.

 

The 2400g does NOT compete with the 8700k, 8600k, or x299 CPUs.

[leadeater]

15 minutes ago, Amazonsucks said:

And as I was saying, small thermal variations, like the one you show in your chart aren't the kind I am referring to. As you can see from your own chart, the delta T of the CPU in your example is ~10F which is insignificant in the context of coefficient of thermal expansion mismatch related delamination or crack formation in the materials being discussed. 

That's a temperature average over all the cores in the CPU and average over time, that's not real time taken every second so that tells you nothing about the instantaneous temperature fluctuations. What I'm showing you is there are big changes in temperature, big and often enough that it shows up in highly averaged wide sample data sets.

 

18 minutes ago, Amazonsucks said:

Not sure what the graph of sensor data is supposed to show about constant vs variable speed fans...

That there are 44 active sensors in the server and that inputs data back to the system controller and it uses that to change the fan speed individually of each of the 6 fans as needed. HPE does this, Dell does it, Lenovo does it, everyone does it.

[Mira Yurizaki]

1 minute ago, The Benjamins said:

What is your point, both the 2200g and 2400g are low end parts, and they are a different die to the other CPUs.

 

The 2400g does NOT compete with the 8700k, 8600k, or x299 CPUs.

Still waiting for actual evidence of how much using thermal paste saves vs. using solder.

[The Benjamins]

10 minutes ago, M.Yurizaki said:

Still waiting for actual evidence of how much using thermal paste saves vs. using solder.

I think other savings is that the die is made to be used with TIM or solder, and example the 2200g and 2400g dies are used in embedded parts (laptops and others). they also have less binning options so it would be desirable for less failures due to solder.

 

But from my view is more WHY did Intel not solder these high end enthusiast CPUs, not why they would use paste for some.

 

At the place I work at they try and save pennies ($0.01) off production costs for their consumer products.

[Amazonsucks]

4 minutes ago, leadeater said:

And the process you just mentioned is at production when the the die and IHS is soldered, when else do you think the solution is solidifying? There's only one time that ever happens.

 

 

There's a really big difference in theory and actuality here, the theory is more than correct however the impact is very small. There are thousands and thousands of Xeon 5100 series through to 5600 series all running perfectly fine with zero thermal issues, that and I have never heard of any Xeon having this problem in real life.

 

The only reason Intel doesn't solder the desktop CPUs is the die size, as is near as much directly stated in the link, so much as so that they postulate that 7nm Ryzen will be paste not solder because of the die size problem.

Since you were being technical, the process of solidification and cooling happens after the actual soldering is done.

 

As to the second part, the rather huge Skylake X, which uses paste TIM instead of solder says hello... Why no solder there if its just size???