AMD Fiji Die shots found – Architectural advantages explained

AMD Fiji Die shots found – Architectural advantages explained

 

Die shots of AMD’s R9 Fury X have emerged, giving us a great chance to explain the many changes made in their newest GPU, including some of the less apparent benefits of HBM memory. 

Here we will look at die shots of AMD’s two most recent GPU designs, the 359mm^2 Tonga silicon and AMD’s newest 596mm^2 Fiji silicon. Firstly we can see that AMD has done a great job in making their new Fiji silicon exceptionally dense, placing a 4096-bit memory bus and 4096 GCN GPU cores in much less than double the space of their Tonga Silicon which is known to have 2048 GCN GPU cores and a 348-bit memory bus. 

This increase in chip density allows AMD to produce these chips using less silicon, making them less expensive to produce. 

    

 

We all should know that AMD’s Tonga based GPUs use GDDR5 memory and Fiji uses the new HBM memory architecture, which both gives Fiji the benefits of lower memory power draw and increased bandwidth, but the benefits do not end there. 

When looking at these die shots we can see that the 4 1024-bit HBM memory Bus segments take up roughly the same space as Tonga’s GDDR5 64-bit memory bus, meaning that HBM also saves the die space required for the memory bus. Tonga uses 6 GDDR5 memory bus segments whereas Fiji only required 4 HBM bus segments, meaning that HBM also saves on-die space, allowing AMD to either use this space to add other GPU parts, or simply use this saved space to reduce chip size and manufacturing costs. 

If we were to consider that the GDDR5 memory bus segments in the lowly Tonga silicon takes up more space than the HBM memory bus takes up on Fiji, imagine the difference in size when compared to the memory bus on GPUs like AMD’s R9 290X or 390X, which use a larger 512-bit GDDR5 memory bus. 

 

    

 

Aside from the space saving on the die from using HBM-memory controllers high bandwidth memory also saves a lot of space on the PCB by allowing these chips to be placed right beside the GPU core and because these chips are much more storage dense than GDDR5.  This allows GPUs using HBM to take up 50% less PCB space than the equivalent GDDR5 GPU, which again saves money for creating the GPUs and gives the GPU a smaller form factor. 

The other benefit of having the HBM memory so close to the GPU core is that this effectively mean that all of the heat output from the GPU is coming from one central location, which allows GPU heatsink designs to be simplified. 

While GDDR5 and the upcoming GDDR5X memory architectures do still look appealing while HBM technology is in it’s infancy, but the benefits of HBM are much greater than mere memory speed increases and power consumption decreases, allowing less silicon to be used on the Core through the use of smaller memory buses and allowing more simplistic heatsink designs by consolidating almost all of the GPUs heat output to a singular location.

HBM is the undoubtedly future of GPU memory, but right now due to supply restrictions it’s adoption may not be a rapid as as it should be.

 

You can join the discussion on AMD’s Fiji Die shots and the Benefits of HBM memory on the OC3D forums.