Tuesday, November 11th 2014
NVIDIA GeForce GTX 960 to Retain Memory Bus from GTX 970
Among other things like CUDA core and TMU counts, NVIDIA was expected to give its next mid-range graphics card, the GeForce GTX 960, a narrower memory bus, with 3 GB of memory, if not less. A sample sniffed out by India's overly transparent customs department, en route testing facilities in the country, reveal that it's not the case.
The GeForce GTX 960, according to description given in the shipping manifest of the sample, features 4 GB of memory, with a full 256-bit wide memory interface. It also reveals clock speeds to be in the neighborhood of 993 MHz core, with 6.00 GHz memory (GDDR5-effective). It doesn't, however, confirm that the GTX 960 is based on a cut-down GM204 silicon. This could still be different chip, the so-called GM206, which succeeds the GK106.
The GeForce GTX 960, according to description given in the shipping manifest of the sample, features 4 GB of memory, with a full 256-bit wide memory interface. It also reveals clock speeds to be in the neighborhood of 993 MHz core, with 6.00 GHz memory (GDDR5-effective). It doesn't, however, confirm that the GTX 960 is based on a cut-down GM204 silicon. This could still be different chip, the so-called GM206, which succeeds the GK106.
22 Comments on NVIDIA GeForce GTX 960 to Retain Memory Bus from GTX 970
$250 price range cards have almost always been the sweet spot for price/performance ratio and usually gets the most sales.
Even though you get huge value for money with the 970, the 960 will still be of interest to a lot of people.
This is from the end of September. Maybe these will be out before the end of the year.
I was hoping that the GTX 960 would be under 100 watts, but with 4GB of VRAM, I doubt it. I'll predict: 25 watts less than the 970.:p
AMD using just 2Gb for the 285 got a bunch of prejudicial comments, even if it showed their compression technology did make 3Gb not necessary for bulk of games. Although, I couldn't imagine whatever a Tonga XT shows as it will pack 4Gb.
The current implementation of GPU memory compression does not allow more data to fit in the same memory space; a 2GB GPU with compression will still only allow 2GB of uncompressed data to fit within VRAM. This is because you can't know the compression ratio beforehand, so you have to assume the worst case scenario - that all data is not compressible - so that you do not unexpectedly run out of VRAM and have to begin flushing to system memory.
What the compression does do is allow the same data to be stored in fewer bytes in the VRAM, increasing effective memory bandwidth. For example, if you can compress a 2MB texture to 1.5MB before storing it in memory, that is functionally equivalent to a 33% increase in memory bandwidth.
An application that is short on VRAM and is overflowing to system memory will not benefit from memory compression. There is still the same incentive as ever to put more VRAM on a card if the application can use it. The problem with benchmarking GPUs is that memory bus width and memory capacity are correlated (you need more chips to populate a wider bus) so it's easy to confuse additional performance as being related to memory capacity as opposed to the increased bandwidth a wider bus offers.
When you see benchmarks where a 2GB Tonga performs the same as a 3GB Tahiti, there are two factors involved. The card is achieving a high memory compression ratio compensating for the 33% decrease memory bus width, and more importantly the application never needed more than 2GB in the first place. An application that needs 3GB of memory will show an improvement on a card with 3GB of VRAM no matter how well compression works on a 2GB card.