It's been a long time coming; there have been rumblings for a while now that AMD's next generation graphics cards would implement a new memory architecture. There was talk about this being a 3D approach to memory design, but there were no clear details about what that meant. At least, that was the case before today.
AMD claims to have managed to create memory that has higher bandwidth than GDDR5, far better power efficiency than we've seen before, and takes up considerably less physical space while doing it.
Although we still don't know when it will come to market, and the company isn't yet talking specifics, we do know that the first iteration of High Bandwidth Memory (HBM) will be used initially in consumer graphics before being implemented across other divisions. The company sees many uses for HBM in the future and expects to see everything from Compute to APUs take advantage.
AMD started work on this project over seven years ago. When the engineers first sat down and started thinking about what problems lay ahead, it became apparent to them that bandwidth per watt would quickly become an issue. As bandwidth demands go up, the traditional solution has been to integrate more into the die. DRAM doesn't have that luxury, and yet its bandwidth demands rise as CPUs and GPUs get faster and faster. For GDDR5 to keep up, more power is required, and at a certain point that is no longer feasible.
HBM was engineered to help solve that problem. Attacking the issue from all directions, the company came up with what it calls a 3D design, stacking four storage chips on top of a single logic die. This logic die is then attached directly to a silicon based interposer with no active transistors. The GPU, CPU or SoC die is also connected directly to the interposer, which itself is connected to the package substrate.
Because this is a very new approach, a whole new type of interconnects was developed for the stacked memory chips which have been dubbed "through-silicon vias" (TSVs) and "ubumps." These TSVs allow one chip to be connected vertically to the next and are also used to connect the SoC/GPU to the interposer.
With the memory chips thus stacked, they take up far less space. Where recent graphics cards have had memory chips surrounding most of the GPU die, an HBM configuration would only require four, and they would be positioned in the corners. In this approach there is far less distance for data to travel in order to reach the processor. Not only are there fewer chips, the HBM stacks use up far less surface area than GDDR5 modules. A single 1 GB stack is only 5 x 7 mm, whereas the same volume in GDDR5 would be 28 x 24 mm.
High Bandwidth Memory resets the clock, so to speak. For years, in order to get more throughput, clock speed increases were necessary. However, the Bus Width that HBM provides negates this need. GDDR5 runs on a 32-bit bus at up to 1750 MHz, which is 7 GB/s. The effective bandwidth can reach 28 GB/s per chip.
HBM, on the other hand, has a Bus Width of 1024-bit and a far lower clock speed of 500 Mhz (1 GBps), and the bandwidth peaks between 100 and 125 GB/s per stack. This is all while improving the bandwidth from 10.66 GB/s per Watt to over 35 GB/s per Watt.
These numbers are for the first generation of the technology, which AMD promised will increase dramatically in the second iteration. Expectations are that the speed will double and the capacity will quadruple by then.
Four times the capacity sounds outrageous, but the first generation is limited to a maximum of 1 GB per stack, and the tech only allows for four stacks to be used. This means the first generation of graphics cards will have frame buffers that max out at 4 GB. This may seem like an issue, but Joe Macri, AMD Product CTO overseeing HBM development, addressed this concern stating his belief that there is no problem that can't be overcome by the company's engineers.
Previously, there were no other options to boost performance other than to increase the frame buffer. There were no engineers working on solving this problem, but now with the way HBM can scale, AMD has put some of its own minds to work on finding more efficient ways to handle memory. It would seem the company is confident that 4 GB of HBM will rival the larger frame buffers that have been popular as of late.
It's only a matter of time before we're able to put AMD's claims about HBM to the test. New GPUs from AMD are expected in the near future, and now that AMD has finally lifted the lid on this well guarded new technology, there must be something coming sooner rather than later. If AMD's claims are to be taken seriously, GPU performance could see a significant leap in the coming year.
Follow us @tomshardware, on Facebook and on Google+.
