David tells us that Nvidia architecture has been quite complex for a long time, especially with fifos and logic to memory transactions, since know that memory is one of the key bottlenecks in graphic cards.
When the good old TNT chip was introduced, Nvidia talked about their patented "Constant Velocity FIFO" which is a set of fifos and buffers that even out "bursty" accesses to memory. This gives a good throughput to the graphic core without peaks and troughs.
Kirk tells us the idea of the Constant Velocity FIFO - CVF -- is to isolate the texturing and shading part of the pipeline from the latency of the read/modify/write operations involved in stencil, Z-buffering, and alpha-blending.
This technology survived to this day and it has evolved into what we call Lightspeed Memory system that will be part of Nvidia's latest offering.
David said that we could not call it "megabytes of memory" we can see that his point is that all the calculations performed on the chip are made by their memory architecture.
So, either, as we suspected, there is some buffer memory, or the CVF is so effective that it can easily process megabytes of information inside the chip with no need for memory to GPU transfer.
There will be more on this, we reckon. ยต