ZOTAC GeForce GTX 750 Ti OC Edition Graphics Card Review
ZOTAC GeForce GTX 750 Ti OC Graphics Card Launch Review
When NVIDIA lauched their GTX 780 Ti a few months ago it cemented their place at the top of GPU performance charts. The GTX 780 was competitive with AMDs high end cards, the GTX Titan was faster but somewhat priced oddly due to its developer focus so the 780 Ti brought the two words together, an unlocked Titan GPU with a more desirable price point than Titan, ideal for enthusiast gamers.
With the GTX 780 Ti being based on the same GPU as Titan and the family re-using parts from previous "generations" such as the 770/680 change it has been some time since NVIDIA has released a "new" GPU. That changes today with the launch of the GTX 750 and GTX 750 Ti, both based on the Maxwell GPU. We have an overclocked version of the Ti on our test bench today and will throw some of the latest games at it, including Battlefield 4, DOTA 2 and Assassins Creed 4: Black Flag.
The NVIDIA GTX 750 Ti 2GB
Before we take a look at the Zotac version of the GTX 750 Ti, it is worth taking a quick glance over the NVIDIA reference design as well as a few details on the New NVIDIA GPU.
So, what the card tells us is that this is a product capable of running on a compact PCB, cooling the GPU is not something that will require a massive/loud heatsink and fan combo and in fact, the card will run from power over PCIe only… no external power required on this 60w TDP product.
Looking at the GPU a little more, this is NVIDIAs first generation Maxwell part with the code name GM107. The core design goal for Maxwell was to create a GPU which excelled on power efficiency and performance per watt. As noted before the GTX 750 Ti is a 60w TDP part which NVIDIA feel will be ideal for the notebook and small form factor PC market (such as Steam machines). Higher performance, enthusiast parts, based on 2nd Generation Maxwell are to be released at a later date.
60w TDP is a low rating, even for a mainstream part, and while NVIDIA are being reasonably quiet on the details of how they have achieved this they do note that the Streaming Multiprocessor (SM) design used on Kepler set them down a path towards this GPU, improving as they went on control logic partitionoing, workload balancing, clock-gating granularity, complier based scheduling and instuctions per clock. With this in mind the new SM architecture allows NVIDIA to use five SM’s on Maxwell with only a 25% bigger die area than that found in Kepler with its 2 SM’s. L2 cache also receives a boost (2048KB, vs. 256KB in GK107) with the end result being fewer requests between GPU and memory.
Here is a little more about Maxwell, direct from NVIDIA:
Dropping down into the SM architecture the new design allows 35% more performance per CUDA core on shader limited tasks and to achieve this NVIDIA re-wrote the algorithms and scheduler to minimise stalls/bottlenecks. In addition the layout has also changed with each SM now split into four processing blocks with their own buffer, scheduler and CUDA cores. During this re-configuration NVIDIA also removed the non-power-of-two CUDA core sharing with the new partitioning simplifying design and scheduling logic to save space, power and latency. Each pair of processing blocks shares four texture filtering units and cache and the L1 cache is combines with texture cache with a separate shared memory unit. So, each SM is smaller which offering 90% of the performance of Kepler SM units. Combine those smaller units and you get 25% more peak texture performance, 1.7x more CUDA cores and 2.3x shader performance.
Finally, there are two other changes of note. Firstly NVIDIA has maximised internal memory system bandwidth and efficiency (in addition to the larger L2 cache). Secondly the hardware based H.264 encoder (NVENC) which powers ShadowPlay (amongst other things) has been enhanced. We get 6-8x real time encode performance compared to 4x in Kepler. Video decode also gets a performance boost (8-10x) and the added local decoder cache as well as higher memory efficiency per stream results in lower power use during video tasks. As a side note Maxwell has an additional power state (GC5) for light workload, such as video playback, for further power savings.