Monarch butterfly prototype

[Threader]

I'm thinking this is in the pipeline for the 600gh Monarch

http://www.asus.com/ROG_ROG/ARES26GD5/

[1715018782]

1715018782

[MrTeal]

From the BFL Forums:


There are some VERY poor design choices on that board:

While a switch to what appears to be a 5-phase regulator per chip is a very good design choice, the choice to use a non-synchronous regulator (you can spot one of these by the need to use a diode(s) in the power path) is very poor.

1/ Synchronous regulators are more efficient than non-synchronous regulators, particularly at low voltages.
2/ Non-synchronous regulators require a diode powerful enough to handle the full current of the output. (That's a BIG diode!)
3/ BFL engineers appear to think they can use lots of smaller diodes in parallel instead of one big diode - that's a big no-no! Google 'diodes in parallel' if you care to know why!

I can't remember the last time I saw a motherboard which had anything other than a multi-phase synchronous regulator for the CPU - and they don't have to deliver as much power as the Monarch does!

What are you talking about, that is a synchronous regulator. You can see the high and low side fets. Those aren't diodes, they're Panasonic polymer aluminum caps.

[joeventura]

From the BFL Forums:


There are some VERY poor design choices on that board:

While a switch to what appears to be a 5-phase regulator per chip is a very good design choice, the choice to use a non-synchronous regulator (you can spot one of these by the need to use a diode(s) in the power path) is very poor.

1/ Synchronous regulators are more efficient than non-synchronous regulators, particularly at low voltages.
2/ Non-synchronous regulators require a diode powerful enough to handle the full current of the output. (That's a BIG diode!)
3/ BFL engineers appear to think they can use lots of smaller diodes in parallel instead of one big diode - that's a big no-no! Google 'diodes in parallel' if you care to know why!

I can't remember the last time I saw a motherboard which had anything other than a multi-phase synchronous regulator for the CPU - and they don't have to deliver as much power as the Monarch does!

What are you talking about, that is a synchronous regulator. You can see the high and low side fets. Those aren't diodes, they're Panasonic polymer aluminum caps.

Just quoting someone on BFLs website

I am not an engineer but I play one on TV

[fractal02]

From the BFL Forums:


There are some VERY poor design choices on that board:

While a switch to what appears to be a 5-phase regulator per chip is a very good design choice, the choice to use a non-synchronous regulator (you can spot one of these by the need to use a diode(s) in the power path) is very poor.

1/ Synchronous regulators are more efficient than non-synchronous regulators, particularly at low voltages.
2/ Non-synchronous regulators require a diode powerful enough to handle the full current of the output. (That's a BIG diode!)
3/ BFL engineers appear to think they can use lots of smaller diodes in parallel instead of one big diode - that's a big no-no! Google 'diodes in parallel' if you care to know why!

I can't remember the last time I saw a motherboard which had anything other than a multi-phase synchronous regulator for the CPU - and they don't have to deliver as much power as the Monarch does!

What are you talking about, that is a synchronous regulator. You can see the high and low side fets. Those aren't diodes, they're Panasonic polymer aluminum caps.

Just quoting someone on BFLs website

I am not an engineer but I play one on TV

Sooo....you are quoting about things you don't understand...oookayy...NEXT !

[augustocroppo]

Don't you see the ETL/FCC and CE logos on the card??

Well, that is most strange part. It is not a prototype, it is just a PCB without the main processor. So how BFL managed to get this certifications without a finished prototype? This is really amazing. Look like everything goes backward in BFL business.

Instead to them...

Design -> Main processor ->  PCB -> Quality Test -> Announcement -> Sales

They go by...

Sales -> Announcement -> Quality Test -> PCB -> Main processor -> Design

[DeathAndTaxes]

Thats not entirely fair. You measure the size of a heatspreader, the actual diesize will determine the thermal density and for KnC's sake, I hope thats no where near 1000mm².

True I was simplifying to avoid turning it into a page long post.  However heat spreaders are larger than the die for a reason.  The thermal conductivity between silicon and the IHS is generally higher than the IHS and heatsink (or waterblock or air).  That is one reason the IHS is generally larger than the die.  If one the die size and die heat flux mattered there would be no reason (well no thermal reason) to use a larger IHS then the die.  

Moreover, the entire surface of these asics will generate about the same amount of heat per mm² very uniformly . On a modern cpu,  something like 90% of the power is used by the cpu core which in most cases occupies only a small fraction of the die (the rest is cache, memory controller, etc). Im too lazy too look up numbers, but Im pretty sure a highend cpu will have a far higher thermal density at its core than any of these asics.  Silicon has fairly good thermal conductivity, mitigating the problem for cpu's to some extend, but its no were as good as copper or even aluminum.

That is a good point on the "even" heat flux however GPU unlike CPU tend to be mostly core engine and thus are more like mining ASICs in that respect.  One ASIC vendor (HF) uses 4 9mm x 9mm dies with 5mm space between them in a single package to lower the heat flux.

So thermal density isnt going to be problem with these chips. Of course you still need to get rid of 100's of watts of power, no doubt that requires a serious cooler, but the size of the chips wont make a lot of difference.

Well no.   Thermal density does matter, it may not be a problem but it does matter.  A Bitfury rig uses ~400W yet needs no heat sink.   200 USB miners use ~400W yet don't even need a fan.   If you have a large lawn it probably has a thermal output (from IR exposure to the sun) of a couple KW (but at a very low heat flux) yet I doubt you run out there on a hot day and crank up a serious cooler to keep the lawn from bursting into flames.

All things being the same the higher the heat flux the more complex the cooling to remove the same amount of energy (watts) from a smaller space.  Baring some exotic tech that means larger heat sinks, more efficient (heat pipes) heat sinks, and more airflow.   With the constraint of a dual slot cooler that puts some boundaries on what is possible in the airflow and size parameters. 

[Puppet]

All things being the same the higher the heat flux the more complex the cooling to remove the same amount of energy (watts) from a smaller space.  Baring some exotic tech that means larger heat sinks, more efficient (heat pipes) heat sinks, and more airflow.   With the constraint of a dual slot cooler that puts some boundaries on what is possible in the airflow and size parameters.  

My point is that BFL isnt going to struggle with hotspots on that chip and that its a smaller chip than KnC isnt their problem. Putting a 1000mm² heatspreader on a 300 mm² die (or set of dies) isnt going to work miraculously better than a 300mm³ heatspreader with a similar sized cooler. In fact, most overclockers will tell you you get better results by removing the CPU heatspreader and putting the cooler directly on the much smaller die (which is the case with most GPU's). The more important reason for that heatspreader is not miraculous heat conductivity, but making it less likely that you crack the chip or misalign the cooler..

Anyway, removing the 350+W of heat in a dual slot GPU sized package will be a stiff challenge, no question about that. Putting more than 2 in a case probably damn near impossible.  But I dont think it would be considerably easier with a monster size chip like KnC. In fact, and slightly OT,  Im looking forward to seeing some close ups of their cooling solution, AFAIk those  intel and amd coolers from Artic dont have a base surface area large enough to cover their entire heatspreader. Did they order a custom batch; or is part of their heatspreader not covered?

[xstr8guy]

Christ that thing is huge. Finally we're going to get more haphazardly kludged together fire hazards.

Don't you see the ETL/FCC and CE logos on the card??

It's safe!!!

 

Well, there aren't actually any traces yet, so I'd imagine it will be quite safe. It might just be an early component layout so that they can get the physical dimensions nailed down.

Yes, that is my opinion, too.  The date on the board is August, as can be seen by zooming up.

They don't need to be mounted to a motherboard.  They can be freestanding or placed in other containers.  They will have USB connectors too.

[DeathAndTaxes]

In fact, and slightly OT,  Im looking forward to seeing some close ups of their cooling solution, AFAIk those  intel and amd coolers from Artic dont have a base surface area large enough to cover their entire heatspreader. Did they order a custom batch; or is part of their heatspreader not covered?

They are stock and the base appears (eyeballing the board and heatsink photos larger than the package size.  Not sure if that is intentional design choice on Artic's part or if KNC just lucked out that Artic's sinks are larger than average.  Due to the direct heat pipe design there likely are some limits on how "short" the base of the pipes can be made.

My point is that BFL isnt going to struggle with hotspots on that chip and that its a smaller chip than KnC isnt their problem.

That wasn't my claim either.  Smaller chip, higher power consumption, constrained device size combined create a significant thermal "challenge".  Not impossible but BFL has had difficulty properly cooling devices with much lower power densities and much more "open" form factors.

BFL is no AMD and even AMD had issues (and 7 month delay) cooling the 7990 and the 7990 still uses less power than BFL design is simulated to use.  If BFL is even over moderately it could go from insanely difficult to (economically) impossible.

[Puppet]

They are stock and the base appears (eyeballing the board and heatsink photos larger than the package size.

According to the drawings on arctic's site:
http://www.arctic.ac/en/p/cooling/cpu/473/freezer-a30.html
(click on compatibility, then scroll down)
it looks like the contact base front to back is no deeper than ~21mm, maybe 25mm if you include the aluminium bracket which Im not sure is supposed to make contact; thats considerably less than the 41mm of the KnC heatspeader. What picture are you looking at?

That wasn't my claim either.  Smaller chip, higher power consumption, constrained device size combined create a significant thermal "challenge".  Not impossible but BFL has had difficulty properly cooling devices with much lower power densities and much more "open" form factors.

BFL is no AMD and even AMD had issues (and 7 month delay) cooling the 7990 and the 7990 still uses less power than BFL design is simulated to use.  If BFL is even over moderately it could go from insanely difficult to (economically) impossible.

I think we are largely in agreement there.

[MrTeal]

In fact, and slightly OT,  Im looking forward to seeing some close ups of their cooling solution, AFAIk those  intel and amd coolers from Artic dont have a base surface area large enough to cover their entire heatspreader. Did they order a custom batch; or is part of their heatspreader not covered?

They are stock and the base appears (eyeballing the board and heatsink photos larger than the package size.  Not sure if that is intentional design choice on Artic's part or if KNC just lucked out that Artic's sinks are larger than average.  Due to the direct heat pipe design there likely are some limits on how "short" the base of the pipes can be made.

My point is that BFL isnt going to struggle with hotspots on that chip and that its a smaller chip than KnC isnt their problem.

That wasn't my claim either.  Smaller chip, higher power consumption, constrained device size combined create a significant thermal "challenge".  Not impossible but BFL has had difficulty properly cooling devices with much lower power densities and much more "open" form factors.

BFL is no AMD and even AMD had issues (and 7 month delay) cooling the 7990 and the 7990 still uses less power than BFL design is simulated to use.  If BFL is even over moderately it could go from insanely difficult to (economically) impossible.

Actually a 7990 will draw ~350W DC running furmark at stock clocks, or 100W more with overclocking. It definitely still won't be easy, but it's not like it's impossible. GPU makers are constrained by useability issues, they can push so far but people eventually put 7990s in cases that can have pretty poor airflow. The also need to at least pretend to care about acoustics. BFL doesn't have that problem, and a lot of cooling issues can be cured by sacrificing your hearing later in life to install a ridiculous amount of airflow.

I think they'll have less trouble cooling these devices than the Singles. The biggest problem with the Singles was the overdriven power supply running inefficiently and being poorly cooled prior to using all those little press on heatsinks everywhere, as well as being forced to use a gapfiller pad rather than paste due to levelness and coplanarity concerns. Remove the massive thermal impedance of the pad, and not only will your core temperatures drop, but you dump much less heat into the PCB and everything else stays cooler.

[Puppet]

cooling those single's is positively childsplay compared to that Monarch. The fact they managed to botch that is telling.

[DeathAndTaxes]

They are stock and the base appears (eyeballing the board and heatsink photos larger than the package size.

According to the drawings on arctic's site:
http://www.arctic.ac/en/p/cooling/cpu/473/freezer-a30.html
(click on compatibility, then scroll down)
it looks like the contact base front to back is no deeper than ~21mm, maybe 25mm if you include the aluminium bracket which Im not sure is supposed to make contact; thats considerably less than the 41mm of the KnC heatspeader. What picture are you looking at?

Well the heat pipes are 8mm diameter each.  So that puts the minimum width more like <32mm. 

I took the KNC board photo measured the pixels across to get the resolution (pixels per mm).  Then measured the pixels between the two mounting holes.  Board has 6 mounting holes 4 inner which I believe are unused and two outer which are used for the heatsink.   That gave me the distance between the two mounting holes.  Then using the heatsink photo did the same thing to estimate the contact surface.  I did this a month ago for my own research but I may have made a mistake.

[MrTeal]

cooling those single's is positively childsplay compared to that Monarch. The fact they managed to botch that is telling.

I would disagree. Of course we don't know how the die is internally attached or even if it's a single monolithic die, but the chances are that the silicon interface to the base of the heatsink will be much, much better than with the Singles. Keep in mind that we're talking similar power levels here, my Single pulls 300W from the wall.

http://www.anandtech.com/show/6830/cpu-air-cooler-roundup-six-coolers-from-noctua-silverstone-and-cooler-master/4
Look at some good air or water coolers, and the whole die->heatspreader->paste->heatsink->ambient thermal path might have an effective impedance in the range of 0.25C/W to 0.5C/W.

Compare that to the even the exceedingly good and expensive thermal pads BFL uses.
http://www.fujipoly.com/usa/products/sarcon-thermal-management-components/thermal-gap-filler-pads/high-performance-gap-filler/xr-m.html (they use the 50X-Hm).
0.06Cin^2/W at a very high pressure of 72psi. Since the BFL die is 0.1in^2, that gives an effective thermal impedance of 0.6C/W just across the thermal pad, in the best case. In the case that one die is lower (or worse one is noticeably higher) those numbers could be much worse.

TL-DR, the thermal impedance of just BFL's expensive ($11/each if you cut 24 out of a full sheet) thermal pads is higher than the entire die to ambient impedance of a decent CPU cooler system.

[erk]

cooling those single's is positively childsplay compared to that Monarch. The fact they managed to botch that is telling.

I would disagree. Of course we don't know how the die is internally attached or even if it's a single monolithic die, but the chances are that the silicon interface to the base of the heatsink will be much, much better than with the Singles. Keep in mind that we're talking similar power levels here, my Single pulls 300W from the wall.

http://www.anandtech.com/show/6830/cpu-air-cooler-roundup-six-coolers-from-noctua-silverstone-and-cooler-master/4
Look at some good air or water coolers, and the whole die->heatspreader->paste->heatsink->ambient thermal path might have an effective impedance in the range of 0.25C/W to 0.5C/W.

Compare that to the even the exceedingly good and expensive thermal pads BFL uses.
http://www.fujipoly.com/usa/products/sarcon-thermal-management-components/thermal-gap-filler-pads/high-performance-gap-filler/xr-m.html (they use the 50X-Hm).
0.06Cin^2/W at a very high pressure of 72psi. Since the BFL die is 0.1in^2, that gives an effective thermal impedance of 0.6C/W just across the thermal pad, in the best case. In the case that one die is lower (or worse one is noticeably higher) those numbers could be much worse.

TL-DR, the thermal impedance of just BFL's expensive ($11/each if you cut 24 out of a full sheet) thermal pads is higher than the entire die to ambient impedance of a decent CPU cooler system.

If you set aside the classic screwup on 65nm ASIC power consumption BFL did, the rest of their work shows they have some pretty competent electronic engineers even their firmware API is more elaborate than the competition.

[klintay]

RUBBISH! bfl please find a tall building and jump off...do the whole scene a favour

[kmtan]

should be a heater

[Bicknellski]

My point is that BFL isnt going to struggle with hotspots on that chip ... (snip)

What chip?

[Unacceptable]

My point is that BFL isnt going to struggle with hotspots on that chip ... (snip)

What chip?

[Bicknellski]

My point is that BFL isnt going to struggle with hotspots on that chip ... (snip)

What chip?