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

[GoldenLag]

25 minutes ago, Amazonsucks said:

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

 

since you are being nitpicky. isnt that strictly speaking still under the process of soldering it? because if it isnt solid. the job isnt done yet. 

 

you arent finished making icecream before the cream is frozen.

 

27 minutes ago, Amazonsucks said:

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???

idk, probably one of the other dozen allready mentioned in this thead.

[leadeater]

1 minute ago, Amazonsucks said:

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???

I wasn't being technical, literally every time I have said the issue is when you solder the die and IHS and not during operation, no idea how that can be misunderstood.

 

Because the Skylake-X die is smaller than Skylake-SP and the original platform and socket design was for paste IHS TIM, that choice was made because in part due to Skylake-X being a lower margin product so lowering the risk of damaging dies was likely required. I have no idea how many dies get damaged during the solder process but that would effect the price Intel is able to set on the product, $3000-9000 USD Xeons are a lot easier to absorb that cost than $500-2000 USD products.

 

Core	Die Type	Package	Dimensions
Skylake SP	LCC	FCLGA-3647	76.16 mm x 56.6 mm
	HCC
	XCC
Skylake X	LCC	FCLGA-2066	58.5 mm x 51 mm
	HCC

Being that the product this topic is about, a Xeon W, which for the actual customers going to use it won't be overclocking it or doing anything extreme paste is fine. For all the extreme overclockers they can all delid which Intel knows. Btw no other Xeon W is unlocked and overclockable like this one is, the product in my view is mostly Intel marketing and not much else.

 

Since you're probably going to bring it up the 9900k, as for why that can be soldered knowing exactly why can only really be answered by Intel. Customer demand and complaints about paste would be a large factor, the cost per die is lower than Skylake-X so each one lost is lesser impact, more could be damaged though due to the size. These are all things I can't answer, I'm not Intel. More people care about thermals and overclock on a 9900k than a Xeon-W 3175X, but if Ryzen can be soldered the 9900k sure can too. Also I don't believe previous Intel desktop CPUs couldn't either, the need to wasn't there so it would only be an all risk high cost exercise so why should Intel do that.

[Drak3]

1 hour ago, M.Yurizaki said:

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

I think the only big tocket savings wpuld come from how many usable dies come out from being soldered vs TIM'd, as from my understanding, the TIM installation itself AMD and Intel use cannot damage the die, whereas the soldering process does have that chance.

[Mira Yurizaki]

58 minutes ago, The Benjamins said:

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.

Which would only affect the manufacturing process until they get to the point where they're bonding the die to the package. Also the 2200G and 2400G are desktop parts, the laptop variant is the 2200U and 2500U.

 

Quote

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

Which is something only Intel knows. We can argue all we want about possible reasons, but none of them are going to be the reason until someone pulls up something from within Intel.

 

Besides that, the first thing that sticks out in my mind is the Skylake X package is a weird beast. Which means you can't exactly apply what you know already with "normal" packages.

 

Quote

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

Probably in places where you can cut costs without affecting the overall quality of the product. Toyota/Subaru put cheap plastic pieces for the sun visor in the FR-S/BR-Z. Mazda didn't bother putting a trunk lid cover on the inside of the MX-5. None of these affect performance of the vehicle itself very much and it's in places people won't really notice, but going nicer would've cost them more.

 

Using solder vs. thermal grease does affect the quality of  the product and should be subject to a group of engineers to make the final call.

Edited October 18, 2018 by M.Yurizaki

[Amazonsucks]

11 minutes ago, GoldenLag said:

 

since you are being nitpicky. isnt that strictly speaking still under the process of soldering it? because if it isnt solid. the job isnt done yet. 

 

you arent finished making icecream before the cream is frozen.

 

idk, probably one of the other dozen allready mentioned in this thead.

Except that soldering is the opposite of making ice cream. You dont really consider the annealing to be the same as melting constituent components into an alloy either, right? They're distinct steps in a fabrication process. 

 

And i was replying to him saying that it was the only reason. Clearly if there are dozens of reasons, that cant be the only one? 

6 minutes ago, leadeater said:

I wasn't being technical, literally every time I have said the issue is when you solder the die and IHS and not during operation, no idea how that can be misunderstood.

 

Because the Skylake-X die is smaller than Skylake-SP and the original platform and socket design was for paste IHS TIM, that choice was made because in part due to Skylake-X being a lower margin product so lowering the risk of damaging dies was likely required. I have no idea how many dies get damaged during the solder process but that would effect the price Intel is able to set on the product, $3000-9000 USD Xeons are a lot easier to absorb that cost than $500-2000 USD products.

 

Core	Die Type	Package	Dimensions
Skylake SP	LCC	FCLGA-3647	76.16 mm x 56.6 mm
	HCC
	XCC
Skylake X	LCC	FCLGA-2066	58.5 mm x 51 mm
	HCC

Being that the product this topic is about, a Xeon W, which for the actual customers going to use it won't be overclocking it or doing anything extreme paste is fine. For all the extreme overclockers they can all delid which Intel knows. Btw no other Xeon W is unlocked and overclockable like this one is, the product in my view is mostly Intel marketing and not much else.

 

Since you're probably going to bring it up the 9900k, as for why that can be soldered knowing exactly why can only really be answered by Intel. Customer demand and complaints about paste would be a large factor, the cost per die is lower than Skylake-X so each one lost is lesser impact, more could be damaged though due to the size. These are all things I can't answer, I'm not Intel. More people care about thermals and overclock on a 9900k than a Xeon-W 3175X, but if Ryzen can be soldered the 9900k sure can too. Also I don't believe previous Intel desktop CPUs couldn't either, the need to wasn't there so it would only be an all risk high cost exercise so why should Intel do that.

I never brought up Skylake SP though... What is its relevance to Skylake X and CPUs that are actually smaller, which use solder, like small Ryzen CPUs? 

 

Wouldnt AMD, who can afford loss even less than Intel, encounter the yield issue if there was a significant risk of damage during or after manufacture?

 

I didnt misunderstand the part about manufacturing vs operation causing failure in the solder interface.

 

The article that you told me to read, as well as extensive metallurgical research begs to differ with your assessment, however.

 

Micro crack formation and delamination does indeed occur during operation.

 

Here's a nice article from Der8auer about it.

 

https://overclocking.guide/the-truth-about-cpu-soldering/
 

"Micro cracks in solder preforms can damage the CPU permanently after a certain amount of thermal cycles and time. Conventional thermal paste doesn’t perform as good as the solder preform but it should have a longer durability – especially for small size DIE CPUs."

 

The technical whitepapers ive read agree with that assessment as well.

[Mira Yurizaki]

3 minutes ago, Drak3 said:

I think the only big tocket savings wpuld come from how many usable dies come out from being soldered vs TIM'd, as from my understanding, the TIM installation itself AMD and Intel use cannot damage the die, whereas the soldering process does have that chance.

That would be a cost saving measure by way of reducing the amount of defects they have. Which is something you should be doing as a manufacturer. But not like say "I need to pinch pennies any chance I can!" vibe that I keep getting.

[GoldenLag]

3 minutes ago, Amazonsucks said:

Except that soldering is the opposite of making ice cream. You dont really consider the annealing to be the same as melting constituent components into an alloy either, right? They're distinct steps in a fabrication process. 

thank you for confirming my entire point. it is still part of the soldering process.

[Mira Yurizaki]

20 minutes ago, Amazonsucks said:

Here's a nice article from Der8auer about it.

 

https://overclocking.guide/the-truth-about-cpu-soldering/
 

"Micro cracks in solder preforms can damage the CPU permanently after a certain amount of thermal cycles and time. Conventional thermal paste doesn’t perform as good as the solder preform but it should have a longer durability – especially for small size DIE CPUs."

 

The technical whitepapers ive read agree with that assessment as well.

This also leads to the reason why the RROD and YLOD of the 7th generation consoles happened: thermal stress caused fractures in the solder, leading to failure. Granted Microsoft claimed they used the wrong lead-free based solder. And also Intel's market is full of customers who don't upgrade their computers for like 4-5 years. Sometimes even longer. At that point reliability trumps absolute performance.

 

But in the end, none of us, as far as I know, work at a manufacturing plant or the engineering team for Intel, AMD, or whoever. So we're all just making guesses here.

Edited October 18, 2018 by M.Yurizaki

[The Benjamins]

36 minutes ago, M.Yurizaki said:

 Also the 2200G and 2400G are desktop parts, the laptop variant is the 2200U and 2500U.

 

yes, but they use the same dies, same as the Ryzen V1000.

AMD is making 2 Different dies for Epyc 7000, Epyc 3000, Ryzen Desktop, Ryzen APU, Ryzen mobile, Ryzen embedded V1000.

 

1 Die is made to be soldered, and the other is not. A die that is created to be soldered can still be used with paste TIM

[leadeater]

5 hours ago, Amazonsucks said:

I never brought up Skylake SP though... What is its relevance to Skylake X and CPUs that are actually smaller, which use solder, like small Ryzen CPUs? 

You asked originally if the reason why they weren't soldered was for thermal cycling reasons, the answer to that is if there was any merit to such a problem then Skylake-SP would not be soldered. It may not be as big of a potential issue on a die larger than the Skylake-X but the customers buying Skylake-SP care far more about long term reliability, pay top dollar and hate down time or unexpected issue meaning Intel would put their best foot forward on all fronts for these product and wouldn't do something that has a known and likely problem.

 

Size is the key factor for how much of a problem this is, the smaller the die the more likely there will be during the solder process a problem with delamination as well as after during operation. All the factors to solder or not revolve around die size, so the summary issue is that, where everything else is consideration points around it and due to it.

 

5 hours ago, Amazonsucks said:

The article that you told me to read, as well as extensive metallurgical research begs to differ with your assessment, however.

 

Micro crack formation and delamination does indeed occur during operation.

Can form cracks, not will form cracks, that might impact cooling in extremely rare cases and get more likely as die size gets smaller.

 

We're talking about metals that have melting points around 160C+ and operating temperatures between 30C-80C typically and you really think the difference in CTE between these different substances is large enough and there is actually enough heat to cause not just significant expansion but also enough difference between the substance to cause cracking that will have actual impact on the cooling capability of the CPU, not just theory but actually.

 

How many failed i7-2700K due to solder cracking exist? These fits your narrative on desktops being worse cases than servers, even though a server is 24/7/365 for 5 years plus with package temps that go between 30C to 60C (actual figures I took since you were so insistent there wasn't large variation).

 

Server CPU coolers aren't actually that good, worse than good gaming air coolers, and use high airflow to compensate however the only difference is the ambient temperature between a server and a gaming desktop. If the delta change on CPU temps is 30C then ambient won't change that.

 

Cracking of solder is just another fear of something that just doesn't happen, you have enough products in the desktop and server side that have been running long enough in enough differing conditions for this to actually show up, if you can show it is actually a wide issue in reality then I might change my view on it. But while Intel solders $9000 USD processors my view on it is solder does not have a reliability concern.

 

TL;DR Jump out of white papers and internet web pages and actually find this issue in reality, bet you won't find it. Real cases of CPU thermal performance significantly degrading due to solder cracking. The only popular one I know of being the Xbox one was caused by faulty solder itself, using a different solder composition and process, and was not a thermal cooling issue at all (of the CPU/GPU) which is not an indicator of a wider issue of soldering the die to the IHS.

[DildorTheDecent]

18 minutes ago, leadeater said:

Real cases of CPU thermal performance significantly degrading due to solder cracking. The only popular one I know of being the Xbox one was caused by faulty solder itself which is not an indicator of a wider issue of soldering the die to the IHS.

Some extreme OC guys have reported solder cracking. Mostly whenever they are reheating the pot after a cold bug IIRC.

 

Of course the key word here is "Extreme OC". I'm sure that no normal user with a capable cooling solution will ever encounter solder cracking.

[Jurrunio]

5 hours ago, Amazonsucks said:

Here's a nice article from Der8auer about it.

 

https://overclocking.guide/the-truth-about-cpu-soldering/
 

"Micro cracks in solder preforms can damage the CPU permanently after a certain amount of thermal cycles and time. Conventional thermal paste doesn’t perform as good as the solder preform but it should have a longer durability – especially for small size DIE CPUs."

 

The technical whitepapers ive read agree with that assessment as well.

de8auer pours LN2 on most things he has on hand though...

 

from the massive number of old soldered Xeons you can see in the used market, let's say quite a lot of them does survive all that thermal cycling with time to spare.

[mr moose]

OMG, guys, if soldering caused that many issues it would not be used in appliance and devices that suffer extreme heat variance let alone places with large physical movement to, like alternators, welders, aircraft etc.    Manufacturers spend a lot of time making sure they don't get a dry solder joint (which occurs during manufacture and not after) because they show up later after heat cycles and physical movement of the product.  

 

Source: my arse hole, where it underwent education on the topic back in 1992 as part of 4 years study* to be a fucking know it all internet forums.

 

*which coincidentally included Processor and chip design.

[MMKing]

Probably not economic. We already know Intel can't produce all the 14nm wafers they want at the moment. Maybe the combination of soldering and low supply of Wafers dominated the decision to go with the TIM solution. The 3175x likely won't be able to compete with the 2990WX in price to performance in the select tasks the 2990WX can handle really well, or should i say, the tasks it's not hampered by it's inherent 4 die, 1 IHS design.

 

I said probably not economic, by that i mean, probably not to save 5 dollars. Intel's ongoing issues with producing 14nm, of which we don't know the extent of, and production errors caused by soldering, which we also don't know the extent of. Had likely allot to do with the decision as well.

[ravenshrike]

14 hours ago, M.Yurizaki said:

This also leads to the reason why the RROD and YLOD of the 7th generation consoles happened: thermal stress caused fractures in the solder, leading to failure. Granted Microsoft claimed they used the wrong lead-free based solder. 

Completely different phenomenon. That was caused by whiskering.

 

https://en.m.wikipedia.org/wiki/Whisker_(metallurgy)

[Amazonsucks]

18 hours ago, Jurrunio said:

de8auer pours LN2 on most things he has on hand though...

 

from the massive number of old soldered Xeons you can see in the used market, let's say quite a lot of them does survive all that thermal cycling with time to spare.

I already explained, and as Leadeater's own chart shows, the temperature of CPUs in servers/HPC tends to vary less than a room temp to 70 or 80C desktop CPU.

 

 He did say that there would be spikes that got averaged into the chart, but hotspotting would similarly be absorbed and "evened out" physically by the cooling solution on the chip, as well as the package itself. Thats kind of the point of having solid copper heatsinks with high thermal mass and a high conductivity interface: to absorb transient hotspots to the point that they dont actually damage the silicon, increase electromigration or other heat related damage to the chip. In addition to serving that function, it keeps the package at a more stable temp.

 

Liquid cooling has become much more common in datacenters as well, and the advantage of the higher conductivity and heat capacity of water helps to keep transients from sufficiently heating the package to cause thermal cycling issues.

[leadeater]

7 hours ago, Amazonsucks said:

I already explained, and as Leadeater's own chart shows, the temperature of CPUs in servers/HPC tends to vary less than a room temp to 70 or 80C desktop CPU.

They vary 30C using real time data and server heatsinks aren't as good as something like a Hyper 212.

 

This is a HPE DL360 heatink, it's not full copper only a copper base plate. I've actually had one faulty where the base plate separated from the heatink, that was weird and unexpected. Came like that.

 

That's really small btw, has to fit in a 1U server. 2U heatink isn't much taller.

 

Also the point was never that a server CPU's heat varies at a greater range than a desktop CPU, the point is they thermal cycle exponentially more times over their life than a desktop CPU. Server CPUs under go more thermal cycling, you ended up making this in to a discussion about how much the variation is which was never the point.

[leadeater]

7 hours ago, Amazonsucks said:

Liquid cooling has become much more common in datacenters as well, and the advantage of the higher conductivity and heat capacity of water helps to keep transients from sufficiently heating the package to cause thermal cycling issues.

Most liquid cooling in datacenters is not direct die, they use heat collectors at the rear of the cabinet. You can't liquid cool switches and disk arrays so liquid heat extraction is more common and more typically preferred because it covers more equipment types and is generally safer, the devices are still air cooled and rely on heatsinks and fans. The heat extraction removes the heat from the room meaning you don't have to cool the room and cold aisle containment is used on the front side of the racks so cold air pockets are contained and goes directly in to the servers in those racks.

 

The video Linus did at the Canadian University shows this if you're interested in getting a better look.

 

 

 

Our container datacenters are liquid cooled, the room itself, so they are air isolated from the atmosphere. That means they can actually be partially submerged in water or ash in the air won't damage the inside of the container.

 

Direct die server cooling certainly exists though, Asetek is a big player in that market. I believe our SGI HPC cluster is water cooled but that's in a different city and operated by another group of people, I haven't actually seen that one. It's old as now though, really old.

[Amazonsucks]

I was talking about cooling more along these lines, which is more common in HPC, high end/mission critical machines.

 

 

https://goo.gl/images/TozC8p


https://goo.gl/images/fVcV2w


https://goo.gl/images/695k1X


https://goo.gl/images/ADqg9T


http://www.fujitsu.com/fts/products/computing/servers/primergy/scale-out/cclc/


https://goo.gl/images/fxAZ16

Although, whats old is becoming new again and the old Cray 80s style of immersion in dielectric liquid cooling is making something of a comeback in addition to the direct liquid.

 

 

 

 

[leadeater]

55 minutes ago, Amazonsucks said:

I was talking about cooling more along these lines, which is more common in HPC, high end/mission critical machines

Yes I know what you were talking about, it's not the most common type. Chilled rear door is because it's more able to fit the needs of more customers and can be retrofitted in to existing server racks with a rear clip on without having to get specific models of servers that are designed for direct chip water cooling.

 

The big datacenter operators who offer hosting can't count on customers buying rack space to be getting those kinds of servers so chilled rear door is an extremely good option while lowering cooling requirements of the facility making it more power efficient.

 

The reference design Nvidia DGX-2 HPC GPU server is air cooled, that's 16 V100's and a total system draw of 10KW and a rack draw of 60KW-80KW depending on if you have the power delivery at the rack level to handle that. Few do so you spread those out, unless you get US Department of [Somthing] money to throw in to top of the line everything every 3-5 years.

 

55 minutes ago, Amazonsucks said:

http://www.fujitsu.com/fts/products/computing/servers/primergy/scale-out/cclc/

That cooling solution is Asetek OEM, they make some really nice stuff.

 

Lenovo also has some really custom stuff as well, even water cooled ram which is ahhh... ok that's just cool (get it ).

https://lenovopress.com/lp0636-thinksystem-sd650-direct-water-cooled-server

 

HPE also has a decent range of direct chip server models too, in the Apollo and SGI product portfolios.

 

Was only a year to two years ago HPE made up 30% of the top 500 supercomputers list, Lenovo has now taken the largest share off them. IBM was king until 2007 with 50% share in both systems and performance.

 

55 minutes ago, Amazonsucks said:

Although, whats old is becoming new again and the old Cray 80s style of immersion in dielectric liquid cooling is making something of a comeback in addition to the direct liquid.

That's super extremely rear, won't see many of those. Been a couple of trial installations, more of study purposes than it actually being a good and cost effective idea. Direct chip water cooling makes far more sense.