Samsung Announces New GDDR6 Memory, 8 TB NGSFF SSD

[Raevenlord]

Samsung managed to snag 36 CES 2018 Design and Innovation awards, and the company took to a press release to acknowledge the honors, and shed some light on some of its upcoming products and technologies. Of particular interest to us enthusiasts is the presence of GDDR6 memory, which Samsung says is "(...) The fastest and lowest-power DRAM for next generation, graphics-intensive applications." this new memory is expected to process images and video at 16 Gbps with 64GB/s data I/O bandwidth, which is capable of up to 5 GB/s speeds. Additionally, Samsung said the new DRAM can operate at 1.35 V, offering increased power efficiency over today's graphics memory (which typically requires 1.5 V).

Samsung also announced a NGSFF (Next-Generation Form-Factor) SSD solution, which "dramatically improves the storage capacity and performance of 1U rack servers," which allows for I/O speeds at 0.5 petabytes per second. Measuring only 30.5mm x 110mm x 4.38mm, the drive also improves space utilization and scaling options in hyper-scale datacenter servers.



View at TechPowerUp Main Site

 

[silentbogo]

Holy crap. It means that in two or three GPU generations we might have an 8GB GTxx30 w/ 128-bit mem. bus in SFF package!
...or I might be wrong and it'll go into PS6 or XBox^2

 

[BadFrog]

8 TB NGSFF SSD? You mean is available because they announced back in August about a 16 TB NGSFF SSD being sampled...

 

[Imsochobo]

Holy crap. It means that in two or three GPU generations we might have an 8GB GTxx30 w/ 128-bit mem. bus in SFF package!
...or I might be wrong and it'll go into PS6 or XBox^2

I think it'll go into geforce and budget stuff where space is not an requirement.
Either Nvidia goes HBM this next round or after that, there is no particular good reason for GDDR over HBM and many reasons for HBM over GDDR.

HBM is not what made vega bad or expensive or w/e.

 

[efikkan]

This looks good for consumer Volta, which is coming next year.

 

[Assimilator]

I thought GDDR5/X already operates at 1.35V?

 

[silentbogo]

HBM is not what made vega bad or expensive or w/e.

It's what made Vega hard to manufacture. Plus, all the CoWoS kinks haven't been worked out yet, and we don't know what those differences in packaging from several manufacturers will do over time.

I thought GDDR5/X already operates at 1.35V?

GDDR5 works at 1.5V, but supports 1.35V low-power operation at slower data rates.
GDDR5x according to electrical specs is pretty much the same as GDDR6 will be, e.g. 1.35V Vdd (supply voltage), and 1.8V Vpp (pump voltage). Even their speeds are quite similar, so the only difference is a totally reworked memory interfacing.

 

[Assimilator]

It's what made Vega hard to manufacture. Plus, all the CoWoS kinks haven't been worked out yet, and we don't know what those differences in packaging from several manufacturers will do over time.

Difficult to manufacture, and contributed to a higher defect rate due to good Vega chips being mated with faulty silicon interposers that caused the whole thing to have to be trashed.

GDDR5 works at 1.5V, but supports 1.35V low-power operation at slower data rates.
GDDR5x according to electrical specs is pretty much the same as GDDR6 will be, e.g. 1.35V Vdd (supply voltage), and 1.8V Vpp (pump voltage). Even their speeds are quite similar, so the only difference is a totally reworked memory interfacing.

Thanks - so if Navi follows the pattern of Fiji and Vega (big, hot chip) then we'll almost certainly see AMD sticking with HBM and its cost and complexity. Although they may have licensed EMIB from Intel as part of the recent deal, which would hopefully give better yields and thus make the pain of HBM less.

Heh, it's kinda ironic that the only reason HBM made its way into the consumer space is due to AMD's inefficient GPU designs...

 

[Steevo]

Difficult to manufacture, and contributed to a higher defect rate due to good Vega chips being mated with faulty silicon interposers that caused the whole thing to have to be trashed.



Thanks - so if Navi follows the pattern of Fiji and Vega (big, hot chip) then we'll almost certainly see AMD sticking with HBM and its cost and complexity. Although they may have licensed EMIB from Intel as part of the recent deal, which would hopefully give better yields and thus make the pain of HBM less.

Heh, it's kinda ironic that the only reason HBM made its way into the consumer space is due to AMD's inefficient GPU designs...

They can scan and check quality while the die, interposer, or memory is still on the wafer, some may have defects not found during testing, but very few. AMD is getting very high yields with everything due to a mature process, and the fact the interposer is made on at least one node size larger for better EC and lower defects.

 

[ppn]

So then we can expect it in 2 years or 1 generation from now. Remember when GDDR5 6Gbps was announced we had to skip GTX460 the next in line and wait GTX 660 in order to get it in consumer products.

 

[oxidized]

there is no particular good reason for GDDR over HBM and many reasons for HBM over GDDR.

HBM is not what made vega bad or expensive or w/e.