AMD’s current series GPU family is a volcanic island, last week broke the next generation of the pirate island R9 300 series, which has Bermuda, Fiji and Treasure Island three core specification is also very impressive, stream processor units from the current maximum 2816 upgrade to 4224. Behind such a massive upgrade, AMD will inevitably have to do to compromise on the GPU core area in the past insisted on small core strategy might change, said that “Fiji” GPU core area is not less than 550mm², may be 600mm ². In addition, the family may spend pirate island faster HBM stacked memory. (HBM stackable memory is designed to support multi-channel bandwidth increased dramatically).

Future of AMD’s Graphic Card Guess the Next Generation: Pirate Island family is the Future? 

Currently there is no exact information on Pirate Island family GPU architecture and specifications, especially process technology and core, previous rumors use TSMC’s 20nm process called, but even so, significantly increasing the size of the operation also had to AMD in the core compromise on the area. 3D center fervent discussions on the ongoing analysis of their given area to be considered the core of Fiji over 550mm ², even up to 600mm ², the level of performance of up to R9 295X2.

AMD had always insisted on a small core strategy, the same generation GPU core area generally be lower than NVIDIA products, Tahiti generation core area of 365mm ², although more than GK104 core 294mm ² to be high, but lower than the GK110 core 561mm ², even R9 290X core of Hawaii is only 438mm ², or no more than GK110 level. Now with the GPU to enhance the scale, AMD’s GPU core area also had to move towards a higher level, whether it is 550 or 600mm ², have reached or even exceeded the level GK110’s.

Of course, this analysis is only one of the words, some people do not agree with the core AMD route would go big, think big heart in the manufacturing and yield are difficult issues, but the big core NVIDIA GK110 has not heard of yield problems, AMD should not have a large core of yield problems. But no matter what kind of advice, Pirate Island GPU core scale will certainly grow, the core area of ​​improvement is inevitable.

More controversial still Pirate Island GPU architecture and technology, the previous Tahiti, Hawaii core architecture is GCN 1.0 and GCN 1.1, pirate island belongs to the process, structure replacement, will be upgraded to the  GCN 2.0 architecture. As computing performance enhancements, GCN 2.0 bandwidth requirements are growing, Pirate Island may spend vendors such as AMD and Hynix jointly developed by HBM stacked memory.

HBM stackable memory
HBM stackable memory

HBM stackable memory is designed to support multi-channel bandwidth increased dramatically.

There are technical background HBM memory before being introduced, AMD and other manufacturers develop joint SK Hynix memory HBM also a long time, through silicon vias TSV technology stack multiple memory channels, memory bandwidth and capacity has improved significantly, not only for memory, but also can be used for graphics and grid computing. Compared GDDR5 memory, HBM can improve the overall performance of 65%, while the power consumption is 40% of the decline.

Pascal previously announced NVIDIA 3D architecture also uses the stack memory, multi-channel package, memory bandwidth, and thus enhance the capacity of the memory is also achieved through silicon via TSV technology, AMD next-generation GPU architecture spend HBM memory, it’s not too unexpected, the only problem is the technical maturity 3D stack memory, similar to the formation of such a universal standard GDDR5 have commercial possibilities.

As Pirate Island GPU’s progress, if AMD really intends to use TSMC’s 20nm process, and at the end of the year, so now it should be almost time to tapeout, especially large core GPU needs more time, NVIDIA’s GK110 core September 2012 for the first time used on the professional card, but in January 2012 on the flow sheet, this schedule can be used as a reference.