The AMD Radeon 300 series “Fiji” with HBM is limited to 4 GB of video memory

In the spring, it’s time for a new graphics card series from AMD, which offers a blissful mix of old and new. Among enthusiasts, all focus is on the mythical graphics processor Fiji, which is expected to be equipped not only with a high number of stream processors but also fresh stacked memory.

The memory that is believed to be provided by SK Hynix is ​​called High-Bandwidth Memory (HBM), to be integrated on the same substrate as the graphics processor. This procedure not only provides record high memory bandwidth, the close integration should also lead to improved energy efficiency compared to the use of GDDR5.

Now Fudzilla reports that the AMD graphics card that will use Fiji, Radeon R9 380X or R9 390X depending on the source, will initially be limited to a capacity of 4 GB. This is said to be a direct consequence of AMD not being able to fit more than four memory capsules next to the graphics processor.

The information is consistent with SK Hynix starting production of HBM at the beginning of the year, where each canister has a capacity of just 1 GB. This includes an IO width of as much as 1,024 bits with a clock frequency of 1,000 MHz, which with four capsules gives a total bandwidth of 512 GB / s. However, rumors suggest that AMD’s top flagship will use a faster variant of 1,250 MHz for as much as 640 GB / s.

At the same time, this does not have to mean that AMD’s next flagship will remain at 4 GB of video memory for its entire life. Although HBM technology is still fresh, there is the possibility that SK Hynix will double its capacity during the year, which would thus enable a graphics card with a total of 8 GB.

In addition to the new memory technology, rumors point to Fiji getting 4,096 stream processors, around 45 percent more than the current flagship Radeon R9 290X. The launch of Fiji with associated graphics cards is expected to take place during the second quarter of the year, probably in connection with the electronics fair Computex 2015.