Nvidia has released its roadmap for the next few years ahead along with a powerful new GPU architecture. Currently, we are on Kepler with the 6 series, which supports dynamic parallelism and enhances the programmability of the GPU. Their next GPU architecture will be Maxwell and it will be used for the next two GPUs from Nvidia. Maxwell supports unified virtual memory, which allows the GPU and the CPU to see each others memory in orther words “All memory is visible to all the processors.”
Now Volta is where things get interesting. Volta is named after Alessandro Volta, who has been credited with the invention of the battery. In fact the unit Volt is named after him. Volta will have stacked memory in which they will stack the DRAM on the same silicon substrate as the GPU chip. To accomplish this they will add another silicon substrate to the GPU, which will extend to host the Stacked DRAM. The DRAM will then be accessed with through silicon vias, which connects each DRAM that is stacked on one another.
The main result of this new approach is 1 terabyte per second of bandwidth. Nvidia’s CEO compares this to taking a full Blu-Ray disc and in 1/50th of a second move the information on that disc through the entire chip. In comparison, the high end GTX Titan only has around 288 GBps of bandwidth.
-
Nvidia Officially Reveals the GTX 460
12 July 2010 2:36 AM | No Comments Read More --> -
NVidia Introduces ARM Processor
06 January 2011 6:56 AM | No Comments Read More --> -
Nvidia Releases Driver to Fix Tomb Raider Performance
16 March 2013 10:06 PM | 1 Comment Read More -->
Related News:
Whoa, talk about bandwidth! So does that mean we’ll need some massive bandwidth increase for pcie too?
Maybe, but keep in mind that what he is talking about is on chip communication for processing graphics. The CPU will be able to get and send information to the GPU at a faster rate, but right now, I believe that PCI-e 3.0 isn’t being completely saturated. Volta may come close, and by then we could in fact see PCIe 4.0.
So, yeah it’s inevitable we are going to need more bandwidth. However, the performance increase is at times negligible. I remember reading an article on Tom’s hardware which they disabled PCIe Lanes on a 8800GTS and a x1900 XTX. This does show that in certain games PCIe bandwith could have a decent impact on performance. But, the performance difference in some cases is negligible.
So, while we may need more bandwidth in the future, we won’t exactly see a huge increase in frames with a jump in PCIe bandwidth (going from 3.0 to something like 4.0), for it may be some time for GPUs to fully utilize and require more bandwidth. Thus, on chip performance and bandwith of the GPU is what matters more and will give us the most performance, but we are going to eventually need greater PCIe throughput.
There is a more recent review of PCIe 2 vs PCIe 3 on hard:
http://www.hardocp.com/article/2012/07/18/pci_express_20_vs_30_gpu_gaming_performance_review/14#.UUo8bVcofIV
Basically, if you are on 2 you are still good and most of the performance gains were due to the fact that clock for clock ivy bridge is a little bit faster than sandybridge. Will a 1TBps card push the 3.0 standard? Well yeah that might happen with that much of an increase, but for Maxwell you should hopefully still be good on 2.0, but we shall see.
The one on anandtech suggested that 7970 crossfire would require pcie3 16s while single card was ok with 8. Considering how power by the time volta hits could jump (like 8800 from the 7 series and 4xxx from the 3 series) and that the big bandwidth between the chip and its ram would perhaps allow for a lot more juice which in turn would need to be communicated to the cpu, is what made me wonder.
We’ll see I guess lol.
Hard had 2 7970s and for the most part the performance was within the 10% margin of the clock for clock gain from going from sandy to ivy.
Also, I found this anandtech article and judging from the results I think people should be somewhat okay with pcie 2.0 for few more years:
http://www.anandtech.com/show/5458/the-radeon-hd-7970-reprise-pcie-bandwidth-overclocking-and-msaa
However, if volta does indeed have 1TBps of bandwidth I would really want to see some of these test with the newer cards repeated. I’m pretty sure 3.0 may be needed for a card of that caliber, but then again volta isn’t expect until what 2015 or 2016? We still have 2 cards based on Maxwell in between according to the presentation.
Exactly, it’s why I’m expecting to see more bandwidth needs.
Yeah, I think in 2015 they are actually going to have PCIe 4.0. So, if this volta card is really that insane there should be a new standard to handle it if in fact it is a problem for PCIe 3.
http://en.wikipedia.org/wiki/PCI_Express#PCI_Express_4.0
But, yeah, we need to play a wait and see game a little bit. Obviously if you were to upgrade now just get a ivy bridge processor with PCIe 3 for the hell of it. My bro and I are just saying that 2 isn’t exactly out of the equation yet.