Graphics memory exploded in 2017 Says Micron - Next in 2018 is GDDR6
Micron on its blog talks a little about the massive graphics memory market in 2017 but also talks what’s next for 2018, yes they will ramp up GDDR6 technology and did not even mention HBM2.
GDDR6 is expected to launch with transfer rates between 12Gb/s and 16Gb/s per channel. Their GDDR5 or G5X memory was for Micron’s graphics team as they had a record-breaking year in GDDR5 shipments and their 12 Gb/s G5X, the highest-performance mass production GDDR memory.
In markets outside of graphics, 2017 provided many opportunities for deep dive conversations around GDDR6. Mobile World Congress America focused discussions in automotive applications. At Linley Processor conference, GDDR6 was looked at for networking applications. The broader interest is a positive sign for what’s to come.
Micron blog --- So, what comes in 2018? Looking forward, a number of trends in each of these key segments will continue to require faster GPUs, hungry for more memory bandwidth and density. That’s one of the reasons we have continued making R&D investments in our GDDR portfolio, to increase memory bandwidth in a discrete component with established packaging, handling, and testing characteristics. We’ve made great progress on a number of programs since my last update, so I’ll share a brief overview of that today. First, we have recently begun mass production shipments of 8Gb GDDR5 on our most advanced production DRAM node. This gives our customers and partners a long-term support option for a proven graphics memory and delivers increased speed margins. In June of this year, we mentioned that our GDDR6 program was on track and progressing well. Since that time, our Graphics teams in Munich and Boise have been hard at work, and both design and device qualification are complete. We are very pleased with the performance this memory will offer our customers across a range of products that crave memory bandwidth.
Being the only memory manufacturer to design, test, and mass produce G5X puts us in a unique position with GDDR6, which shares a very similar specification. Our initial progress is very promising, as you can tell in the data eye and wafer schmoos I’ve posted with this blog. Designing devices with the level of high-speed signaling required for these graphics memories is no simple task. We are confident our GDDR6 is leading the industry in these metrics, and we share them because we think it’s important for our customers to see this progress.
Simulated data eye openings at 16 Gb/s – Combined de-emphasis and DFE (left) and stronger (3dB) de-emphasis (right).
Measured link margin shmoos at 16 Gb/s without and with single-tap de-emphasis enabled.
For the techies reading this, please drop by Designcon to see the detailed paper and presentation from Micron outlining techniques for designing for GDDR6 at 16 Gb/s. Like G5X, we see a path for GDDR6 to offer a 16 Gb/s bin over time, with initial mass production bins of 12 and 14 Gb/s. With developments like these on the horizon, 2018 is going to be another big year for anyone who loves graphics tech. We can’t wait to see it, thanks SH SOTN foir the news-submit.
NVIDIA Titan V Graphics Card Benchmarks - 12/09/2017 09:43 AM
The first benchmarks of the new NVIDIA Titan V Volta Graphics Card have arrived. The numbers include Ashes of Singularity, Ashes of Singularity DX12, Rise of The Tomb Raider - Max 1440p, Gears of War...
NVIDIA Launches TITAN V "Volta" Graphics Card With 21 Billion Transistors - 12/08/2017 08:18 AM
Completely out of the blue, Nvidia just announced its new flagship graphics card, the TITAN V, the card is based on a "Volta" GV100 graphics processor, based on a GV100 is a multi-chip m...
Review: Destiny 2: PC graphics benchmark analysis (updated) - 11/24/2017 11:35 AM
We test that PC enhanced Destiny 2 for Windows relative towards graphics card performance with the latest AMD/NVIDIA graphics card drivers. Multiple graphics cards are being tested and benchmarked. We...
GALAX SNPR External Graphics Enclosure Based on GTX 1060 - 11/22/2017 08:02 AM
GALAX has launched its SNPR External Graphics Enclosure, much like offerings from like Gigabyte, this is an external box with power supply and graphics card. It connects (and requires) Thunderbolt 3....
ELSA GeForce GTX 1080 8GB S.A.C R2 Graphics Card - 11/20/2017 08:52 AM
Everey once on a while you stuble into an ESLA card, they are still arounf and have released their GeForce GTX 1080 8GB S.A.C R2 Graphics Card. With its 2560 Shader Cores, a 256-bit memory interfac...
Senior Member
Posts: 13234
Joined: 2004-05-16
Micron was backing a different technology along with Intel/Nvidia called Hybrid Cube Memory, which is similar to HBM but AMD/SK Hynix beat them to the punch and IIRC HBM was better suited for graphics workloads where as HCM was more for general purpose ones. As far as I know they gave up on the tech, so it's possible they will start shipping their own HBM parts soon.
Senior Member
Posts: 1385
Joined: 2004-02-14
With memory sizes on video cards getting bigger and bigger. I bet with GDDR6 we'll see a shift toward 2GB per chip on the PCB.
This would also mean Nvidia would need to move up to 64-bit memory controllers if they want enough actually memory bandwidth for these larger memory footprints.
I bet the GTX 1180 will likely have 12GB DDR6 12Ghz.
Which would fit it with 6 x 2GB chips linked with 64-bit controllers for 384-bit memory interface. That would net 576GB/sec memory bandwidth!
On the GTX 1180Ti would likely upsize to 14GB DDR6 13Ghz or 14Ghz.
That would come out to 7 x 2GB chips again with 64-bit controllers for 448-bit memory interface. Which would get 728GB/sec to 784GB/sec bandwidth depending on the memory speed.
If you compare that to the HBM on Titan V with 653GB/sec memory bandwidth it would really make more sense to just use GDDR6 with a larger memory interface.
Senior Member
Posts: 2513
Joined: 2012-04-16
I won't mention the h-word either.
Senior Member
Posts: 12500
Joined: 2010-05-22
I wondered what woke me up.
Senior Member
Posts: 2599
Joined: 2006-09-02
i do understand that hbm is a bit complicated to manufacture, i believe that micron is not leaning towards the risk of hbm right now, but i do believe hbm is the future. it greatly reduces board complexity. it just needs to mature just a bit more.