Back to all articles

AMD breaks the mold with high bandwidth memory


The integrated circuitry you’ll find in your CPU or GPU is an astounding technological feat. Billions of microscopic logic gates squeezed onto a semiconductor wafer only a few hundred millimeters across. However, that circuitry is two dimensional, which is why die sizes are expressed at square units and not cubical ones. 3D integrated circuitry on the other hand, stacks multiple connected silicon wafers on top of each other, packing much more into a more efficient space.

Until now we haven’t really seen many commercial applications, although believe it or not Intel built a 3D Pentium 4 back in 2004. Of course, Samsung’s 3D V-NAND can be found in its 850 PRO SSDs which has also brought a significant improvement to the performance of that technology.

Now, AMD has put 3D circuitry products into the consumer space. The R9 Nano graphics card from AMD is a flagship, high-end graphics card performing at about the level of a Nvidia GTX 980, but comes in at only six inches in length. This is possible mostly thanks to the new High Bandwidth Memory (HBM) that AMD has created.

This memory enjoys the advantage that 3D stacked circuits: a smaller package size, significantly lower power consumption and a massive improvement in bandwidth. While the GDDR5 memory in the GTX 980 is 256 bits wide, the HBM on AMD’s card is 4096 bits wide. This gives the memory 512 GB/s of effective bandwidth, despite only being clocked at 500Mhz. In other words, unlike GDDR which uses a narrow pipe (bus width) but very high flow (clock speed) HBM does the opposite. For one thing this means parts than run much cooler, with the total amount of bandwidth better than the GDDR solution. Time will tell if speeds on this wide bus can also be ramped up, but it seems that memory certainly won’t be a bottleneck. One current limitation for HBM is a 4GB memory limit, so it’s unclear when we would see Quadro or FirePro type cards with 12GB HBM allotments. This is a limitation of the stack height, the number of stacks the GPU can can connect to and of course the memory density of the actual stack layers. It also remains to be seen if HBM will will meet the memory precision requirements for professional workstation cards.

What is clear, is that this is a practical, reliable technology that might be set to change the size, energy and performance limits of consumer computing devices in a significant way. HBM is very impressive for a first-generation product and can only get better from here.

What do you think about HBM? Let us know in the comments.