Graphics, Hardware

Can Prolimatech GPU Heatsink Cool Down Quadro and FirePro?

There’s no beating around the bush – 3rd party GPU [Graphics Processing Unit] cooling is a fast growing market. While this was a tiny niche in the past, everything started to change in 2002, when I showed Sapphire’s marketing rep ATI’s reference 9700 Pro with the passive Zalman heatsink [and used no fans to cool the board]. Dan Forster of Sapphire Blue managed to persuade the management to launch passively cooled 9500 and 9700 graphics cards and the rest is history.

Today, every GPU vendor has at least one model in its lineup that uses customized cooling, usually overclocked and carrying a good price premium over the stock parts. Some go as far as creating completely customized board which can be compared to a Formula One car. However, there is one market segment where customized cooling still hasn’t arrived, mostly due to reliability fears: professional graphics.

Just like liquid cooling, hard-working professionals demand top performance, yet the experience in the offices that feature powerful, dual-socket, 32GB+ system memory carrying, often multi-GPU enabled systems created an industry of rackable workstations, accessible through expensive cabling that leads to a workstation. Yet, isn’t it better to solve the problem at the source, rather than beating around the bush?

Professional Applications are hardly a place where a 3rd party cooler would be used...can Prolimatech break the mould and become viable in content creation industry?
Professional applications are hardly a place where 3rd party cooling is used… can Prolimatech break the mould and become viable in engineering?

As you probably know, one of teams inside Bright Side Network Inc. is heavily involved in automotive development and working on multi-CPU systems equipped with AMD FirePro and nVidia Quadro boards. After a while, our engineers complained that some models produce noise. Given that the professional boards are usually under-clocked consumer parts, default cooling usually does a pretty good job and keeps the GPUs quiet. However, when the fan on PNY Quadro CX started to produce "funny noises", and two weeks after brand new AMD FirePro V8800 joined in and created a cacophony in the working area, we knew that the time was right to test can a 3rd party GPU cool down these beasts.

Meet Prolimatech MK-13
Prolimatech MK-13
Unlike a lot of competitors that use plastic packaging, Prolimatech uses recyclable cardboard

We looked around for an ideal solution, when we noticed that Taiwanese progressive cooling expert, Prolimatech Ltd. announced MK-13, their massive GPU heatsink. In a matter of days, we received Prolimatech MK-13 from, one of largest distributors of computer equipment for enthusiasts in Europe. Caseking offers MK-13 in several different packages, varying from very powerful to near-silent fans, as well as stand-alone, passive version. Our version came with two ultra-silent 120mm S-Flex fans from Schyte.

The box is sturdy and will keep its contents well protected during the long travel from Taiwan to Western markets, as well as reaching its final destination – your home or office. We checked the sturdiness with our traditional drop’n’kick test and we can report that nothing happened with the retail packaging, not even a dent.

The manufacturer took good care about protecting the contents inside the box, with plenty of protective packaging. Thus, we saw no damage on the delicate aluminum fins – 75 of them, to be more precise.

The contents of the box: GPU Heatsink, Paste, 5800 Adapter, DRAM heatsinks, fan rails and adapters

The box features every possible combination you could imagine: besides the MK-13 heatsink, there are also PK-1 thermal paste, 18 2.5cm [1"] heatsinks for the DRAM memory, and eight lower heatsinks, intended to cool down external chips such as nVidia’s NVIO and BR02 bridges, special heatsink for AMD Radeon 5800 series of cards [we’ll have to see if its compliant to AMD’s upcoming HD 6800 and HD 6900 series of cards], as well as several adapters for add-on fans. Overall, we’re quite sure that you should have no issues installing this part on standard reference or on custom cards.

DRAM heatsinks on Prolimatech MK-13 ship with factory adhesive. As it happens, this is where the product makes it or breaks it.
DRAM heatsinks on Prolimatech MK-13 ship with factory adhesive.

Install Procedure: AMD FirePro V8800 2GB
AMD FirePro V8800 2GB - AMD's high-end workhorse
AMD FirePro V8800 2GB – AMD’s high-end workhorse

Prolimatech states that 28 graphics cards are compatible with the MK-13 heatsink, but the interesting bit is that allegedly, GeForce GTX 480 is not supported, while GTX 470 is. Given that we installed this part on a GTX 480 and it works with no major issues, it looks like on its own [without external fans], MK-13 can’t cool down the hot’n’power hungry GF100 "Fermi" chip. However, Caseking’s kit includes two silent S-Flex fans and they significantly reduce the temperature achieved by the parts. Naturally, the company didn’t list a single workstation-class graphics card.
Prolimatech also states that there are known incompatibilities with some odd-shaped cases on the market, such as SilverStone’s RV0
2 "Raven" case. Given that we use the aluminum version [FT02], we checked that incompatibility claim as well.
But first and foremost, the reason why we acquired the MK-13 heatsinks was very simple: can they reduce the noise level of our SuperMicro and HP workstations?

The first board that came under scrutiny is AMD’ FirePro V8800 2GB. Even though this card is not the fastest card on the professional market, it comes with one significant advantage over its green competition: true multi-display support. While even the latest Quadro boards drive only two displays, FirePro V8800 can drive four displays at the same time.

Cypress GPU ticks at 825MHz, which is only 25MHz less than the Cypress chip powering the consumer board [HD 5870], while 2GB of GDDR5 memory tick at 1GHz in QDR mode, i.e. 4 GT/s. Memory is downclocked by 200MHz [HD 5870 = 1.2GHz QDR, 4.8 GT/s].

AMD FirePro V8800 2GB laid bare - The board is filled with components
AMD FirePro V8800 2GB laid bare – The board is filled with components with almost no room to spare

The install procedure is very simple: we took the FirePro V8800 apart and while there are no major changes in removing the heatsink, there is one part you have to be very careful about. When you look at the card from the back side, the lower left corner [front of the card] is connected with a screw from inside – thus you need to show a bit flexibility in getting that screw out and liberating the backplate. Thus, you need to remove all the screws from the back, turn the board upside and remove the main heatsink. Unscrew that screw and the backplate will come off too.

Custom PWM Heatsink for AMD Radeon HD 5800 Series does a perfect job with FirePro as well
Custom PWM Heatsink for AMD Radeon HD 5800 Series does a perfect job with FirePro as well

After that, clean the GPU area from the factory default paste, put the Prolimatech PK-1 thermal compound on the GPU and placed the custom-built heatsink for the V8800 power regulation [identical to consumer Radeon HD 5830, HD 5850, HD 5870 boards].

Tall heatsinks installed on every memory chip [front first, then back]
Tall heatsinks installed on every memory chip [front first, then back]

Next step was installing the heatsinks for the eight front GDDR5 chips, with the ninth heatsink positioned on the power regulator on north west side of the GPU chip. Given that the board features 16 memory chips, MK-13 package will leave one of these tall heatsinks to spare.

Firming the rear mount. Note that you will experience clearance issues with first and last memory heatsinks on top side. Simply position heatsinks a bit higher to avoid any spacing issues
Firming the rear mount. Note that you will experience clearance issues with first and last memory heatsinks on top side. Simply position heatsinks a bit higher to avoid any spacing issues

After everything was placed firmly, we positioned the MK-13 heatsink over the GPU and turned the board upside down to place the adapter and firm the heatsink.

Ready for Action: Passively Cooled AMD FirePro V8800 2GB, powered by Prolimatech MK-13
Ready for Action: Passively Cooled AMD FirePro V8800 2GB, powered by Prolimatech MK-13

Install Procedure: NVIDIA Quadro CX  / FX 4800 1.5GB
nVidia Quadro CX is identical to FX 4800, with the only difference being the bundle. After CS5 development took off, the product was discontinued and replaced with FX 4800.
One of our workhorses finally gave up: the fan went bust. Luckily, here’s Prolimatech to the rescue.

The second board today to receive the Prolimatech MK-13 treatment was nVidia Quadro CX, followed by another FX 4800. However, we’ll only show pictures of CX, given that the boards are identical: 55nm GT200b GPU, 1.5GB GDDR3 memory, 192 cores. All performance results come from FX 4800, though – as the fan on CX was broken.

Unlike the consumer brothers, GT200b GL chip in this Quadro board operates just like newer Fermi-based GPUs: 1:2 ratio. Quadro’s GPU ticks at 602MHz, while the 192 cores operate at 1,204 MHz, i.e. 1.2 GHz. Memory clock is identical to almost every Quadro board in the past couple of years: 1.5GB of GDDR3 memory operate at 800MHz in DDR mode, resulting in 102.4GB/s of video memory bandwidth.

The only real difference between the two was the bundle anyways [and the price]: CX originally shipped with Elemental Accelerator, professional transcoder for Adobe Creative Suite 4, developed by Elemental Technologies. After the introduction of Mercury Playback Engine, the product was discontinued and replaced with FX 4800.

Opening the backplate isn't easy as it seems...removing the screws is easy, but you need to use phyisical force for plate removal.
Removing the backplate isn’t easy as it seems…removing the screws is easy, but you need to use physical force for plate removal.

The board isn’t physically different than regular GT200-based GeForce cards: the two plastic covers are firmed with 12 screws, but the removal of them is just the beginning of the process – you have to use certain level of physical force to separate the two plates.

After the backplate is removed, you need to liberate the front heatsink, which is nearly glued to the GPU surface thanks to the amount of thermal interface material nVidia used. In case anyone is interested, thermal paste was identical to the one used by GeForce GTX 285 1GB.

In case of nVidia boards, Prolimatech advises that you use lower memory heatsinks for the memory and the NVIO chip, while use the tall ones for the power regulators
In case of nVidia boards, Prolimatech advises that you use lower memory heatsinks for the memory and the NVIO chip, while use the tall ones for the power regulators.

Unlike the FirePro V8800, which utilizes 2GB highly-clocked GDDR5 memory, Prolimatech advises owners of GT200b-based boards to use low heatsink, and use nine tall heatsinks for the power circuitry. The manual is well-laid out and it is easy to see where to position the heatsinks.

The 55nm GT200B chip takes a lot of paste, make sure you spread it thinly across the whole IHS [Integrated Heat Spreader]
The 55nm GT200B chip takes a lot of paste, make sure you spread it thinly across the whole IHS [Integrated Heat Spreader]

nVidia’s design decision to go with an IHS is a clear evidence that the company is fighting "hot spots", present on almost all CPUs and certain GPUs. That’s the consequence of architectural design, with processing cores working at a higher clock than the rest of the chip. If you see an heatspreader on your GPU, you have to apply the thermal material across the whole heatspreader.

Putting the MK-13 on the GPU was very simple: position it properly, turn it upside down, firm the bolts and that's it.
Putting the MK-13 on the GPU was very simple: position it properly, turn it upside down, firm the bolts and that’s it.

Overall, the replacement of the stock heatsink to a completely modified board lasted for 14 minutes on an AMD FirePro V8800 and around 16 minutes on nVidia Quadro CX/FX 4800. The difference is solely the amount of time required to put the thermal paste onto the very large piece of copper [nickel-plated].

As stated before, our primary goal was to put the FirePro and Quadro back into machines, in order to keep the high level of productivity within the team. The workstations in question are based on the following chassis. Do bear in mind that all published scores are based on tests done in SilverStone RV01 Raven Case:

Fortress, or FT02 is without any doubt, one of most beautiful cases we have laid our hands on. The quality of the finish is something that original Raven series just missed, and it does offer evolutionary positioning for the motherboard. Given that the motherboard sits at a 90 degree angle, rules of physics apply and the heat naturally flows upwards, without any risk of gaining hotspots on the components inside. The only real problem is that the chassis does not support full sized E-ATX and WTX motherboards, meaning ATX up to 12 inches in height [305mm] is supported.

On the other side, Raven RV01 was the first case that started this trend. The case is bigger than the FT02 and fits full sized eATX motherboards, meaning you can install even dual socket motherboards. This case was developed in cooperation with Maingear Computers, who are using higher-end version of this case for their Shift systems. However, the case has one major advantage over any other case on the market [with the exception of Raven RV02 and Fortress FT02]: thermals flow naturally, and you can count on lower system temperatures, even with dual-socket systems.

In contrast, SuperMicro workstation systems are all about efficiency, instead of looks. This workstations are heavy-duty workhorses that fit dual-socket systems, are capable of supporting multiple hot plug hard drives [in our case, multiple Intel 160GB SSD drives].

Our test systems were as follows:

  • SilverStone Raven RV01 [provided by SilverStone]
  • Intel Core i7 Extreme 965 3.2 @ 4.0 GHz [provided by Intel]
  • ASUS P6T7 SuperComputer [provided by ASUS Computer International]
  • 12GB Kingston HyperX DDR3-1600 [3x4GB]
  • 4x 160GB Intel SSD
  • 2x 600GB Seagate Cheetah 15k7 [provided by Seagate]
  • BeQuiet Dark Power PRO 1.2 kW [provided by Listan/BeQuiet!]
  • SilverStone Fortress FT02 [provided by Caseking]
  • Intel Core i7 Extreme 975 3.33 GHz
  • EVGA FTW2 Classified [provided by EVGA Europe]
  • 12GB Kingston HyperX DDR4-1600 [provided by Kingston]
  • 120GB OCZ Vertex 2 [provided by OCZ Technology]
  • 4x 2TB Seagate Constellation ES
  • Antec TruPower 1.2 kW [provided by Antec]
  • SuperMicro SC747TQ-R1400B
  • 2x Intel Xeon
    E5640, 32nm 4C/8T 2.2GHz
  • Supermicro X8DT3-LN4F
  • 24GB Samsung DDR3-1066 Reg. ECC
  • 4x 160GB Intel SSD
  • 2x 1TB Samsung
  • Supermicro 1.4 kW Gold Level PSU

As stated above, our tests were made using the SilverStone Raven RV01 case and the appropriate components. Even though it may sound odd that a solidly overclocked system is serving as a workstation in an automotive development studio, the reason for that is quite simple: clock beats multi-threading. In the case of the Siemens NX-7 a.k.a. Unigraphics software, the application simply doesn’t care if you have one or sixteen cores running at 2.26GHz – the performance will remain the same. Thus, the only logical choice was to go with a liquid cooled setup for the CPU, and clock it as high as it can go stable. The next upgrade for that system will be an i7-990X CPU, hopefully with a stable 4.5GHz clock. Or perhaps Siemens will discover the wonders of multi-threading and GPGPU acceleration.

The way how we test is quite simple: room is kept at 22 degrees Celsius by a powerful air-conditioner. Idle measurement is taken after 30 minutes of inactivity after load of Windows 7 [3D-accelerated Aero interface is enabled]. After that, we load up MachStudio Pro and leave the GPU render "on" in our test scene. After 30min of rendering, we measure the temperature. Secondly, we leave the board to cool off by leaving the system idle until the original idle temperature is reached, and then start Furmark "Hotter than Hell" Edition. From our experience, even though MachStudio Pro results with somewhat lower temperatures than Furmark, we do believe MachStudio Pro is the better choice, as it is actually working on a real world scene.

Prolimatech MK-13 and FirePro V8800: Who needs fans?
AMD FirePro V8800 2GB reached an idle temperature of 36C. By using a "power virus" called FurMark, the V8800 recorded 82C in FurMark. Real-world test didn’t lag behind: 77C in MachStudio Pro under 100% load. Given that in the very same case Radeon HD 5870 Eyefinity6 2GB recorded 91C in FurMark and 89C in MachStudio Pro, the temperature difference is impressive. AMD obviously did a great job with cooling down the card, especially given the fact that the GPU is only clocked 25MHz less than the desktop part. 25MHz less for 9C difference in FurMark is not an easy task to accomplish.

By using Prolimatech MK-13 with two 120mm Scythe S-Flex fans, the temperature dipped to 74C in FurMark and only 62C in MachStudio Pro. The interesting bit was that MK-13 with S-Flex fans only reaches 24C at idle – meaning that the GPU warmed by only two degrees when compared to the air temperature. Given the environment [closed case, 4.0GHz CPU sitting right next to it], we expected much more.

Given the performance, and the fact that there were two 180mm fans blowing straight onto the graphics card, we removed the fans and left the board passively cooled. The temperature rose to 85C in FurMark and 79C in MachStudio Pro, with idle temperature being 41C.

Prolimatech MK-13 and Quadro FX 4800: Who needs fans, part deux
Given that the fan was broken, we did not measure the Quadro CX, but rather tested the FX 4800 board with stock cooling and later, with the MK-13 heatsink.

By using default cooling, Quadro FX 4800 is not exactly a "hot board". The card idles at only 30C, with Furmark going up to 74C, but not a single degree more. Then again, it’s not that 0.6/1.2GHz clock [compare that to 0.64/1.47GHz for stock GTX 285, or EVGA’s 0.7/1.58GHz] can heat up the heatsink that was built to cool down a GTX 285. Temperature in MachStudio Pro only reached 65C, but given that FX 4800 doesn’t support Tessellation we’re not surprised at such a difference. AMD FirePro had to handle the Tessellation aspect.

Going from stock cooling to MK-13+Scythe, we experienced an idle temperature of only 24C, identical to one achieved by FirePro V8800. FX 4800 survived FurMark heating to only 62C, i.e. 12 degrees less than with stock cooling, while MachStudio Pro Render Test reached 56C. As you might expect, we removed the two Scythe fans and checked how the board performs while living on a subtle breeze from the lower fans only.

The results? 29C in idle mode, i.e. one degree less than with stock cooling, and five more than with fans. When the difference is as such, you might ask yourself why even bother and lose time with putting those fans on? The confirmation came with our full load test. Furmark? 70 degrees Celsius. MachStudio Pro? 62 degrees Celsius.

Thus, there was no doubt in our minds – after testing, V8800 kept the MK-13 heatsink and now, several weeks after testing has completed – the board still happily runs MK-13 with no additional fans.

Given that we only received one Prolimatech MK-13, and after seeing the test results achieved on both boards, we went out and purchased additional MK-13 heatsinks to cool down our FirePro V7800 and Quadro FX 4800 boards, and now have four passively cooled professional GPUs.

We’d say Prolimatech actually experienced direct benefit, as all of our future purchases will now include a requirement for ultra-silent workstation – passive cooling for the GPU and mandatory liquid cooling [or passive cooling] for the CPU. Unfortunately, no company manufactures liquid blocks for the Quadro and FirePro, and we think its a missed opportunity.

Now, Prolimatech MK-13 isn’t without sin. First and foremost, it is absolutely unacceptable that the company used low-quality TIM on DRAM heatsinks, two of which actually fell from the V8800. Interestingly enough, Quadro CX and FX 4800 did not experience such a drop. The only difference between Quadro and FirePro is the memory type [GDDR3 on Quadro and GDDR5 on FirePro], thus we feel it might be that FirePro isn’t warming the memory enough and that the bond simply gives away after a while. Just in case, we replaced the DRAM heatsink bond material with an after-market purchase.

Value for Money distributes the Prolimatech MK-13 with two Scythe S-Flex fans for 69.90 Euro, including German VAT. If you want the passive version, the recommended price for Europe is 49.90 Euro [again, including German VAT]. We have used fans inside the SilverStone case and additionally lowered the noise coming from the system but in reality, if you want to cool down professional GPUs such as FirePro and Quadro, it adds an unnecessary cost. Cooling down the overclocked consumer GPUs is all together – a story for another article. Getting silence and better-than-stock cooling performance on $1,200-1,800 parts for 50 Euro is nothing, compared to the peace of mind and the fact that no longer you hear complaints from designers and engineers that they require "graveyard silence".

One of the major concerns for any business is the "Void your warranty" factor. The R&D studio runs on hundreds of thousands of dollars in equipment and software, but at one point, it stops being a concern. $1000 GPU in a $10,000 workstation with $15,000 in software is a little concern if you know that psychological factor of keeping engineers and designers is actually in
the "high productivity mode".
In the end, we can conclude that Prolimatech made a big step forward when it comes to delivering excellent cooling performance even in passive mode. The only reason why we cannot give an award for the part is the fact that thermal adhesive "let go" on a $1,100 card is a show stopper. We have contacted the manufacturer and we expect that the new versions of the product come with better adhesive for the DRAM heatsinks.

Based on this experience, we’re now making a case study on the impact of GPUs in a traditional working environment and how GPU increase productivity, as well as impact of modded GPUs. This is a field that nobody covered before, and you can be sure we’ll publish our findings when we finish with the research.