Posted by Marc Greenberg on June 17, 2015
AMD announced their new line of GPUs are using the new HBM (High Bandwidth Memory) DRAM technology yesterday. I have known these were coming for a while but the thing that surprised me the most was the relatively reasonable cost for the performance that they deliver – at least, the relationship between cost and benefit of adding HBM to the system appears to be almost linear.
The high-end GPU using HBM, the Radeon R9 Fury X, has a recommended price of $649 and has 512GB/s of DRAM bandwidth to 4GB of HBM DRAM connected to 4096 stream processing units. (source)
The nearest GDDR5-based system, the Radeon R9 390X, has a recommended price of $429 and has 384GB/s of DRAM bandwidth to 8GB of GDDR5 DRAM connected to 2816 stream processing units. (source)
So the Radeon R9 Fury X has about 33% more memory bandwidth and 45% more stream processing units than the 390X for about 50% more recommended retail cost. I am assuming the AMD engineers did their homework to balance the number of stream processing units with the bandwidth available, and therefore we could assume that they are using the available bandwidth from the HBM memory more efficiently than they did with GDDR5. Yes, there is half the amount of DRAM capacity available, but in general the GPU applications are more limited by bandwidth than capacity, and the 8GB of DRAM in the R9 390X may partially unused just because that’s how much capacity they needed to buy to get the number of pins required to transmit the 384GB/s bandwidth in the 390X. We’ll need to look at the relative performance analyses that come out from the gamer labs across the internet but on the face of it, it looks like there is a pretty linear relationship between cost and benefit when adding the HBM technology to the system.
Then there’s the issue of heat. The Fury X is capable of doing more work than the 390X and therefore you might expect it to get hotter. To that end, AMD’s specification sheet says that the Fury X is liquid cooled. If you’ve heard me talk about heating in DRAM devices before, then you’ve heard my super-secret retirement plan: I’ll retire wealthy as soon as I invent a DRAM that works better when it’s hot!
I have had concerns over HBM in the past for the reason that DRAMs don’t like to be hot, and one of the last places I would choose to put a DRAM device would be in close thermal proximity to a high performance computing element. Unfortunately the nature of HBM – and one of the reasons it can provide so much bandwidth – is it’s requirement to be placed close to the computing element it serves. So it appears that AMD have addressed this with liquid cooling and some of the cost of the Radeon R9 Fury X may be due to the liquid cooling system rather than the cost of the memory.
Finally, it’s very important to note that this cost/benefit relationship applies to AMD (and specifically AMD GPUs) and not for every system out there – you couldn’t build HBM into a low-volume Enterprise product and expect the same cost/benefit. AMD can benefit from their consumer volume pricing on DRAM, their consumer-speed inventory turns, and they can amortize the NRE cost of the silicon interposer required for HBM across a large volume of devices. Someone building a lower volume product with longer inventory turns could expect a very different cost/benefit…
You can read the AMD press release here: http://www.amd.com/en-us/press-releases/Pages/new-era-pc-gaming-2015jun16.aspx
Graham Allan is the Sr. Product Marketing Manager for DDR PHYs at Synopsys. Graham graduated from Carleton University's Electrical Engineering program with a passion for electronics that landed him in the field of DRAM design at Mosaid in Ottawa, Canada. Beginning at the 64Kb capacity, Graham worked on DRAM designs through to the 256Mb generation. Starting in 1992, Graham was a key contributor to the JEDEC standards for SDRAM, DDR SDRAM and DDR3 SDRAM. Graham holds over 20 patents in the field of DRAM and memory design.