AMDs latest Linux drivers did not only come with the usual performance increases, tweaks and fixes but with a rather interesting addition. Within it a new chip was referenced:
Seeing that in the drivers prior to the Vega architecture the then soon to be released cards were codenamed GFX9 we’re most likely getting a small hint that Navi, Vegas successor, is on its way. But don’t get too excited: Linux drivers tend to be months ahead of their time concerning chip compatibility to ensure a stable transition so we’re definitely not talking about a few weeks or even a couple of months.
That would further be enforced by AMDs roadmap that already mentioned Navi, based on the new 7nm process and most sites and their sources, place Navis release to the end of 2018, like PCGameHardware, put the earliest possible release of Navi into Q3 2018 which would also explain the super secret codename to that there is not even a hint about the lineup at all and if we can expect the lower end models first or the other way around.
But despite this secrecy there are already some details that got us thinking within these slides. It’s pretty obvious, thanks to the 7nm process that Navi will deliver more performance per Watt than Vega, which still uses the 14nm version. But Nexgen Memory and Scalability are the word that are most intriguing. The memory will most likely not be HBM2, otherwise it would be there, as it is with Vega, leaving HBM3 and GDDR6 as possibilities. But with Samsung placing HBM3 into 2019 we’ll most likely see GDDR6 and up to twice the memory speeds and data rate compared to GDDR5. That would surpass even AMDs current HBM2 performance.
Even more interesting than the question about memory is the scalability. Could we see another example of AMDs infinity fabric and if so, how will this influence GPU costs? With Threadripper and Epyc, also “glued together” chips, based on normal Zeppelin dies that work together thanks to Infinity fabric, we saw not only how easy it is for AMD to put out processors with hugely different performance values without much additional effort, but also an enormous drop in prices for high end chips. Without the need to set up a whole processing line for a smaller quantity of high end dies production costs drop dramatically. In addition to that the yield can be higher thanks to the fact that dies with defective/ disabled cores can still be coupled and result in yet another chip like we saw with the Ryzen 6 1600, sporting two dies with only 3, out of the that 4 that are physically there, active cores each.
There is much to be hyped about, but sadly we still have to wait quite a bit to finally see what AMD has been cooking up there.