HBM (High Bandwidth Memory) — The Future of Graphics Cards?

When HBM (High Bandwidth Memory) was announced, it shocked the world with its substantial performance improvement over GDDR5. It allowed the 4GB AMD Fury card to go toe-to-toe with Nvidia’s seemingly much better-endowed cards with 8GB GDDR5. Now HBM3 has been announced, and the performance gains are just as extraordinary.

We all want more power. In our systems, I mean. But often times, we fail to see the bottlenecks in our system. It was only a few years ago that SSDs were scarce and expensive, only coming in low capacity packages. But HDDs just couldn’t keep up with the state of CPUs. Your HDD was, more often than not, the single biggest bottleneck of your PC.

Now, with respect to high-end graphics cards, it seems to be the Video RAM (VRAM). It’s a profound statement to say that the processing power exceeds the speed of the working memory available to it. HBM can push 128GB/s, this is stunning in comparison to the 10GB/s GDDR5 that is still often used in high-performance cards today.

The problem with High Bandwidth Memory was its expense. It was simply too costly for use in all cards. But with the recent announcement of HBM3 (HBM2 was very limited in availability, only on select Nvidia cards) at the Hot Chips Symposium, the price is expected to be reduced to an affordable point, below that of HBM upon its first availability, while giving massive performance improvements.

HBM3 has solved some of the packaging problems of HBM and HBM2, namely increasing the number of chips in a memory stack, and has reached an astronomical throughput of 512GB/s. Samsung and Hynix have both announced the production of this incredible memory, though, Samsung has opted for the moniker of Extreme HBM for its product. Also worth mentioning is that Samsung will be producing a “Low Cost” HBM that will push a blazing 200GB/s, and in addition, while it does seem almost pointless to mention in contrast, they will be producing GDDR6 which will perform at 14GB/s, a 40% increase over the GDDR5 standard.

Regardless of how much HBM3 will cost it’s surely good news for computer enthusiasts the world over because the technology always trickles down. Excitingly, in the case of HBM, is that AMD’s new Zen architecture has been designed with the potential to house HBM (and presumably future iterations of the product) directly on the die which can give untold performance enhancements to integrated GPUs and even CPU if AMD so chooses to use HBM for system RAM.

Do any fossBytes readers own an HBM or HBM2 equipped product? Let us know about your experience in the comments below.