
New shuffle Instructions Īt a low level, GK110 sees additional instructions and operations to further improve performance. With a 48KB space, the texture cache can become a read-only cache for compute workloads. Register file space has increased to 256KB compared to Fermi. The SMX also sees an increase in space for register file. The single biggest change from GK104 is that rather than 8 dedicated FP64 CUDA cores, GK110 has up to 64, giving it 8x the FP64 throughput of a GK104 SMX. With GK110, Nvidia opted to increase compute performance.

Performance in register-starved scenarios is also improved as there are more registers available to each thread. Both the L2 cache and register file bandwidth have also doubled.

As for the L2 cache, GK110 L2 cache space increased by up to 1.5MB, 2x as big as GF110.

At the SMX level, GK110's register file space has increased to 256KB composed of 64K 32bit registers, as compared to Fermi's 32K 32bit registers totaling 128 KB.

With GK110, increases in memory space and bandwidth for both the register file and the L2 cache over previous models, are seen. This model also attempts to maximise energy efficiency through the execution of as many tasks as possible in parallel according to the capabilities of its streaming processors. GK110 was designed and marketed with computational performance in mind.
