Quantcast
 


NVIDIA GeForce GTX 550 Ti Roundup: ASUS, EVGA, Gigabyte & MSI

Author: SKYMTL
Date: March 14, 2011
Product Name: NVIDIA GTX 560 Ti 1GB
Share |

GF106 to GF116; Another Revised Architecture


Unlike past revised Fermi cores, NVIDIA really didnít have to make too many changes to the GF106 architecture in order to bring its features in line with the rest of their product stack. Nonetheless, NVIDIA still decided to revise the GF116ís layout so more of the faster, higher leakage transistors were placed on the critical rendering paths instead of being used for periphery tasks. Meanwhile, the slower low leakage transistors were placed where speed wasnít a primary concern.

Strategically distributing the transistors in this way allows for a small speed-up in overall rendering performance. More importantly it also means the fastest transistors will now be fully utilized instead of being used for non critical tasks and thus lowering overall performance per watt. In addition, some other not so insignificant tweaks were made.


The GF116 core looks very much like that of the GF106 but there are a number of noteworthy additions. The basic layout of four individual SMs bringing with them 192 CUDA cores 16 texture units and a quartet of all-important PolyMorph engines has remained unchanged from the GF106.

Meanwhile, NVIDIA has gone to work expanding other portions of the architecture. The GF106ís remaining ROP partition and 64-bit memory controller were enabled which brought along an additional 128KB of L2 cache. This means GF116-based products will have 24 ROPs, 384KB of L2 cache and a 192-bit memory interface.


With the addition of a 192-bit memory interface, NVIDIA was faced with a bit of a challenge. Usually, memory controllers and drivers are meant to function with balanced memory allotments. Take for example the GTX 560 Tiís layout; it uses four 64-bit memory controllers each of which is paired up with two 128MB GDDR5 modules equaling 256MB per controller and 1GB when all of the modules are combined.

The above-mentioned formula led a version of the GTX 460 having a 192-bit memory interface along with 768MB of memory and the 8800 GS having a 192 / 384MB layout. NVIDIA wanted to avoid reducing the overall memory allotment from the GTS 450ís 1GB to 768MB on the GTX 550 Ti so they implemented an obvious yet innovative solution.

NVIDIA now has a way to allow for mixed memory allotments on a per-channel basis. Since the technology is proprietary and will presumably a closely guarded secret, they declined to discuss the specifics with us. What we do know is that two of the GTX 550 Tiís memory controllers are populated with 256MB of memory (in two 128MB modules) while the other is paired up with 512MB of GDDR5. Presumably, there is some sort of load balancing going on behind the scenes which is facilitated by a slightly revised driver stack but weíre sure that some core changes were implemented as well.

In our opinion, this newfound ability to mix memory sizes is truly a game changer that could have a huge impact upon upcoming NVIDIA products.


With the refined process bringing the ability for increased clock speeds and the addition of an expanded memory interface, the GTX 550Ti performs above and beyond the GTS 450 which is likely why this new card isnít going to replace its predecessor in NVIDIAís lineup. The 192-bit interface in particular brings massively increased memory bandwidth (an area in which the GTS 450 was lacking) which could lead to some dramatic increases in games.
 
 
 

Latest Reviews in Video Cards
July 27, 2014
ASUS' new STRIX is one of the few 6GB GTX 780 cards on the market but it has another trick up its sleeve: with 0db fan technology the STRIX runs completely silent in some situations....
July 7, 2014
NVIDIA's TITAN Z was launched without much fanfare but drew both admiration and derision from gamers due to its steep price tag. But how does it actually perform in games? We find out....
June 25, 2014
PowerColor's Devil 13 seems to do the impossible: take a pair of R9 290X cores, stick them onto a single PCB and air cool them. Have they been successful? We find out....