Volta Sensor Decoding
With Volta GV100, each thread can access memory in the host memory address space directly, and independently of other threads, sharing memory through shared memory buses. This eliminates the need to synchronize thread accesses to global memory, reducing the synchronization overhead of global memory access, enabling more fine-grained synchronization and cooperation between parallel threads in a program, and thus more efficient programs.
Volta Sensor Decoding
Volta GV100 supports a number of new instructions for control and access to shader registers, such as fetch and store through the execution units of an instruction-scheduler, which greatly improve the efficiency of program execution. These new instructions for program control are available at the instruction level and thus do not require any new resources, as an example, 32 registers for 32 integer instructions is one new resource.
Volta also supports a new resource called the Lite Shader Core, a space of 8 x 8 KB of VRAM that is attached to the engine and available to the Graphical Processing Unit to the same extent as the Application Memory. Lite Shader is intended to support the development of Vulkan compliant drivers, as well as game-related use-cases, such as the rendering of static geometry.
NVIDIA has described the Volta GV100 SM as "a powerful combination of 28nm process technology, next-generation process node and advanced FPGA technology. These innovations are all designed to create an extreme scaling process platform for advanced compute hardware design from those who want the best performance to those who value cost."