AMD's CTO of Computing and Graphics Joe Macri addressed concerns around potential memory capacity limitations with HBM, especially on high end GPUs. As AMD's upcoming ultra enthusiast GPU code named FIji is rumored to feature 4GB of memory. The same capacity as Nvidia's current high end GTX 900 series offerings and AMD's last generation Hawaii ( R9 290 series ) graphics cards.
The potential concern is that on such a powerful GPU as Fiji is rumored to be, 4GB of memory may pose a limitation. Fiji is AMD's upcoming flagship ultra enthusiast class GPU. It 's also the world’s first GPU to leverage stacked High Bandwidth Memory.
The logic is that the faster the GPU, the more likely that users will push the graphical settings of a game. And thus more memory capacity will be utilized to render the effects of these higher in-game graphical settings. This however isn't the issue here because 4GB of memory is actually quite plenty. But a question around the possibility of 4GB posing an issue at very high resolutions such as 4K and high resolution multi-monitor setups can arise.
AMD's Joe Macri Addresses Potential 4GB HBM Capacity Concern With High End GPUs Such As Fiji
Arstechnica managed to get Joe Macri on the phone to talk more about HBM. And in his comments Macri addressed multiple aspects of the concern mentioned above. Is HBM limited to 4GB on high end GPUs ? and if so will 4GB be enough ? The short answers are no and yes. But let's hear what Joe Macri had to say for a more comprehensive view.
"You're not limited in this world [die stacking world] to any number of stacks, but from a capacity point of view, this generation-one HBM, each DRAM is a two-gigabit DRAM, so yeah, if you have four stacks you're limited to four gigabytes. You could build things with more stacks, you could build things with less stacks. Capacity of the frame buffer is just one of our concerns. There are many things you can do to utilise that capacity better. So if you have four stacks you're limited to four [gigabytes], but we don't really view that as a performance limitation from an AMD perspective."
For obvious reasons Macri did not reveal the actual memory capacity of Fiji. Specifications will come once the GPU is launched which AMD says you'll be able to buy this summer. Macri pointed out that there is no 4GB limitation. Products can be designed with as many memory stacks as is desired. Note, with "stacks" Macri is referring to HBM cubes so to speak. Because each cube is still limited to 4 DRAM dies, each 256MB large, layered on-top of one another. What is not limited however, is the number of cubes / stacks you want to add to the design.
With 4, 256MB dies, you get a total of 1GB of memory with each HBM1 stack. Four of these would net 4GB of memory. For higher capacities more HBM stacks can be arranged around the host processor, in the case it's the Fiji GPU.
Macri also addressed the second question, which is if 4GB is all we had ,is it going to be enough ? Macri says yes.
"If you actually look at frame buffers and how efficient they are and how efficient the drivers are at managing capacities across the resolutions, you'll find that there's a lot that can be done. We do not see 4GB as a limitation that would cause performance bottlenecks. We just need to do a better job managing the capacities. We were getting free capacity, because with [GDDR5] in order to get more bandwidth we needed to make the memory system wider, so the capacities were increasing. As engineers, we always focus on where the bottleneck is. If you're getting capacity, you don't put as much effort into better utilising that capacity. 4GB is more than sufficient. We've had to go do a little bit of investment in order to better utilise the frame buffer, but we're not really seeing a frame buffer capacity [problem]. You'll be blown away by how much [capacity] is wasted."
According to Macri, GDDR5 fed GPUs actually have too much unused memory today. Because to increase GPU memory bandwidth, wider memory interfaces are used. And because wider memory interfaces require a larger amount of GDDR5 memory chips, GPUs ended up with more memory capacity than is actually needed.
Macri also stated that AMD invested a lot into improving utilization of the frame buffer. This could include on-die memory compression techniques which are integrated into the GPU hardware itself. Or more clever algorithms on the driver level.
We have no doubt that once the GPU is released, we'll see very extensive tests of the memory system. Both to analyze the effect of the huge increase in memory bandwidth that is enabled by HBM has on performance. And to find if the memory capacity will be as sufficient as Joe Macri believes. The truth is that we haven't seen a single GPU yet that can actually make use of 4GB all on its own all the while maintaining playable framerates. You'll simply run out of performance way before you run out of memory.
Multi-GPU setups of course are the exception. You'll actually have enough performance to turn the settings up and fill the memory buffer while maintaining playable framerates. And that's when it may very well become a limitation. Fortunately however that's where low level APIs such as DX12 and Vulkan may prove to be the solution. With advanced multi-GPU rendering techniques, the memory buffers of multiple GPUs can be combined to form a larger memory pool that can be utilized and that would solve the issue.
There are so many moving parts in the GPU industry today that are sure to be important in the future. A new memory standard, multiple new low level APIs and various rendering innovations do bring back much needed excitement to this space.