2016 will be marked with the arrival of two memory standards, which should spread across the mainstream and high-end / enthusiast line-up like fire. First, we have the HBM2, an improved version of HBM memory which debuted (and so far, only ships inside) with AMD R9 Fury family of cards. HBM2 promises a four times increase in capacity and double the memory bandwith – meaning a single card can go from 4GB and 512GB/s to 16GB and 1TB/s.
Given the low volume of HBM and HBM2 memory, those two will probably remain only on enthusiast graphics cards, such as recently renamed Greenland, high-end Polaris graphics processor from AMD – and Pascal, next-generation GPU from NVIDIA. For the sub-500 dollar world, Micron is readying GDDR5X memory, which is serving as a stop-gap to GDDR6, scheduled to debut in 2017-18. When we talked to Micron last year, they made sure to point out that GDDR5X and GDDR6 are different products altogether.
GDDR5X is a memory type developed by Micron’s Graphics DRAM Design Center located in Munich, Germany. Originally set up to develop memory standards aligned with ill-fated Intel’s Larrabee graphics processor, Micron expanded the role of the lab and focused on becoming one of best memory laboratories in the world. Graphics DRAM Design Center is now closely working with all three majory players – AMD, Intel and NVIDIA. GDDR5X could also see adoption in high-end switches from companies like Cisco. While not a common knowledge, high-end network switching also utilizes GDDR memory, originally intended for graphics processors.
Couple of weeks ago, memory standards body JEDEC published the official GDDR5X Graphics Memory Standard, which you can download here; Graphics Double Data Rate (GDDR5X) SGRAM Standard. Micron’s GDDR5X is manufactured in-house, using the 20nm process technology. First generation GDDR5X chip comes with 8Gbit / 1GByte capacity, while we should be seeing 8/16 Gbit chips as the yielding improves. Supported memory configurations are as follows:
- 4GB – 4x1GB; either 64-bit or 128-bit memory controller
- 6GB – 6x1GB; 384-bit memory controller
- 8GB – 8x1GB; 256/512-bit memory controller
- 12GB – 12x1GB front and back; 384-bit memory controller
- 16GB – 16x1GB front and back; 256/512-bit memory controller
Due to arrival of HBM2 memory, we don’t believe we will see 12GB and 16GB GDDR5X cards, but it is possible that companies such as NVIDIA and AMD will continue manufacturing (and rebranding) Maxwell and GCN-based graphics chips, with the updated memory configurations.
With the first chips now coming ‘from the fab to the lab’, Micron’s memory experts started to test GDDR5X memory chips, and they’re getting very encouraging results. Kristopher Kido (Global Director, Graphics Memory Business), Michael Richter and Ralf Ebert published a blog with promising statements:
“Micron’s GDDR5X program is in full swing and first components have already completed manufacturing. We plan to hit mass production this summer. The team at our Graphics DRAM Design Center in Munich, Germany is doing a fantastic job, too. Not only do we have functional devices earlier than expected, these early components are performing at data rates of more than 13Gb/s (1.63 GB/s)! Memory components mature as they move through the development and manufacturing process, so to see first silicon performing at nearly full performance specs was a pleasant surprise—these early results are incredibly promising.”
Standard states that GDDR5X is expected to yield between 10-14Gbps per pin (1.25-1.75 GB/s), meaning a theoretical maximum from a single GDDR5X chip should be 237.5 GB/s per 32-bit memory interface (190 FBGA package), but that’s when power contacts are excluded. If we would compare a GeForce 980 Ti equipped with GDDR5 and GDDR5X, we would get the following bandwidth performance
- GeForce GTX 980 Ti GDDR5 stock – 336.5 GB/s
- GeForce GTX 980 Ti GDDR5X Prototype – 624 GB/s
- GeForce GTX 980 Ti GDDR5X Production – 672 GB/s
As you can see for yourself, both today’s GDDR5X prototype and estimated production units would achieve results higher than the current world’s highest memory bandwidth champion, AMD’s Radeon R9 Fury X with HBM memory and 4096-bit memory controller, which achieves ‘only’ 512 GB/s. In our internal testing, we’ve been able to overclock the memory to 607MHz, getting 621.57 GB/s – less than GDDR5X can achieve.
The future for Virtual Reality, and Augmented Reality – as well as regular 1080p, 1440p and 4K gaming is looking very bright – 6/8GB of GDDR5X memory with 448-672 GB/s bandwidth should help GPUs achieve that magical goal of rendering VR content at 90/120 Hz.