Nvidia GeForce GTX 660 (OEM) Specs Surface
Nvidia quietly released the GeForce GTX 660 GPU for OEM use. Is this signs of what to come with the Retail Version?
With the recent release of the GeForce GTX 660 Ti, Nvidia quietly released its GeForce GTX 660 for OEM use only. The graphics card bares similar specifications as its bigger brother the GTX 660 Ti, which might be signs of what is to come if a retail version is released. However, take that with a "grain of salt", as Nvidia has in the past changed specifications from OEM version to retail version, such as different clock speeds or CUDA cores.
GeForce GTX 660 (OEM)Taking a look at the GTX 660 OEM specs, we see it's based on the same GK104 silicon as the GTX 660 Ti with two SMX units disabled, resulting in 1,156 CUDA cores and 96 texture units. The GPU core is clocked at 823 MHz (888 MHz GPU Boost), which is about 10 percent lower than the GTX 660 Ti. The memory is clocked at 5.80 Gbps GDDR5, with a 192-bit memory bus width. The GTX 660 OEM offers memory options of 1.5 GB and 3 GB compared to the GTX 660 Ti's 2 GB configuration. With the decrease in specifications, we see that the GTX 660 OEM only requires one PCIe power connector and a maximum power requirement of 130 W. The other specifications are consistent with the GTX 660 Ti and the other members of the Kepler-based GPUs. With its specifications, it should put the performance of the GTX 660 OEM around about 70 to 80 percent of the GTX 660 Ti.
Far from weak. Maybe it's on par with the last generation GTX 580. Anyway, Only one pin PCIe power connector is great!
Far from weak. Maybe it's on par with the last generation GTX 580. Anyway, Only one pin PCIe power connector is great!
If it is OEM, will it be available in the retail at all?
It's not that complicated. 660 Ti > 660 vanilla. If you thought this was bad, you must not have been gaming during the 7000/8000/9000 era, where you'd be trying to choose between the 8800 GS, 8800 GT, 8800 GTS (three different models under one name!), 8800 GTX, and 8800 Ultra. I wouldn't mind if they dropped the Ti suffix and renamed one of the two (e.g. GTX 660 -> 655, 660 Ti -> 660), but the x60 is the only one they're currently using suffixes on.
Oh it was worse than that, recall Nvidia renamed the 8800 GT to 9800 GT to 240 GT with only minor bois revisions.
BestBuy sells NVIDIA brand video cards, but they are the only non-OEM I know of that does that.
What they lack in bus width, they make up for in speed. Have you noticed how fast this new cards' memory clocks are? You can push just as much data through a smaller pipe if you increase the speed accordingly. This saves on board layout and complexity (reducing cost). It's like comparing SATA to PATA. Just because SATA only has 7 connecting pins, doesn't make it slower than a 40 pin IDE setup.
Not saying GK104 or GTX680 is bad. I have a GTX680 and it's a beast!!! I am just hoping to see some competition in the midrange market.
The *poor man's* 680 is the 670. This is the *poor man's* 660 TI.
It's the only reasonable way to differentiate Kepler's performance because the memory bus on the upper cards is so weak that decreasing GPU performance significantly doesn't make much of a difference in gaming performance. Look at the 680, the 670, and the 660 TI for outstanding proof of this. The 670 dropped nothing but GPU performance and did so by about 20% (rough math, I could give more accurate numbers if asked), yet it hardly changed gaming performance at all relative to the 680.
Decreasing the bus width to 192 bit was the only way that Nvidia could make a card that is actually significantly slower than the 680 without decreasing core count more significantly. Even this GTX 660 with what is probably closer to a 35-40% drop in GPU performance (again, rough math, I could give more accurate numbers if asked) would likely perform around the GTX 680 if not for the 192 bit bus. It probably performs very close to the 660 TI, much closer than the 560 is to the 560 TI.
I don't agree with this method of decreasing performance because it makes the card more unbalanced, but it is effective.
SATA is a different technology from PATA/IDE and transfers data differently. It is like going from PCI to PCIe. Dropping the bit width of an interface without changing the technology of it is like taking PATA and decreasing the pin count. Unlike with SATA, it would reduce performance because you're using the same technology, but with inferior, rather than superior, hardware.
Cutting the bus width by 25% and not changing the memory frequency (which is already very high) means a 25% reduction in memory bandwidth. The 670 has more bandwidth than the 660 TI and it shows in performance quite significantly.
I made that comment based on the radeon 6850 card that was mentioned using a 256 bit width. Even with a wider bus width, it only has a throughput of 128Gb/s where as the cut down 192 bit width of the 660ti can still put more through at 144Gb/s. They can do this by using higher clocked memory. I suppose the SATA to PATA reference was not that good because of the underlying protocol differences.
Anyway, the 660ti can be produced/sold cheaper by easily cutting bit width of the memory. It takes up less manufacturing resources and less space on the card, and is probably more effective than cutting more cores from the GPU itself. It's a great card for the price, but if you want the bandwidth too you'll have to pay the extra $100 for a 670.
The 6850 has maybe half the gaming performance of the 660 TI, if even that much! Of course it can have lower memory bandwidth. That a far lower end card has almost as much memory bandwidth shows that the 660 TI might not have enough. That similarly modern cards to the 660 TI have far more memory bandwidth and tests have confirmed that it makes a substantial difference proves that the 660 TI has too little bandwidth. Your examples are irrelevant and your reasoning only shows that the Kepler cards were made purely to maximize profits, not make excellent graphics cards.
The 660 TI is not a great card at all. It is decent and I won't go any farther than that. There are far better cards at lower prices, such as the 7870, and even greater cards for about the same price as the 660 TI and going to be cheaper (7950). The 660 TI's memory bandwidth has been proven to be a huge Achilles's heel when put to the test and its memory is so highly clocked that it can't be pushed substantially higher, so it can't even be helped. TXAA is the only saving grace and unlike the 670 and 680, not even TXAA can help it enough. The 660 TI has a lower mid-ranged memory interface paired with a very high-end GPU that already had poor tessellation efficiency and crap DirectC performance with the GTX 670, giving it even worse AA efficiency.
Using the Radeon 6850 as your example to try to disprove the important of the memory bit width, you ignore the fact that even the 6850's memory bandwidth can be increased substantially through overclocking, granted it doesn't have a GPU that is fast enough to improve performance similarly substantially. Overclocked to the same memory frequency, the 7870 has a 33% memory bandwidth advantage over the 660 TI and the 7950 has a huge 100% advantage.