by Anton Shilov
02/19/2014 | 11:55 PM
Nvidia Corp. has introduced its first graphics solutions based on the code-named Maxwell architecture that was designed to further improve performance per watt. The new GeForce GTX 750-series graphics cards are based on the GM107 graphics processing units and are aimed at the mainstream market.
"We know that to advance performance, we must advance performance per watt, because every system we design for has a power limit -- from supercomputers to PCs to smartphones. That is why we architected Maxwell to be the most efficient GPU architecture ever built," said Scott Herkelman, general manager of the GeForce business unit at Nvidia.
To boost efficiency of the first-generation Maxwell architecture and to boost performance per stream processor/per watt, Nvidia completely redesigned streaming multiprocessor (SM) compared to Kepler architecture. With Kepler, the each SM contained control logic that routed and scheduled traffic for 192 cores. This was complex for a single piece of control logic. With Maxwell, Nvidia divided the SM into four blocks, each with its own piece of control logic. Each block features its own control logic (i.e., instruction buffer, warp scheduler, dispatch unit), register file and 32 stream processors. Each Maxwell SM features its own PolyMorph Engine 2.0 unit (one vertex fetch, one tessellator, one viewport transform, one stream output, etc.).
Improvements to control logic partitioning, workload balancing, clock-gating granularity, compiler-based scheduling, number of instructions issued per clock cycle, and many other enhancements allow the Maxwell SM (also called “SMM”) to far exceed Kepler’s SM (SMX) efficiency. The new Maxwell SM architecture enabled Nvidia to increase the number of SMs to five in GM107, compared to two in GK107, with only a 25% increase in die area.
The GM107, the first GPU based on Maxwell architecture, integrates 5 SMs (640 stream processors), 40 texture units, 16 raster operating units, 2MB L2 cache as well as 128-bit memory controller (divided into two 64-bit controllers). The chip is manufactured using 28nm process technology at TSMC and contains 1.87 billion transistors. By contrast, the GK107 only sports 384 stream processors.
By breaking up one giant problem of the control logic doing the scheduling and threading on each SM into smaller work, and by performing the global partitioning in software, we were able to design the chip to be more efficient. As a result, we’ve actually increased the amount of peak performance per core by 35% when compared to the cores in chips built with the previous generation Kepler architecture.
Since Nvidia uses less overall cores to get more performance, less power is ultimately used, increasing the performance/watt for the Maxwell architecture. A number of additional optimizations were made inside the chip to increase power efficiency.
The GM107 GPU supports all modern high-end features like DirectX 11, OpenGL 4.3, OpenCL 1.2, stereoscopic-3D, multi-monitor capability, PhysX, ShadowPlay, G-Sync, GameStream and so on. Unfortunately, it does not seem that the first-gen Maxwell supports DirectX 11.2.
The GeForce GTX 750 Ti and GTX 750 graphics cards are small, only 5.7” long (14.5cm), and do not require an internal power connector. Since the GTX 750-series graphics boards consume so little power, they run extremely quietly and generate very little heat, making them perfect for use in small form-factor and home theater PCs.
The Nvidia GeForce GTX 750 Ti and GTX 750 graphics boards are now available from add-in card suppliers, including Asustek Computer, Colorful, EVGA, Gainward, Galaxy, Gigabyte, Innovision 3D, MSI, Palit, PNY and Zotac. Pricing is expected to start at $119 for the GTX 750, $139 for the 1GB GTX 750 Ti and $149 for the 2GB GTX 750 Ti. The GeForce GTX 750 Ti and GTX 750 GPUs will also be sold in fully configured systems from leading U.S.-based system builders as well as other system integrators outside North America.
According to Nvidia, the GeForce GTX 750 Ti significantly outperforms its main rival on the market, the AMD Radeon R7 260X which is based on Bonaire XTX graphics chip (896 stream processors, 56 texture units, 16 raster operating units) powered by second-generation GCN architecture.