GP108 also sports 16 ROP units to GK208’s eight. Meanwhile, GK208 employs two larger SMXes, each with 192 CUDA cores, eight functional texture units, 64KB of shared memory and L1 cache, and a separate texture cache. Again, each SM includes 128 CUDA cores, eight texture units, 24KB of L1/texture cache, and 64KB of shared address space. GP108 utilizes a single Graphics Processing Cluster with three Streaming Multiprocessors. So why is GP108 so much more complex of a GPU than GK208? The two GPUs do span a couple of different architectures, so they’re organized differently, for starters. GeForce GTX 750 Ti is a 60W card, and we haven’t seen evidence that it can be passively cooled in a low-profile form factor. Also, GT 730’s 38W lands a lot closer to GT 1030’s 30W specification. Spec-wise, the Kepler-class GeForce GT 730 is a much closer match with 384 CUDA and 16 texture units across two SMXes, eight ROPs, 512KB of L2 cache, and a 64-bit interface. In reality, GeForce GT 1030 is the spiritual successor of GT 730 since Nvidia never created a 900-series product below the $160 GTX 950. We’re only fascinated by the comparison to GT 1030 because of their similar transistor counts.
GeForce GT 730: The Real Competitionīut remember that GTX 750 Ti launched as a $150 card, and even now sells for $100 and up. Of course, a 64-bit aggregate memory bus cripples GT 1030’s peak bandwidth to 48 GB/s using 6 Gb/s GDDR5 GTX 750 Ti’s wider interface facilitates up to 86.4 GB/s. In contrast, GeForce GTX 750 Ti starts at 1020 MHz and boosts just slightly to 1085 MHz. Our sample employs a 1227 MHz base frequency and a typical GPU Boost rating of 1468 MHz. Nvidia goes a long way to overcoming those deficits in GT 1030 with higher clock rates. That’s a lot of lost resources for a ~4% difference in complexity. And whereas GeForce GTX 750 Ti utilizes two 64-bit memory controllers, GT 1030’s specs break the memory bus down into a pair of 32-bit controllers, adding up to a 64-bit interface. That means GeForce GT 1030 includes 512KB L2 total-a big reduction from GTX 750 Ti’s 2MB. However, those partitions are aligned with 256KB slices of L2 cache on GP108 and 1MB slices of L2 on GM107. The two GPUs feature a pair of ROP partitions, giving you up to 16 32-bit integer pixels per clock. Both designs also expose eight texture units per SM, totaling 24 on GeForce GT 1030, while GTX 750 Ti gets 40. Given 128 CUDA cores per SM/SMM in the Pascal and Maxwell architectures, that’s 384 cores for GT 1030 and 640 for GTX 750 Ti. Here’s the thing, though: whereas GeForce GTX 750 Ti employs five Streaming Multiprocessors, GT 1030 comes equipped with three. Or how about the GeForce GTX 750 Ti, which we’re making the GT 1030 battle in today’s benchmarks? That card’s GM107 GPU has a similar transistor count as GP108, but in a 148mm² die, owing to its 28nm manufacturing process.
Compare that to GeForce GT 730’s GK208 chip with 1.02 billion transistors in an 84mm² die. It’s a teeny thing at just 70mm², thanks to the same 14nm FinFET process used to manufacture GP107. GeForce GT 1030 utilizes an all-new graphics processor called GP108, composed of 1.8 billion transistors.
0 kommentar(er)
