Manufacturing Process size: what does it mean?

[Amywalker730]

There is a lot of talk and rumors that make grand claims that Nvidia's Pascal will be x10 faster than the current Maxwell and chief among the reasons seems to be the anticipated 16nm manufacturing process.

At first I thought that this would mean a large jump in performance. Then I looked into the history a bit and compared the GTX 580 and the GTX 680, 40nm vs 28nm. In terms of frames per second there was a 24% boost. Yet when I looked at benchmarks for the 780 and 980, both of which used 28nm, there was still a 25% boost.

Finally I looked at the 480 and 580 jump and found a boost of just 14%, despite both of them having a 40nm manufacturing process.

Each generation seems to enhance performance, and despite rhetoric and marketing, this never seems to change the in-practice performance improvements. So what exactly does the 14nm manufacturing process and 8GB if HBM 2.0 mean in terms of FPS gain?

Is there any reasonable argument to not go ahead and upgrade my 680 with a 980TI?

 

[Shaun o]

In lithography the nm scale relates to how many transistors you can create and form in the die area of the silicone substrate used to make a cpu or a Gpu die.

The more transistors you can pack into that space, the density the more powerful the cpu or Gpu can be given the process in nm that can be imaged to the silicone.

Just like a negative of say a old picture camera, printing of the cpu or the layout of the transistors and electronic pathways.
Are stenciled on to the silicone substrate in layers using different chemicals to build up the transistors and pathways.

The optics to do this and focus the stencil image during each chemical stage of building the cpu or Gpu die and it`s capability to project a clean and defined image is where you get the value in nm processing.

The more transistors packed into a space the more powerful.
High Bandwidth Memory.
The closer you have the available memory the faster it will be.
Because HBM for example is part of the GPU die on for example current 390X Ati graphics cards.
The gpu has direct communication to the memory, it does not have to send a request for example from the Gpu Via the cpu of the system through the Pci-e bus. There for the memory and it`s bus width can be wider.
The wider a Data bus is in bits, means you can process from memory a larger block of memory either reading or writing data to it in one clock cycle of the memory.
This is the other way to increase how fast memory works other than the rated clock speed in Mhz that for example are stated of the memory of a graphics card.

512 bit wide data bus = 1 clock cycle of the memory.
256 bit wide data bus= 2 clock cycles = less data processed in a set time = slower memory performance.

 

[utroz]

All manufactures tend to over state the boost from their next gen cards.. I would be more than happy if they can double performance from the latest gen 28nm to first gen 16nm. Most likely the first 16nm cards will only be around a 50% increase in performance but with much lower power usage, once 16nm matures then the GPU dies can get way bigger without horrible yields so by then end of 16nm I expect at least double the performance if not 3X performance with around the same power usage as current top end 28nm cards.. As far as HBM 2 it will allow really fast VRAM in a smaller area allowing more card like the Fury Nano.. As far as waiting, if you can wait 6 months for new cards that is cool, if not get what you want now. There will always be the next GPU just around the corner no matter how long you wait..