[News] SK Hynix unveils GDDR6 Memory @ nVidia's GTC event

[EternityZX9]

SK Hynix has taken off the wraps off their next generation GDDR6 memory technology at NVIDIA?s GTC 2017. The new GDDR6 memory will be used to power NVIDIA?s most advanced GPUs to date, designed with the Volta chip architecture that is due for launch in 2018.

Coming to the talk, while NVIDIA uses the fastest GDDR5 solution on their cards, known as GDDR5X, they will soon be shifting to GDDR6. NVIDIA and Micron fine tuned the G5X memory architecture to hit really incredible speeds, topping out the first generation of HBM. With GDDR6, the chip giant plans to top out HBM2, bypassing any production issues associated with the specific memory type.

If we compare competitor solution, Fiji, the first graphics card to feature HBM had 512 GB/s bandwidth while the upcoming HBM2 based Vega is said to feature 512 GB/s bandwidth too along with a effective bandwidth and cache sub-system. For comparison, a Volta GPU with a 256-bit wide memory interface using GDDR6 clocked at 16 Gbps transfer speed would result in 512 GB/s bandwidth. A 384-bit Volta chip with similar rated memory chips will feature 768 GB/s bandwidth.

Not only would the 256-bit solution be faster, cheaper and more easy to produce than first generation HBM based solutions, but the 384-bit solution would be able to topple even HBM2 cards such as those based on current generation Pascal GP100 cores that have 720 GB/s bandwidth. For the folks wondering about efficiency, well the good thing about GDDR6 is that it brings 10% power saving over current gen GDDR5 memory chips.

The new chips feature low voltage of 1.35V compared to 1.55V and have twice the bandwidth. GDDR6 would allow for graphics card with increased capacities of 16 GB and 24 GB and it looks like gaming enthusiasts will have no problem hitting higher resolutions in their favorite titles.

Link to full article

[StyM]

volta ftw!

[EternityZX9]

Yeah so basically while AMD has to deal with the HBM2 problems nvidia is going to be coming out with GDDR6 16GB/24GB video cards next year. Really can't wait to see the monster combinations they could come out with next year.

[BenchZowner]

Day by day the choice to go HBM2 seems to be biting AMD in the a$s.

[Particle]

The bandwidth is impressive, but a 10% power savings isn't going to unlock the extra logic power budget that HBM's much larger savings does. We need both. Use HBM/HBM2 for high end cards and use GDDR6 for everything else.

[EternityZX9]

The bandwidth is impressive, but a 10% power savings isn't going to unlock the extra logic power budget that HBM's much larger savings does. We need both. Use HBM/HBM2 for high end cards and use GDDR6 for everything else.

I think that depends. Like, on what architectural improvements Volta brings + GDDR6 memory power savings, and everything else they put into the next generation of cards. It could really start to add up and make the path for the power savings by route of HBM2 diminish in importance.

Should have some more clarity in regards to this tomorrow after 9 am PST at the GTC keynote.

[BenchZowner]

Power saving for the flagship sounds stupid to me.
Ok for laptops and handhelds, but... who would prefer a for example 10% slower card for 5W of power saving ?

[zanzabar]

Power saving for the flagship sounds stupid to me.
Ok for laptops and handhelds, but... who would prefer a for example 10% slower card for 5W of power saving ?

it is also much lower latency and acts as a cache layer with direct access. if you look at the amd road map the point was to move to hbm, then have add in cards or extended ram configurations adding gddr to cheaper cards. on the cpu side look at what intel wants to do with 3d x-point and hbm offers the same thing. the problem is that hbm1 was basically useless at 1GB per stack, and hbm2 while offering 4-16x more density (depending on the stack format) at similar speeds per layer; there was no one making it or even had it ready to go small batch. even now samsung is the only one shipping it and they are listed as limited samples that you cannot order as a normal person (or company with access to their parts portal.) in the past anyone with access could order samples if it was about to enter bulk for ddr, gddr, or nand.

going all in so you have no way to make something like a gp102 was clearly a bad idea, but nvidia is also heavy on hbm2 just not all in.

[Sparky]

OK pet peeve - they are HBM not HMB

[Particle]

Power saving for the flagship sounds stupid to me.
Ok for laptops and handhelds, but... who would prefer a for example 10% slower card for 5W of power saving ?

The numbers are larger, and you're also missing the point. The point is to save power on the memory array without sacrificing memory array performance and then reinvest that savings into logic power. The logic can be larger (ie more shaders) or clocked higher without increasing total board power. The HBM array on the Fury saved something like 25-30 watts of board power and yet exceeded the performance of what GDDR5 memory arrays at the time were capable of.

Instead of getting a 10% slower card for 5 watts of power savings like you suggest, it's more like getting a 10% faster card for the same power due to having been able to dedicate more power to logic without increasing total board power.

[BenchZowner]

It's not like they're restrained to the 250W envelope...

@Zanzabar

Lower latency isn't beneficial to any GPU for gaming.

[Particle]

No conclusion would seem to follow from your statement unless you're suggesting that what I've outlined isn't desirable--which it is.

[zanzabar]

OK pet peeve - they are HBM not HMB

fixed it, sorry.

It's not like they're restrained to the 250W envelope...

@Zanzabar

Lower latency isn't beneficial to any GPU for gaming.

you are looking at latency boundaries in the range 10-100x lower than gddr. it should make a big difference as rops can pull data from the hbm instead of having to bounce data from the ram to the cache then to the ROP. that changes the operations from an interrupt to a wait and is much more efficient so long as the memory is fast enough.

[BenchZowner]

With the traditional design the data you request is at the VRAM already when it is asked by the ROPs, not from the RAM ( that would make even a GV100 look like a S3 Virge DX )