Previous | Next --- Slide 20 of 65
Back to Lecture Thumbnails
LilWaynesFather

For those that are interested in what current high-end GPUs are composed of, the NVIDIA GTX Titan GPU has 2688 cores and about three times as many transistors as a top of the line i7. Also not surprising, the actual graphics card is really a giant heatsink.

Image: http://cdn3.wccftech.com/wp-content/uploads/2013/02/er_photo_184846_52.png

kayvonf

@LilWaynes. Question: How do you think the "2688 CUDA cores" that NVIDIA describes relate to the "cores" in this diagram, and to the SIMD ALUs we discuss in this lecture? Something seems a bit fishy doesn't it? Hint: read this.

LilWaynesFather

@kayvonf After a bit of reading it seems that CUDA cores are more comparible to ALUs rather than a true CPU core, although I'm hesitant to say that they are hardware-wise anything related. Nvidia seems to describe them as parallel shading calculation units, so could it possibly be described as multiple (or one) giant core with thousands of mini "ALU-ish" calculation units?

kayvonf

Yes, that is correct. NVIDIA uses the term "CUDA core" to refer to what amounts to an execution unit responsible for "one lane" of a vector operation. It's a more complicated than that if we want to get into micro-architectural details, but it is perfectly reasonable to equate one lane of SIMD unit in an Intel CPU with a "CUDA core". If you're thinking, why does NVIDIA say "core" to mean one thing when the rest of the industry uses the term to mean something else... the answer: good marketing. Larger numbers sound better.

By NVIDIA's definition, the GTX 480 shown here has 15*32 = 480 CUDA cores. Just count the yellow boxes.