• Competitor rules

    Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.

Nvidia's GTC 2014 Thread

nIjTgCZ.png
 
Orangey said:
I tuned in when he was halfway done talking about the new form factor, he said something like "3x smaller than PCI-E"? What's the scoop?
Click to expand...

Nothing much, interposers/tsv's are where the next few gens of computing parts are going. When you look at the pci-e card, the massive majority of complexity and size is due to routing memory data. When you get all the memory on package you get rid of 2/3rds of the pinout on the back of the package, you get rid of most of the required pcb, and use less power, with more stability so you can package it all much smaller.

But this is primarily for custom built supercomputers. 3d packaging has been doable for realistically 3-4 years but the cost is so high it's just not feasible. I don't know exactly when it's predicted that everyone is going for it but the chance of it appearing on $1000+ professional cards before $300 consumer cards is pretty high ;)

Ultimately when 3d packaging, stacked mem, interposers are ready they are HUGE leaps forward, to a significantly larger degree than a process node drop. It's just a case of cost. 3d stacked mem is probably the cheapest/easiest, but without being on package it isn't massively useful. Good for density but not bandwidth. With stacked chips you could double or quadruple memory on a "normal" card with a different chip used, but to get higher bandwidth you need many more connections to each set of stacked chips, and that is where interposer/on die stuff comes from. Rather than going off package through bumps and copper traces, you go via silicon made etches that are a magnitude smaller. You can fit say 10 traces at silicon level where you could only fit one at pcb level.

Very interesting things will happen when everything can be connected with on silicon level connections, insane bandwidth, vastly reduced power.

It's not Nvidia specific, nor AMD, or Intel, just cost prohibitive, when it's not everyone will use it. I would be very surprised if the next gen gaming stuff wasn't pci-e and very similar in size. Professional stuff will probably mostly be pci-e but they will do uber expensive versions like they showed on stage. No one seems really sure when it will become mainstream, again because of cost. In general it involves taking a bunch of working dies, sticking them together, and some of those will fail at that stage wasting a whole bunch of stuff, the effective yields drop massively. There will be a cross over point where yields after sticking everything together are high enough that performance beats the cost.

You're talking about going from 512bit memory being highly complex at pcb level, to 5120bit memory on die being more than doable..... /drool.
 
Last edited:
Never want to hear

"The titan is not a gaming GPU" Jen just said the Titan runs 4K all games, with all effects turned on!

Ohhh just announced the Titan Z
 
You must log in or register to reply here.
Back
Top Bottom