Micron Fast-Tracks Superfast GDDR6 Graphics Memory

Megalith

24-bit/48kHz
Joined
Aug 20, 2006
Messages
13,000
Thanks in large part to the booming popularity and demands of e-sports, GDDR6 graphics memory is reportedly debuting as early as this year. Originally intended for 2018, Micron has also decided to speed things up for reasons such as shorter console hardware cycles and virtual reality. As the technology matures, we should expect GDDR6 to provide double the performance of GDDR5.

Over time, GDDR6 will transfer data at 16Gbps (bits per second), making it two times faster than GDDR5, which is used in most GPUs today. It will also be faster than GDDR5X, which had an initial target speed of 12Gbps. Users will want higher bandwidth and low-latency memory to get better graphics on their PCs and gaming consoles, Eby said. GDDR memory is targeted at mainstream computing, Eby said. However, it's not yet known when GPUs with GDDR6 will appear. Micron will make GDDR6 chips and is already making GDDR5X memory. The GDDR5X memory will continue to be used as a low-end graphics memory replacement to GDDR5.
 
So will there be two different artchitectures moving forward: HBM and GDDR? I thought both AMD and nVidia were going HBM?
 
Are they saying that GDDR6 will have lower latency than GDDR5? That would be huge since higher bandwidth memory almost always has a higher latency.
 
I believe HMB and GDDR will complement each other, like a hybrid SSD/HDD drive. Speeding up GDDR still makes sense. Why they have to say e-sports and vr have anything to do with it is a head scratcher. Faster is better and has pretty much motivated the market thus far.
 
So will there be two different artchitectures moving forward: HBM and GDDR? I thought both AMD and nVidia were going HBM?


it's really going to depend on costs i feel. you may end up with high end being HBM and low end being GDDR 6. but it's hard to say since HBM allows for a much smaller footprint than using GDDR.
 
Are they saying that GDDR6 will have lower latency than GDDR5? That would be huge since higher bandwidth memory almost always has a higher latency.

No, it just states that users want higher bandwidth with low latency, nothing specific with gddr6. In other words I wouldn't make that jump cuz nothing is pointing to what you implied.
 
HBM3 is going to be a lot cheaper than HBM2 reportedly, but they're making it cheaper by making it somewhat slower. HBM2 is spec'd for stacks of 4 GB (1024 bit) running at 2 GT/s and 256 GB/s bandwidth. Nvidia's Tesla cards using it only run the memory at 1,406 MT/s and come in 12 GB and 16 GB configurations with 540 GB/s and 720 GB/s of bandwidth.

At 16 Gbps GDDR6 would provide 512 GB/s of bandwidth on a 256 bit bus and 768 GB/s on a 384 bit bus. GDDR5X was aiming initially for 12 Gbps and only managed 10 Gbps, so if GDDR6 "only" manages 14 Gbps it would be 448 GB/s and 672 GB/s.

The Titan X Pascal has 480 GB/s of bandwidth, so I think a Geforce 1080 successor with 16 GB of GDDR6 on a 256 bit wide bus with some new efficiency voodoo under the hood (similar to Pascal's color compression and other bandwidth saving stuff) would make sense if Nvidia's not already planning to switch their almost top of the line to HBMx.
 
Back
Top