Threadripper BUILDERS thread

[ub4ty]

Take this garbage to PM, no one cares about your being coy about being insulted by a meme, allegedly appropriated or not. Memes don't belong to or represent a single group, no matter what media tells you to think. We're on the internet. Deal with it.™

As far as the discussion about IHS coverage, heatpipe contact, etc:
1) The gains from optimizing either or both of these things will be marginal at best
2) Marginal gains can certainly help when you are pushing the limits
3) For the vast majority of users outside of enthusiasts/hardcore OCers, I doubt it makes a difference

That said, I should have an opportunity to compare a 240mm AIO (AC Liquid Freezer 240) versus an EK Supremacy EVO Threadripper edition water block with full custom loop, 360mm rad, etc. We will see if it makes any significant difference or not. And yes I will be pushing the limits and voiding warranties, as usual.

I will be pushing my build to an unworldly level. So these specific details matter significantly. That being said, the broader community profits immensely from people who call manufacturers out on their b.s and do reviews beyond a grade school level. The only site that has made me proud in such a way is Gamer's Nexus and funny enough they are directly inline w/ my comments. A margin is a margin... When you can engineer a proper product w/o significant impact to your bottom dollar it's what you do. At the core counts and sophistication level were currently at and heading towards, the review community would be far better served from non-shill based reviews and informed analysis based on real engineering beyond populist nonsense. Maybe I might tip my hat and trained eye into the mix to disrupt yet another stagnant ecosystem..

That being said, in case the new entrants forgot, if you want to measure surface temp, you have to get to the surface. Clearly w/ the mounting pressure and lack of die access, there are restrictions. However, the concept is there :

Not with silly flir imaging systems provided by paid sponsors pointed at surfaces that wont give you any information about what's going on above or even at the heatspreader. I am happy that you are going to provide very high level information about the respective coolers. However, what will be missing in the case of the inefficient AIO compatible coolers is the temperature gradient across the die that it doesn't accurately covered. CPU temps at the level that you can access them don't give you the whole picture.

The people manufacturing these solutions know exactly what the product's issues are :

Sadly, instead of being honest with their customers, they fill the gaps with handwaving and marketing b.s (compatible/ok margin). This is the kind of crap that results in crappy VRM heatsinks and M.2 heat insulators...

All this being said, there is level of analysis that can occur beyond any of these approaches that is far more accurate and will get these details. However, this involves hardware at a much lower level that is far beyond the reach of most e-celebs/populist review sites. Maybe this is where I enter into the fold. Maybe not. However, there is much to be desired from the generic expose' reviews and I've frankly had enough.

P.S - My money is where my mouth is on this platform. So, I have ever right to critique such obvious shortfalls in 3rd party add-ons. You can do analysis to confirm the obvious. However, since a thing called physics exists. You're not going to achieve better cooling by not contacting your heat-source vs. contacting it. Whatever margin that results in.. It's a margin. There is no reason for a margin or amateur hour gradient across a die Due to insufficient cooler engineering. The current cooler plates were designed for a single central die. If you want to make one for threadripper you put the copper fins that the water flows through above the friggin dies and expand the contact plate. In the case of air, you expand the contact plate and extend the heatpipe mating to it. Although it may seem like it is to some, this isn't rocket science. It's engineering 101.

Now, since it didn't occur to any of the talking heads to do something as simple as insert a 3600Mhz rated RAM speed stick while doing their review in order to confirm its even possible, I have to try to sort out what I'm going to do about RAM.

As the complexity of hardware increases, it would serve every well if we started behaving accordingly.

 

[ub4ty]

I ordered G.Skill Samsung B-die 3600Mhz RAM CL16 2x8GB
So, i'll test this 3600 OC+ ram claim shortly.
I am also doing so in order to understand Tom's hardware claim about :

Both of AMD’s CPUs are designed for a maximum power consumption of 180W at their default settings. If the memory’s overclocked, then the CPU gets 15W less, which might affect performance in usage scenarios that employ all of the cores and, consequently, get too close to the limit.

As 3600+(OC)/3200(OC)/2933(OC)/2667/2400/2133
are all OC, I fail to see how 15W just magically disappears w/o scale from 2933 to 3600. Obviously there's a power draw gradient (if one actually exists)... One that I was unaware even existed. As such, it appears that OC'ing or doing some kind of power unlock on the CPU and breaking it out of the 180W TDP lock is imperative. I'm not sure what mechanism is causing it to lock-out extra wattage usage but it'd be great to disable it and still have XFR w/ no set CPU OC.

My understanding is that 3600 CL16 can easily drop to 3200 CL14. So, there is a cushion w.r.t to timings and you're necessarily in the lowest CL14 timing scenario if you must drop clocks for some reason.

So, this is a good module to evaluate. If the module doesn't hit its spec'd rating, it's getting returned.

 

[Atari2600]

While the heatspreader effect won't be fully quantified until folks can do back to back tests (with the same chip) using different heatsinks, we can make a bit of a guess.


AMD have used indium solder onto gold plate onto the main heatspreader which is primarily copper with a Nickel plate on top.

So,
- The solder has a thermal conductivity of 80 W/mK.
- The gold has a thermal conductivity of around 300 W/mK
- The copper has a thermal conductivity of around 400 W/mK
- The Nickel has a thermal conductivity of around 90 W/mK


Good thermal paste has a thermal conductivity of around 8.5 W/mK.
Intel's TIM has a thermal conductivity of around 6 W/mK.

http://www.techarp.com/articles/delidded-threadripper-secrets-revealed/


Now, lets assume that the heat will pass through the gold/copper/nickel mix. Because the copper is most conductive, we will further conservatively assume it passes only through the copper - which we'll assumed is 1.25 mm thick in an IHS of overall thickness 1.5mm.

If we further assume only 70% of the die has normal conductance through the heatspreader to heatsink then the average distance the heat must travel is approx ~.0042m (based on full diameter of 51.9mm to enclose all 4 dies and thus 43.5mm dia to cover 70% of die area).

Take the conductive area as the circumference of 43.5mm cylinder with thickness 1.5mm and for copper you get a thermal resistance of 0.16 K/W.

Then, with 30% of 180W passing through this, the outside corners of the CPU dies will see a temperature rise of between 10 and 17 Kelvin (worst case scenario since we haven't added in the nickel or gold or indium solder).

Given that reviews are recording TR temperatures ~65degC at stock and ~80-85degC on overclock, and compare that to the Intel Core i9s... I don't think the TR heatsink/heatspreader incompatibilities are much of an issue.


[Saturation should be a non-factor as the heatsink will be continually extracting heat from the system and the overall thermal gradients should be large enough to render the localised effects rather minor.]

 

[Paratus]

So I'm going with a whole new build at the end of the year. I'll be replacing just about everything including the case.

I'm coming from an i7920 @3.6 onX58. I'm going to spec out an Intel mainstream and HEDT setup and an AMD mainstream and HEDT setup.

My use cases are office work, games, some video transcoding, and my wife does a bit of photo work. My intention is to keep it for several years. I'm also limited to a mid tower case.

I'm leaning towards an HEDT setup since that's what I currently have and I'm sort of choking on the idea of going from 4C8T, triple channel RAM and 40 PCIe lanes to 4C8T, dual channel RAM and 16PCIe lanes after 7 frickin years.

Can you guys provide some suggestions for what hardware to look for to get the most out of Threadripper?

I'm gathering high speed mem and high end cooling maybe a must?

Assume a Vega / 1080 type card.

With my use cases I'm also leaning towards the1900X 8C/16T (do we know anything about that one yet?) or at most the 1920X.

Thanks!

 

[raghu78]

So I'm going with a whole new build at the end of the year. I'll be replacing just about everything including the case.

I'm coming from an i7920 @3.6 onX58. I'm going to spec out an Intel mainstream and HEDT setup and an AMD mainstream and HEDT setup.

My use cases are office work, games, some video transcoding, and my wife does a bit of photo work. My intention is to keep it for several years. I'm also limited to a mid tower case.

I'm leaning towards an HEDT setup since that's what I currently have and I'm sort of choking on the idea of going from 4C8T, triple channel RAM and 40 PCIe lanes to 4C8T, dual channel RAM and 16PCIe lanes after 7 frickin years.

Can you guys provide some suggestions for what hardware to look for to get the most out of Threadripper?

I'm gathering high speed mem and high end cooling maybe a must?

Assume a Vega / 1080 type card.

With my use cases I'm also leaning towards the1900X 8C/16T (do we know anything about that one yet?) or at most the 1920X.

Thanks!

For your use cases I think Skylake-X and Threadripper are not the good choice. Your best bet is a core i7 8700k (6C/12T Coffeelake) if its priced at USD 350-400.

 

[krumme]

While the heatspreader effect won't be fully quantified until folks can do back to back tests (with the same chip) using different heatsinks, we can make a bit of a guess.


AMD have used indium solder onto gold plate onto the main heatspreader which is primarily copper with a Nickel plate on top.

So,
- The solder has a thermal conductivity of 80 W/mK.
- The gold has a thermal conductivity of around 300 W/mK
- The copper has a thermal conductivity of around 400 W/mK
- The Nickel has a thermal conductivity of around 90 W/mK


Good thermal paste has a thermal conductivity of around 8.5 W/mK.
Intel's TIM has a thermal conductivity of around 6 W/mK.

http://www.techarp.com/articles/delidded-threadripper-secrets-revealed/


Now, lets assume that the heat will pass through the gold/copper/nickel mix. Because the copper is most conductive, we will further conservatively assume it passes only through the copper - which we'll assumed is 1.25 mm thick in an IHS of overall thickness 1.5mm.

If we further assume only 70% of the die has normal conductance through the heatspreader to heatsink then the average distance the heat must travel is approx ~.0042m (based on full diameter of 51.9mm to enclose all 4 dies and thus 43.5mm dia to cover 70% of die area).

Take the conductive area as the circumference of 43.5mm cylinder with thickness 1.5mm and for copper you get a thermal resistance of 0.16 K/W.

Then, with 30% of 180W passing through this, the outside corners of the CPU dies will see a temperature rise of between 10 and 17 Kelvin (worst case scenario since we haven't added in the nickel or gold or indium solder).

Given that reviews are recording TR temperatures ~65degC at stock and ~80-85degC on overclock, and compare that to the Intel Core i9s... I don't think the TR heatsink/heatspreader incompatibilities are much of an issue.


[Saturation should be a non-factor as the heatsink will be continually extracting heat from the system and the overall thermal gradients should be large enough to render the localised effects rather minor.]

We know the temp on the cpu. And its what matters.
Either the plate is thicker than 1.5mm2 or more likely imo its less than 15% going through edges.
What else can there be?
What i find crazy is those full size plates were not ready but we will have to wait a week or two for them. I mean come on. Such a simple thing.

 

[ub4ty]

While the heatspreader effect won't be fully quantified until folks can do back to back tests (with the same chip) using different heatsinks, we can make a bit of a guess.

Then, with 30% of 180W passing through this, the outside corners of the CPU dies will see a temperature rise of between 10 and 17 Kelvin (worst case scenario since we haven't added in the nickel or gold or indium solder).

Given that reviews are recording TR temperatures ~65degC at stock and ~80-85degC
[Saturation should be a non-factor as the heatsink will be continually extracting heat from the system and the overall thermal gradients should be large enough to render the localised effects rather minor.]

This is the kind of analysis I like to see...
So : http://www.portvapes.co.uk/?id=Latest-exam-1Z0-876-Dumps&exid=threads/threadripper-builders-thread.2512631/page-12#post-39028004
10-15*C Differential from the die area that's covered by the cooler and the area that's not.
This is an issue IMO, a pronounced one when considering OC as the temps are 80-85*C and we have no clue what aspect of the CPU die is being attributed to that temp.
Quickly throwing 10-15 more degrees on 80/85 you get 90-95*C temps. vs a more tolerable 80/85*C.
I think the issue here is clear. It is clear enough when you look at how Coolers are compared and many times 5*C difference makes the decision. You're talking about 2-3x this difference simply because of insufficient uncomplicated engineering. I deleted your comments about intel because it isn't applicable to the discussion. Were speaking about threadripper and the insufficient coolers that are on the market.

Saturation is not an issue. However, thermal inefficiency is. That 30% uncovered region has to cross over somehow and it does so by going back across the already loaded 70% region. Simply stated, if you can cover the active heat region, you do. If you don't, you'll have thermal inefficiency. How much? You've done some calculations and arrived at my back hand conjecture .. ~10-15*C which is nothing quite frankly to sneeze at given that the processor nears thermal limits in OC. You're looking quite frankly at an OC limiter and running at unsafe temps simply due to some manufacturers not making a custom and sufficient thermal surface and using one intended for one central CPU die. I'm sorry but I can't wave that away. Given what temps the CPu is at in OC, the numbers you presented are even more damming as the inefficiency delta is just enough to throw the CPU into degradation and unsafe temps.

 

[ub4ty]

We know the temp on the cpu. And its what matters.
Either the plate is thicker than 1.5mm2 or more likely imo its less than 15% going through edges.
What else can there be?
What i find crazy is those full size plates were not ready but we will have to wait a week or two for them. I mean come on. Such a simple thing.

And where pray-tell is the sensor that gets you this reading? On the middle of the die? An average of multiple ones? On the edge? therein lies the problem.

 

[Markfw]

For your use cases I think Skylake-X and Threadripper are not the good choice. Your best bet is a core i7 8700k (6C/12T Coffeelake) if its priced at USD 350-400.

I think he is better off with a 1700 or 1700x

 

[Atari2600]

This is an issue IMO, a pronounced one when considering OC

I'd say its only an issue when overclocking.

I'd also wager the vast majority of these CPUs will be used in a stock clock configuration.

 

[Paratus]

For your use cases I think Skylake-X and Threadripper are not the good choice. Your best bet is a core i7 8700k (6C/12T Coffeelake) if its priced at USD 350-400.

Well that one is on my list of potentials for a mainstream Intel build. (Intel: 7700K, 8700K, 7820K; AMD: 1700X, 1900X, 1920x)

Although if they don't improve the motherboard support it'll still be16PCIe and dual channel RAM. (My current box has a GPU, a capture card, and an PCIE 4X USB 3.0/SATA 6.0 card). I feel like the mainstream boards are pretty constraining if I have to add anything other than a GPU.

Plus I'd still like to spec out a Threadripper build for comparisons sake, (at least once the 8700K and 1900X comes out since I've got the time).

 

[krumme]

I will be pushing my build to an unworldly level. So these specific details matter significantly. That being said, the broader community profits immensely from people who call manufacturers out on their b.s and do reviews beyond a grade school level. The only site that has made me proud in such a way is Gamer's Nexus and funny enough they are directly inline w/ my comments. A margin is a margin... When you can engineer a proper product w/o significant impact to your bottom dollar it's what you do. At the core counts and sophistication level were currently at and heading towards, the review community would be far better served from non-shill based reviews and informed analysis based on real engineering beyond populist nonsense. Maybe I might tip my hat and trained eye into the mix to disrupt yet another stagnant ecosystem..

That being said, in case the new entrants forgot, if you want to measure surface temp, you have to get to the surface. Clearly w/ the mounting pressure and lack of die access, there are restrictions. However, the concept is there :

Not with silly flir imaging systems provided by paid sponsors pointed at surfaces that wont give you any information about what's going on above or even at the heatspreader. I am happy that you are going to provide very high level information about the respective coolers. However, what will be missing in the case of the inefficient AIO compatible coolers is the temperature gradient across the die that it doesn't accurately covered. CPU temps at the level that you can access them don't give you the whole picture.

The people manufacturing these solutions know exactly what the product's issues are :

Sadly, instead of being honest with their customers, they fill the gaps with handwaving and marketing b.s (compatible/ok margin). This is the kind of crap that results in crappy VRM heatsinks and M.2 heat insulators...

All this being said, there is level of analysis that can occur beyond any of these approaches that is far more accurate and will get these details. However, this involves hardware at a much lower level that is far beyond the reach of most e-celebs/populist review sites. Maybe this is where I enter into the fold. Maybe not. However, there is much to be desired from the generic expose' reviews and I've frankly had enough.

P.S - My money is where my mouth is on this platform. So, I have ever right to critique such obvious shortfalls in 3rd party add-ons. You can do analysis to confirm the obvious. However, since a thing called physics exists. You're not going to achieve better cooling by not contacting your heat-source vs. contacting it. Whatever margin that results in.. It's a margin. There is no reason for a margin or amateur hour gradient across a die Due to insufficient cooler engineering. The current cooler plates were designed for a single central die. If you want to make one for threadripper you put the copper fins that the water flows through above the friggin dies and expand the contact plate. In the case of air, you expand the contact plate and extend the heatpipe mating to it. Although it may seem like it is to some, this isn't rocket science. It's engineering 101.

Now, since it didn't occur to any of the talking heads to do something as simple as insert a 3600Mhz rated RAM speed stick while doing their review in order to confirm its even possible, I have to try to sort out what I'm going to do about RAM.

As the complexity of hardware increases, it would serve every well if we started behaving accordingly.

Yeaa. I follow you about sites a loads of pr bs they transport from manufacturers to us.
And i also really like Steve on gamers nexus because his style and reviews is not like say an apple/intel/nvidia review on Anandtech eg on skl x "intel goes to their tdp". Lol. Such crap is not my style. I like when Steve hammered amd for their inmature ryzen paltform. Because it was inmature even for a new platform.

But steve also makes mistakes. You do that when you take firm standpoint all the time.
Tr temp was ok. Even if the drawings made us beliewe otherwise.
Be real here. Did you expect those temps with those initial coolers?

 

[dnavas]

... I fail to see how 15W just magically disappears w/o scale from 2933 to 3600. Obviously there's a power draw gradient (if one actually exists)... One that I was unaware even existed.

See The Stilt's post http://www.portvapes.co.uk/?id=Latest-exam-1Z0-876-Dumps&exid=thread...ores-of-awesome.2506474/page-58#post-39025304

Hopefully that's the only bump, and there is no other between 3200 and 3600.

 

[Paratus]

I think he is better off with a 1700 or 1700x

The 1700X was the mainstream AMD chip I was looking at. So good call.

 

[ub4ty]

I'd say its only an issue when overclocking.

I'd also wager the vast majority of these CPUs will be used in a stock clock configuration.

Think about this a little more and you see the blaring issue here.
It is a clear and defined significant issue in OCing.
It's a clear and defined issue even in stock.. The reason why it is in stock is because of XFR which pushes cores to 4Ghz and even beyond to 4.2Ghz (beyond what you can achieve in OC). It does this dynamically and has heat to go w/ it. Now imagine this occurs on Cores that are not covered at all by a cooler surface. You're getting dimensioned performance simply due to non-purpose built colors. There are no excuses for this. So, please stop looking for them . No matter what you can come up with, I'll point out exactly why it's still an issue. The reason why I can do this is because the issue is significant and fundamental. If it weren't, you'd find tons of ways to wiggle out from underneath my critiques.

 

[krumme]

I ordered G.Skill Samsung B-die 3600Mhz RAM CL16 2x8GB
So, i'll test this 3600 OC+ ram claim shortly.
I am also doing so in order to understand Tom's hardware claim about :


As 3600+(OC)/3200(OC)/2933(OC)/2667/2400/2133
are all OC, I fail to see how 15W just magically disappears w/o scale from 2933 to 3600. Obviously there's a power draw gradient (if one actually exists)... One that I was unaware even existed. As such, it appears that OC'ing or doing some kind of power unlock on the CPU and breaking it out of the 180W TDP lock is imperative. I'm not sure what mechanism is causing it to lock-out extra wattage usage but it'd be great to disable it and still have XFR w/ no set CPU OC.

My understanding is that 3600 CL16 can easily drop to 3200 CL14. So, there is a cushion w.r.t to timings and you're necessarily in the lowest CL14 timing scenario if you must drop clocks for some reason.

So, this is a good module to evaluate. If the module doesn't hit its spec'd rating, it's getting returned.

Its a good idea to be able to break eg 3000 specs and 180w tdp and still have xfr. Write amd!. Perhaps an option to set tdp yourself?
Let us know if you get to 3600.

 

[scannall]

The 1700X was the mainstream AMD chip I was looking at. So good call.

Microcenter has the 1800X at $349 now. Plus their bundle deal. Looks tempting.

 

[Atari2600]

It's a clear and defined issue even in stock.. The reason why it is in stock is because of XFR

A non-issue IMO.

XFR works based on the core properties, so it'll know if any particular core in any particular CCX delivers poor performance at higher clocks. Since its also only throttling up 2-4 cores, then it has ample choice of cores with optimal thermal paths.


You are making a mountain out of a molehill. The current situation is not ideal, but its likely not noticeable at all at stock speeds* and only noticeable under some workloads when overclocking.


*proof of the pudding will be the reviews that come with different heatsinks.

 

[ub4ty]

See The Stilt's post http://www.portvapes.co.uk/?id=Latest-exam-1Z0-876-Dumps&exid=thread...ores-of-awesome.2506474/page-58#post-39025304

Hopefully that's the only bump, and there is no other between 3200 and 3600.

Ah', I see :

Around 40W on average it seems (two separate VRMs combined) in this load, which is obviously the worst-case scenario.

3200MHz MEMCLK, which results in high SoC power draw due to the automatic VDDCR_SoC control present in Threadripper. The SMU adjust the SoC voltage automatically (unlike on AM4 parts), based on the set MEMCLK. 2133MHz = 0.850V, 2400MHz = 0.900V, 2666MHz = 0.950V, >= 2800MHz = 1.10625V.

So, we have : 2133/2400/2667/2933+(OC)/3200(OC)/3600(OC)

Breaking this out. We have Non OC clocks :
2133MHz = 0.850V, 2400MHz = 0.900V, 2666MHz = 0.950V
which have a gradient voltage and subsequent power draw but within non-cpu clock limiting range.

Then you shoot up to 1.10 SOC at anything above 2800Mhz with :
>= 2800MHz = 1.10625V for all OC whereby
3600/3200/2933 are all impacted by a 15W draw that allows for less for the CPU?

My fundamental question here is, if there is a clear gradient for non-oc RAM speed ranges and there isn't one for OC ranges, is there any instability or is the big voltage jump due to covering enough of the OC range that you have enough for 2933 up to 3600?

This indeed answered my question but now I have more about how exactly you can get more juice to the CPU to cover the 15W loss...

One thing I noted in the reviews it that it indicates that stock settings on the CPU actually outperform a 3.9Ghz OC. I wonder what the nature of this is considering the limits 3200 ram setting put on the CPU and that a number of reviewers did their reviews w/ 3200.

 

[ub4ty]

A non-issue IMO.

XFR works based on the core properties, so it'll know if any particular core in any particular CCX delivers poor performance at higher clocks. Since its also only throttling up 2-4 cores, then it has ample choice of cores with optimal thermal paths.


You are making a mountain out of a molehill. The current situation is not ideal, but its likely not noticeable at all at stock speeds* and only noticeable under some workloads when overclocking.


*proof of the pudding will be the reviews that come with different heatsinks.

Yet again, you're cornered. So you mean to tell me that I should deal with my threadripper only being able to use XFR on inner cores that are sufficiently covered simply because some clown third party manufacturer decided, instead of making a sufficient cooler for threadripper that they were going to repurpose an insufficient one made for single dies? No. That is unacceptable and is a mountain.

Even worse if the area of the die that's left uncovered is the CCX phy that is roasting and causing instabilities because I have it clocked at 3600 and its not getting sufficient cooling? Give up.

The compatible coolers are insufficient. Someone with their hands on it whose doing thorough analysis has already stated this :
http://www.portvapes.co.uk/?id=Latest-exam-1Z0-876-Dumps&exid=thread...ores-of-awesome.2506474/page-58#post-39025343

The situation with Threadripper compatible coolers is extremely pathetic at the moment. There is a single cooling solution (20+ variants, all made by Asetek) available for it.
Air coolers exist from Arctic, Noctua and Coolermaster but most of them are insufficient (Arctic, Coolermaster and the smaller Noctuas) and none of them are actually available..

The verdict is already out. These coolers are insufficient holdovers

 

[raghu78]

The 1700X was the mainstream AMD chip I was looking at. So good call.

You are going to have to look at 8700k max clocks for 24x7 vs 1700X at 4 Ghz in the workloads you care about. Since most of what you care about is going to be single thread performance driven the 8700k will probably be the better pick. Do your research but as far as I can say a 8700k at 5 Ghz will beat a 1700x at 4 Ghz anyday for your use cases.

 

[Markfw]

You are going to have to look at 8700k max clocks for 24x7 vs 1700X at 4 Ghz in the workloads you care about. Since most of what you care about is going to be single thread performance driven the 8700k will probably be the better pick. Do your research but as far as I can say a 8700k at 5 Ghz will beat a 1700x at 4 Ghz anyday for your use cases.

This is a threadrippers builders thread, not an 8700k builders thread, take your advice elsewhere.

 

[raghu78]

This is a threadrippers builders thread, not an 8700k builders thread, take your advice elsewhere.

The post I replied to was off topic for this thread. Anyway I will not post further off topic.

 

[Paratus]

This is a threadrippers builders thread, not an 8700k builders thread, take your advice elsewhere.

So since this is a Threadripper builders thread....

I've seen 3200 RAM recommended and used in various reviews online. Now G.Skill has a 32GB pack, (on the front page of AT), for Threadripper at 3600. Does generically faster RAM improve Threadripper performance where latency is concerned or is 3200 the ideal speed?

(I'm assuming latency is the issue. Quad channel 2400 and above should provide massive amounts of bandwidth)

 

[StefanR5R]

(concerns about contact area of Noctua and Arctic air coolers for socket TR4)

The bottleneck of these coolers will be radiator size.

The thick copper base plate of the Noctua coolers in particular will be fine. I also trust that Noctua got the shape of the bottom face right (probably needs to be plane) and made the mounting procedure as painless as possible for this socket. Both of these points are important for good thermal contact, and IME they got all this right on other sockets.

However, my opinion is that a 150 mm single tower radiator (NH-U14S) is probably OK for non-OC 180 W, but may be too little for a CPU which you want to overclock to, let's say, 250 W. (Personally I like it quiet, which is why I would prefer a 150 mm dual tower on a 180 W CPU, or even larger water coolers.)

I wouldn't consider a 120 mm or 90 mm tower cooler for a 180 W CPU. In fact, neither do Noctua according to their TDP guide (which has not been extended for TR4 and SP3 yet). I am puzzled why they mention TR4 compatibility of NH-U12S TR4-SP3 and NH-U9 TR4-SP3 if no TR4 CPUs with matching low TDP currently exist.

@ub4ty, with all your worries related to heat conduction in metal, you seem to forget about the bottleneck which is located in the boundary layer between metal and air.