Graphics going to GDDR 6 next year

Advanced Graphics DRAM
SK Hynix Develops World’s Fastest Graphics DRAM


"SEOUL,KOREA
24 April 2017 - 2:30pm
Cho Jin-young
SK Hynix developed a next-generation graphics dynamic random-access memory which has the world’s fastest graphics processing speed with lower power consumption. The latest DRAM will be used in a next-generation high-end graphics card to be released early next year.

SK Hynix announced on April 23 that it has developed the double data rate 6 (GDDR6) memory chip designed for graphic processors.

The latest 20-nano level 8 gigabit (Gb) GDDR6 can transfer data at 16 Gb per second (Gbps) per pin, which is by far the fastest speed offered in the industry. By utilizing 384 input and output (I/O) modules that are most commonly adopted by high-quality graphic cards, the memory chip can send graphic data of up to 768 gigabytes (GB) per second.

SK Hynix plans to start to mass produce the latest graphic memory chip early next year when its customers are expected to launch their next-generation high-quality graphic cards loaded with high-performance graphic DRAM chips.

The Graphics DRAM is a memory chip designed to speedily process videos and graphics upon demand from graphic cards in personal computers, work stations, video-playing devices, and high-performance game machines.

The GDDR6 is considered a next-generation and high-performance graphics DRAM that can offer twice the maximum speed of the existing GDDR5 while saving operating voltage by more than 10 percent. An international organization called “JEDEC” is now setting standards for the semiconductor.

SK Hynix expects that the GDDR6 will replace the market dominance of the GDDR5 and GDDR5X from next year."

http://www.businesskorea.co.kr/engl...-hynix-develops-world’s-fastest-graphics-dram

 

Heat generation

 

some back ground on Hynix

SK Hynix Posts Record Profit as It Pursues Toshiba's Chip Unit
by
Sam Kim
April 24, 2017, 4:52 PM PDT April 24, 2017, 7:12 PM PDT

SK Hynix Inc. posted its best-ever quarterly operating profit after strong demand for computer-server chips pushed up prices, boosting confidence in the Apple Inc. supplier’s bid for Toshiba Corp.’s semiconductor business.

Hynix, the largest provider of DRAM chips after Samsung Electronics Co., is targeting the troubled Japanese company’s flash memory division as growth in its main market tapers off. Demand for DRAM may “ease somewhat” in the second half as supply tightness eases, though the market will remain strong overall in 2017, President Kim Joon-ho said Tuesday during a call with analysts.

Operating profit was 2.47 trillion won ($2.2 billion) in the three months ended March, the Icheon, South Korea-based company said in a statement. That compares with the 2.25 trillion-won average of estimates compiled by Bloomberg. Sales were 6.29 trillion won, topping projections for 5.98 trillion won. Shares of Hynix fell 1.2 percent as of 11:10 a.m., and have nearly doubled in the past year.

“Memory prices are going strong well into the first half of this year,” Lee Jae-yun, an analyst at Yuanta Securities, said before the announcement. “The question is whether the momentum will stay strong in the second half, and what the potential impact would be on Hynix should it succeed in acquiring Toshiba’s unit down the line.”

Hynix is keen to expand its footprint in the faster-growing market for NAND chips used in everything from smartphones to connected appliances. The company however won’t let its bid for the Toshiba unit, which is expected to fetch more than 2 trillion yen ($18 billion), detract from a 7 trillion won investment in chipmaking capacity this year, Kim said.

Hynix and Samsung control more than two-thirds of the market for DRAM, commoditized memory employed in phones and personal computers. But Hynix ranks just fifth in NAND, needed in smartphones and tablet computers for functions from playing videos and multitasking to storing books and photos. If Hynix acquires Toshiba’s chip unit, its NAND market share would rise to more than 30 percent, based on 2016 data, overtaking Micron Technology Inc.

It would also get in a better position to serve Apple as it prepares to launch its much-anticipated 10th-anniversary iPhone this year. The U.S. company is Hynix’s largest customer, providing about 8 percent of its sales, according to data compiled by Bloomberg. Intel Corp. and Samsung also buy from Hynix, which on Tuesday said it expects 20 percent growth in demand for DRAM and 30 percent growth for NAND this year.

Prices of DDR3 4-gigabit DRAM chips rose to $2.98 at the end of March, up from $2.79 at the end of December, and 77 percent higher than a year earlier, according to data from In Spectrum Tech Inc. That helped propel the bottom line: net income for the quarter was 1.9 trillion won, topping analyst projections of 1.76 trillion won.

https://www.bloomberg.com/news/arti...s-beat-estimates-as-memory-chips-extend-rally

 

Yes, GPU memory is going to GDDR6 next year -- Samsung and Micron have also said that they'll switch over either in late 2017 or early 2018. However, it's not clear what the benefits from this are or whether this will be the standard for the next few years. The best graphics memory being used today is not GDDR at all, but HBM. HBM2 is available right now in the high end Quadros and the bandwidth there is significantly greater than even the Titan X variants. In fact, the 720GB/s of the Quadro GP100 is already nearly as fast as the projected maximum of GDDR6 (768GB/s from the article in the first post). We'll see.

 

Why is GDDR still being used?
Its ancient.
HBM should be taking over.

 

Gaming cards don't need HBM yet, why waste money on its more expensive manufacturing qualities?

 

once HBM manufacturing has become cheaper then GDDR will be phased out.

 

Call me when HBM comes to GPUs. We were supposed to see it in Maxwell, then it got pushed to Pascal, then now Volta or Vega.

 

Yeah but Samsung doesn't expect to mass produce HBM3 until 2019-2020.

Nvidia was wise to pursue GDDR6.

AMD has, thus far, gained no benefit from betting the farm on HBM.

 

Wise to pursue GDDR6? More like sly. It is likely that they chose GDDR6 to keep production volumes of HBM relatively low, and destroy AMD's game of Vega and Navi.

More production volume translates to lower unit cost.

As for no gains on HBM? Their Fury X and Fury Nano have performance on par with a GTX 1070, and that's nothing to sneeze at.

 

This was expected, It's no surprise long back the news came in 2016 August. The HBM2 made to the only GP100 chip non mass market, plus the HBM cannot be overclocked at all, Also the memory chip being closest to the die might factor in more heat to transfer to the heatsinks, Overclocking capabilites. This reason is alone enough to ignore HBM on the mobile market where we are now going reverse (Look at that Triton 700 - a gimped machine which banks on it's 19mm thickness with a GTX1080 lol, I bet it'd perform at 1070 level or worst depending on the temps) advertising the thinness & buzzwords over the performance stability and holding clocks to the optimal temps, sigh..(Side note - The power consumption is one of the advantage but AMD Fiji needs a powergrid to run it lol despite HBM, that's the reason AMD are on this path)

G6X will come too and iirc, afaik the G6 beats HBM by a huge margin, 512GBps of HBM vs G6 with 384bit at 760GBps speeds, Plus another point is the heatsink and OEM R&D costs, they always want to keep them as low as possible, the mobile market followed a standard design principles since a decade, but the desktop market is a whole another case, So changing that would mark a new era but at the expense of lot of resources doesn't seem plausible, this is a corporate pov ofc if they can cool it with more robust cooling and efficient machines with non BGA machines with more better power delivery systems will pan out more better for consumers for the servicing or upgrades from this point which might come showing us a new standard, but with caveats - that was all thought of when the MXM standards changed and the speculation at the future of the GPUs after Pascal, But let's save this talk for another day..

Bring it on !!

 

You dont need HBM with any graphics card that render games.
GDDR6 will not only be cost effective but will be readily available against HBM, which by the way requires a complete overhaul of existing PCB and package design. It requires changes to almost everyone involved which makes it a difficult transition when you output millions upon millions of graphic cards to the market.

448GB/s on a 256bit GDDR6 card should be more than enough when GTX 1080 offer 323GB/s with GDDR5X.
For 384-bit gamer cards with GDDR6 we are talking around 740GB/s. No way you will ever need more than that. Its overkill.

HBM is better in bandwidth than GDDR5X and GDDR6, but it belong on professional cards that actually needs the bandwidth and is produced in a lot less numbers than gamer cards for the average joe

 

Stop posting sensible facts. This is the internet.

 

See, I never said anything about the bandwidth. I was talking about PCB footprint, whereby HBM certainly wins out. Especially in the mobile gaming space where, well, space is at a premium, all the more we ought to be pushing towards HBM instead of using oddly-shaped MXM cards like Clevo, MSI and Asus are doing, just to fit desktop power into a mobile device. There also have been attempts to introduce low-cost HBM into the mass market. I, for one, would like to see an MXM-A GPU that is as powerful as a GTX 1070, for example, which cannot be done with the current GDDR standard.