NVIDIA Quietly Releases OEM GeForce GTX 660
by Ryan Smith on August 22, 2012 8:30 AM ESTAs is typically the case for NVIDIA when it comes to OEM products, they have once again quietly released their newest OEM video card. Their latest OEM addition is the GeForce GTX 660 OEM, which comes hot on the heels of last week’s launch of the retail GeForce GTX 660 Ti.
GTX 680 | GTX 670 | GTX 660 Ti | GTX 660 OEM | |
Stream Processors | 1536 | 1344 | 1344 | 1152 |
Texture Units | 128 | 112 | 112 | 96 |
ROPs | 32 | 32 | 24 | 24 |
Core Clock | 1006MHz | 915MHz | 915MHz | 823MHz |
Boost Clock | 1058MHz | 980MHz | 980MHz | 888MHz |
Memory Clock | 6.008GHz GDDR5 | 6.008GHz GDDR5 | 6.008GHz GDDR5 | 5.8GHz GDDR5 |
Memory Bus Width | 256-bit | 256-bit | 192-bit | 192-bit |
VRAM | 2GB | 2GB | 2GB | 1.5/3GB |
FP64 | 1/24 FP32 | 1/24 FP32 | 1/24 FP32 | 1/24 FP32 |
TDP | 195W | 170W | 150W | 130W |
Transistor Count | 3.5B | 3.5B | 3.5B | 3.5B |
Manufacturing Process | TSMC 28nm | TSMC 28nm | TSMC 28nm | TSMC 28nm |
Launch Price | $499 | $399 | $299 | N/A |
The GTX 660 OEM is another GK104 based video card, and like NVIDIA’s other OEM Kepler parts it’s a fairly conservative configuration. NVIDIA is shipping this card with only 6 of 8 SMXes enabled, the first time we’ve seen a desktop GK104 part with fewer than 7 SMXes. This further reduction in SMXes brings the GTX 660 down to 1152 CUDA Cores and 96 texture units, while on the raster side of things it’s unknown whether NVIDIA has disabled a whole GPC, or if they’re disabling SMXes in two separate GPCs. Meanwhile like the retail GTX 660 Ti this part has also had a ROP/L2/memory cluster disabled, giving it the same combination of 24 ROPs, 384KB of L2 cache, and a 192bit memory bus.
Looking at its specs NVIDIA seems to be particularly interested in getting a sub-150W Kepler card out – the GTX 660 OEM is rated for 130W and only requires 1 PCIe power connector – so compared to the other desktop GK104 parts the clockspeeds have also taken a hit. The GTX 660 OEM is clocked at just 823MHz core with an 888MHz boost clock, which is about 10% lower than the GTX 660 Ti. The memory clock is also a hair lower at 5.8GHz, an odd configuration since it means NVIDIA still has to equip the card with 6GHz GDDR5. At the same time NVIDIA is equipping the GTX 660 OEM with a fully symmetrical 1.5GB or 3GB of RAM, which coming from the asymmetrical 2GB GTX 660 Ti is probably a combination of cost-cutting and recognition of the fact that the OEM market isn’t quite as finicky about memory capacity as the retail market is.
Taken altogether, this puts the theoretical compute/rendering performance of the GTX 660 OEM at around 75% of the performance of the GTX 660 Ti, with the wildcard once again being the impact of memory bandwidth, which is almost unchanged. This is a larger step between cards than what we saw in the past generation of products (e.g. GTX 560 vs GTX 560 Ti), but at the same time NVIDIA’s OEM products are usually underspeced compared to their retail counterparts. For that same reason however we’d caution against looking too hard at the GTX 660 OEM for a sign of what the eventual retail GTX 660 will be like. NVIDIA’s been known to use different clockspeeds, different core configurations, and even different GPUs entirely, so everything is still on the table for now.
Source: Fudzilla
36 Comments
View All Comments
CeriseCogburn - Saturday, August 25, 2012 - link
The 660Ti has been shown here to beat the 7970 sometimes and to have exactly zero performance hit - but we can all thank the reviewer for making his DISPROVEN speculation foremost in your mind by mentioning it almost every page and then commenting how surprising it was when it didn't show up.More fuel for idiot amd fanboys, exactly what the pr doctor ordered.
RickLaRose - Saturday, August 25, 2012 - link
Oh brother...Beat the Radeon HD 7970 in what?
You do realize that with new drivers the Radeon HD 7970 Ghz Edition is the single fastest GPU overall right? and that the Radeon HD 7970 gives the GTX 680 a run for its money with the new driver + lower price point?
You do realize these things right?
And we're still talking video games here. If we look towards GPGPU tasks it's not even close. Kepler is a horrible GPGPU architecture. It is actually a step backwards from the GTX 580 in terms of Computational Performance on both Single and Double precision workloads.
Unless all you do is gaming or there is some specific nVIDIA feature you require and are only willing to purchase a single card (Other than PhysX... no other useful features really... maybe CUDA but everyone uses OpenCL now anyway) or you have some aversion to AMD drivers (odd considering nVIDIA has been having more issues for the last year/year and a half in that dept) then there is no real compelling reason to purchase an nVIDIA card.
I'm just being honest. Some people care about GPGPU performance.. you really ought to consider that before you post non-coherent garbles of text that make you appear as though you're suffering from some sever mentally illness.
CeriseCogburn - Monday, August 27, 2012 - link
1. Go read the 660TI charts to find out mr clueless. Oh, forgot , you're being honest, and therefore immense ignorance is required.2. compute / amd = winzip
3. after that nVidia wins everything
4. Go look at a real compute amd cpu, they need software, the release announcement/ review is up
5. If you disagree tell the author of that article you haven't and likely won't read, heck, you didn't even look at the 660Ti charts here in it's release review
6. The above proves you're a loon, with a big mouth.
Midwayman - Friday, September 7, 2012 - link
Sure it beats the 7970 on a couple test, but I wouldn't venture to say the 660ti review puts it anywhere near the 7970 on average. Its much closer to the 7870/7950.CeriseCogburn - Monday, August 27, 2012 - link
http://www.anandtech.com/show/6191/amd-announces-f..." The real differentiator may come down to software, with AMD having invested virtually everything into OpenCL.( LOL - THAT MEANS ALMOST NOTHING SINCE THEY ARE BROKE) Success here means that AMD needs to continue turning developers away from CUDA and towards OpenCL, as they can’t sell hardware to developers until developers can run their projects on AMD’s hardware in the first place." ( And cannot offer a 10th of the support nVidia does, they will fail, and whomever falls for the lure ( poor open source neckbeards who dial , email and complain a lot LOL ) will get burned.
Looks like amd will continue to lose money and then go bankrupt.
Nvidia is ahead in OpenCL anyway. ROFL
Jim746 - Sunday, February 10, 2013 - link
I purchased a retail version to see if it was faster and it was actually slower, so I returned it.