AMD Fiji GPU ‘Not Arriving Soon Enough’ – Will Possibly Have Stacked dRAM (HBM) Memory

Usman Pirzada
Posted Aug 15, 2014
30Shares
Share Tweet Submit

[Rumor] Before we begin, a short reminder that this post is tagged as rumor, so take it with that famous pinch of salt. The aptly named SemiAccurate.com has published something that sounds quite interesting. The article hints at some unique feature in Fiji GPU and then just states that it wont be arriving for a while. Because I didn’t cough up a thousand bucks, I can literally see only 3 lines of text and I am not sure what to make of it.AMD Pirate Islands GPUs Bermuda Fiji Treasure Island

AMD Will Probably Feature HBM Stacked Memory in Fiji GPU – Not Coming Soon

I will attempt to deduce anything and everything from these lines, so this article is mostly pure educated speculation. If this isn’t your cup of tea, you might want to skip this write up. Basically the article hints that: “AMD’s upcoming Fiji has something fun, something that will provide some very interesting numbers”. Now depending on when Fiji launches, it might be based on the 20nm node but that isn’t really something unique or ‘fun’ as Charlie puts it. Its just standard lithographic progression so I very much doubt he is referring to that. No, in my opinion there is only one feature that fits the bill and that is HBM or stacked dRAM if you may. There is a very very huge amount of evidence on AMD working on stacked dram for graphical uses and this seems like the ideal fit to “something fun”. An alternative would be an ARM core on the GPU but somehow I think thats too much of a long shot.

We know for a fact that AMD is working with Hynix to introduce HBM into its graphical solutions and I have found evidence of as big as 8-Hi HBM Stacks. For the nomenclature averse, this means 8 stacks of dRAM core dies with the control die at the bottom and TSVs (through silicon vias) connecting the stacks. HBM uses a very low amount of voltage (1.2V) and delivers upto 256 GB/s (8-Hi) per chip of throughput which is quite an insane amount. In comparison modern GDDR5 ram delivers only 28GB/s per chip. Now here is the thing, the very shoddy leak we received a long time back was only good for a few facts, and one of these was the nomenclature of AMD’s R9 300 GPUs. Because of that we know that Fiji XTX is not actually the highest end chip, rather Bermuda XTX is (unless AMD has decided to do a shuffle). So why exactly is the source talking about Fiji like that? Well, if I were to take a guess at AMD’s road map, I would say that Fiji XTX is probably 28nm with HBM while Bermuda XTX  is 20nm with HBM. Or it could be that Fiji is 20nm with HBM while Bermuda is 28nm without HBM, that would fit in with the leak mentioned above since it said that Bermuda would arrive first. The 3 lines say its not arriving soon, but of course I have no idea weather not arriving soon is described as by months or years or what.

Share Tweet Submit