World's biggest water-cooled graphics card tested

LAST WEEK WE HAD Peltier cooling the GF8800 world - this week, we go water cooling, and with a truly massive example at that.

BFG tailor-made a water block for the 8800GTX cards. Its Claudia Schiffer-slim, sleek, water-cooled GF8800GTX provides us with 1U overall thickness with good performance.

The only problem is power. To an average of 200Watts TDP per card (400W in SLI, therefore) let's add, say, 160 WTDP for an overclocked 3.33GHz quad-core Kentsfield chip. If we add ~80W for the extra-hot Nforce 680i North Bridge, South Bridge and two VRM blocks, you'll need a pump, radiator block and water reservoir (or, more precisely, swimming pool or at least jacuzzi tub) to efficiently dissipate up to 700W of theoretical peak heat at any time.

With this arrangement, cooled with a single water source, one goes, it all goes, especially if running the system on a hot day in Singapore.

Asus took a different approach with its EN8800GTX Aquatank. Its engineers took the standard card, replaced the basic cooling with a pretty hefty GPU water block with extra memory heat spreaders, then added dedicated fan to keep that cool. That all takes up close to a two slots in the motherboard

Then, you got two black tubes leading to a separate reservoir/pump/ radiator/fan block, full of copper and green liquid, which by itself takes another 1.5 slots or so (which effectively means another two slot spaces). The kit, knocked together by Thermaltake, is physically the largest among all the graphics cards we've ever seen.

So, in total, you got to spare nearly four slots for a single graphics card - even before thinking of SLI, we barely had space for a Sound Blaster X-Fi card in there. Of course, the card supports SLI, but simply, unless you have a casing with, like, 10 slots of space provided, or stack things on top of each other using the fourth spatial dimension, you can forget about SLI in this case - the laws of physics stand in your way.

Oh yes, not to forget, besides the usual twin 6-pin power connectors for the GPU, both the extra GPU fan and the water block have their own added 4-pin HDD-type power connectors, which of course you got to feed if you want the card running. Overall, this is one big, hungry baby.

What about its performance numbers? Do they justify the extra power & case space required? I was itching for a direct comparison against the last week's contender, Sparkle's Calibre P880+: while that card comes pre-set with 630MHz GPU and 980MHz DDR (x2) clocks, Asus is set at the same 630MHz GPU, but 1030MHz DDR (x2) rates, or GDDR3-2060 effective memory speed.

I used the card in the very same quad-core Intel Xeon 3220 system running at 3.333GHz CPU and FSB1667 (the rest of the specs are in the last week's story). Everything was identical, including the 3DMark 06 runs.

Benchmark	Asus UXGA	Calibre UXGA	Asus QWUXGA	Calibre QWUXGA
Clock	657 / 1049	655 / 1034	647 / 1049	650 / 1031
3DMark06
Overall	12001	11553	8438	8119
SM2	4827	4632	3398	3249
SM3	4681	4495	2979	2858
CPU	5002	5001	5000	5002
GT1	39.6	37.5	26.5	25.2
GT2	40.9	39.7	30.1	28.9
HDR1	43.9	42.0	28.9	27.7
HDR2	49.8	47.9	30.7	29.5
Note: Overall at same speed
Asus 630 / 1030	11604	?	?	?
Calibre 630 / 1030	11186	?	?	?

Oh, surprise! At the identical GPU & memory clock rates and settings (both at GPU 630 and memory 1030), Asus had around 3% - 4% faster 3DMark numbers compared to the month-older Calibre, across several runs. Since it is the same PCB, either Asus used some kind of newer, improved G80 stepping, or maybe the RAM chips used were with a lower latency. Anyone's got a better explanation?

More importantly, I could overclock the card somewhat more on both CPU and memory front - the absolute limit, in the pre-defined Ntune increments, was 657 MHz GPU and 1064 MHz (GDDR3-2128) memory. Interestingly, even after three runs, the 3DMark06 score at 657 MHz GPU and 1049 MHz (GDDR3-2098) timings was a tiny bit faster, crossing the 12,000 3DMark06 UXGA mark for the first time in a single-GPU system! So, I guess, the additional RAM latency does kick in after crossing the GDDR3-2100 line.

Noise-wise, this humongous setup was practically silent - there must be some use for the bucket of green water and a brick of copper fins, after all. In summary, it is the fastest and, definitely, bulkiest single-card 3-D PC GPU setup I know of, excellent for those who want the ultimate without going to (or, afterwards, able to fit in) SLI.

Now, why not power up that reservoir a bit more, and provide dual pipe sets to connect two GPUs? Then we could talk about that SLI, again - maybe with the 8900GTX? ?

See Also
Ultra overclocking pushes Peltier-cooled graphics beyond limits