And so did quite a number of its card vendors. Most of these like BFG, Asus were water-cooled, but just one was still using the Peltier thermoelectric cooler - you guessed it right, it was Sparkle again with its Calibre card.
While water-cooled cards had default factory settings of 630 MHz GPU, and 1030 MHz (DDR-2060) memory, the Calibre P880+ had the same GPU speed, but slightly lower memory speed setting, 980 MHz (DDR-1960). Hmm, let's see if this spec difference led to any real-world performance drop.
From the outside, the first difference you notice is the new boxes' stark black, compared to the creamy white of its immediate non-overclocked predecessor - the new colour suits the "enthusiast" market better. The Peltier cooler on the card is now black as well, as you can see - there aren't any other external changes. Except, oh yes, there is now a large (black, of course) heat spreader for the twelve memory chips and NVIO ASIC this time.
I ran the card in the reference test system, based on the desktop quad-core Intel Xeon X3320 (the real top-grade Kentsfield), running on the Asus Striker Extreme Nforce 680i board at 3333 GHz CPU / FSB 1667, with the usual 2 GB of Corsair XMS6400CL3 RAM at DDR2-833 CL3-3-3-5 latency. WindowsXP Pro SP2 with ForceWare 97.92 driver and Futuremark 3DMark06 did the job for comparative testing.
Upon the boot, Nvidia Monitor confirmed the 630 GPU / 900x2 memory speed settings - the card completed the 3DMark runs on both 1600x1200 UXGA on the 20-inch Philips and 2560x1600 WQXGA resolutions on Dell's 30-incher. Interestingly, the card temperature was two degrees Celsius higher (up to 65 C) when running the higher resolution test run.
Not surprisingly, this difference applied in the overclocking limits, too. At the UXGA level, the card could (3 times repeatedly) complete the run at 654 MHz GPU testing and 2 x 1034 MHz (2068 DDR) memory. However, at the WQXGA resolution and double the total pixels, the maximum clocks at which the 3Dmarks completed flawlessly were 650 MHz GPU and 2 x 1031 MHz (2062 DDR) memory - could it be due to the overheated NVIO chip this time? And yes, this time I ran the system in an air-conditioned setting.
But in general, the Peltier's advantage was that the difference between idle, post-full-load and full-load operational temperature was six and fifteen degrees C respectively most - on the normal 8800GTX cooler, this can be as much as 13 and 30 C respectively, and on a water cooler, we're still talking about 8 and 20 C respectively.
Here are the results:
| Calibre P880+ |
Std
UXGA |
OC
UXGA |
Std QWUXGA
|
OC
QWUXGA |
|
| Clock |
630 / 980
|
655 / 1034
|
630 / 980
|
650 / 1031
|
|
| 3DMark06 | |||||
| Overall |
11109
|
11553
|
7795
|
8119
|
|
| SM2 |
4427
|
4632
|
3105
|
3249
|
|
| SM3 |
4290
|
4495
|
2734
|
2858
|
|
| CPU |
4997
|
5001
|
5000
|
5002
|
|
| GT1 |
35.9
|
37.5
|
24.0
|
25.2
|
|
| GT2 |
37.9
|
39.7
|
27.7
|
28.9
|
|
| HDR1 |
40.1
|
42.0
|
26.5
|
27.7
|
|
| HDR2 |
45.7
|
47.9
|
28.1
|
29.5
|
|
In summary, by going from the standard to OC edition of GeForce 8800GTX, you get some 10% extra 3DMark performance, which is expected as both GPU and memory clocks are pushed roughly by that amount. If you push it to its limits, at least in this Calibre P880+ case, you get an additional 4%, consistently on both UXGA and WQXGA resolutions. But there is no gain in the 3DMark CPU scores, they still stay around 5,000 on my system - which is pretty good on its own, anyway. Most importantly, even at the 2560x1600 WQXGA level, the game frame rates never drop below 25 fps, which in my observations is the comfort-level margin for games playability - sorry I'm not a fast-paced teen anymore, so 25 FPS does it for me.
I also ran a test where GPU was overclocked at the same speed, but memory ran at 'only' 2 x 1010 MHz . Guess what, there was near-zero result difference - 11538 3DMark06 vs 11553 3DMark06, or 0.1% total speed loss for 2% memory clock loss. Therefore, at ~GDDR3-2000 speeds, the GeForce8800GTX memory requirement is already maxed out - slight RAM speed changes, up or down, dont' affect the results.
So, Calibre's Peltier does a very good job in getting the G80 GPU to go quite a bit beyond even its OC settings - 650 MHz is not bad at all. Its memory cooling performance isn't yet up to par with the water cooled siblings. My belief is that Sparkle could have done better there, as the simple heat spreader used doesn't even have full cooling fins to maximise the heat dissipation area for the RAMs or NVIO. On the other hand, well, GDDR3-2060 working performance is more than good enough for the overclocked G80 anyway - the benchmarks don't improve with the memory speed bumps. ยต