OpenCL (and vulkan-compute) are for the programmable units on a gpu. This is possibly not what they're referring to when they talk about accelerators (especially if comparing to cuda)
Aside from typical arbitrary computation stream processors (still can't believe industry called them shaders, even in the graphical context fragment programs made more sense, I blame microsoft) nvidia have fixed function blocks. Being fixed function they are a lot more efficient per power and silicon area, so you might have a function that multiplies two 8x8 matrices for example.
The fixed functions change depending on the hardware, and there has to be a means of handling it, an abstraction layer. CUDA handles this.
Nvidia went all in with dedicated functions with the 3000 series cards, amd still had general computing with 'shaders' with the 6800 line. Some fixed function hardware was added in the 7000 line just to compete.
I'm not that much of a fan of fixed functions with deprecation rather than programmable units.. but you can see there's performance benefit to be had if you know those fixed functions will be used in specific ways.