Micron Pushes GDDR5X To 16Gbps, Expects To Launch GDDR6 In Early 2018

Subject: Memory | June 7, 2017 - 01:02 AM |
Tagged: micron, gddr6, gddr5x

JEDEC made the GDDR5X memory standard official almost a year and a half ago where it launched at 10 Gbps and quickly hit 12 Gbps. Set to bridge the gap between GDDR5 and the upcoming GDDR6, the “G5X” standard is quickly catching up to and matching the speeds that GDDR6 will run at.

Specifically, Micron’s Graphics Design Team in Munich was able to achieve an impressive 16 Gbps in their high speed test environment. The team was able to hit 16 Gbps on a “meaningful sampling” of its mass production GDDR5X silicon which makes the feat much more impressive as it means these higher speeds are moving closer to reality than theory. Micron measured a PRBS11 (psuedorandom binary sequence) pattern read at 16 Gbps using an oscilloscope and also showed off a chart that compared the stable data rate timing margin versus data rate from 10 Gbps to 16 Gbps.

Micron GDDR5X.png

In addition to teasing the 16 Gbps memory speed (it will be awhile yet before we see products like graphics cards running memory at those speeds), Micron announced that it expects to being mass productions of GDDR6 chips in early 2018. GDDR6 will see a new (larger) FBGA1180 package, faster base sort speeds (GDDR6 will start at 12Gbps vs G5X's 10Gbps), and moving to a dual channel approach with channels that will have half as many I/O links (GDDR5X is x16/x32 while GDDR6 will be x8/16 per channel). It will be interesting to see how this move will stack up to G5X, but in theory Micron will be able to push clocks even higher (maybe even higher than 16 Gbps) by having more but simpler channels (and it may be easier for graphics card manufacturers to wire up their cards to the memory chips.

SK Hynix, who showed off its first GDDR6 chip at GTC, appears to be following the same I/O design as Micron with two channel memory at x8 or x16 per channel.

Are you ready for faster GDDR5X? Hopefully these new faster G5X chips come out soon to give AMD and NVIDIA a more appealing alternative to HBM and HBM2 for mid-range and high end consumer graphics cards since High Bandwidth Memory seems to still be suffering from limited supply and is holding the GPU guys back on being able to crank up the production lines!

Also read:

Source: Micron

New AMD Polaris 10 and Polaris 11 GPU Details Emerge

Subject: Editorial, Graphics Cards | May 18, 2016 - 01:18 PM |
Tagged: rumor, Polaris, opinion, HDMI 2.0, gpu, gddr5x, GDDR5, GCN, amd, 4k

While Nvidia's Pascal has held the spotlight in the news recently, it is not the only new GPU architecture debuting this year. AMD will soon be bringing its Polaris-based graphics cards to market for notebooks and mainstream desktop users. While several different code names have been thrown around for these new chips, they are consistently in general terms referred to as Polaris 10 and Polaris 11. AMD's Raja Kudori stated in an interview with PC Perspective that the numbers used in the naming scheme hold no special significance, but eventually Polaris will be used across the entire performance lineup (low end to high end graphics).

Naturally, there are going to be many rumors and leaks as the launch gets closer. In fact, Tech Power Up recently came into a number of interesting details about AMD's plans for Polaris-based graphics in 2016 including specifications and which areas of the market each chip is going to be aimed at. 

AMD GPU Roadmap.jpg

Citing the usual "industry sources" familiar with the matter (take that for what it's worth, but the specifications do not seem out of the realm of possibility), Tech Power Up revealed that there are two lines of Polaris-based GPUs that will be made available this year. Polaris 10 will allegedly occupy the mid-range (mainstream) graphics option in desktops as well as being the basis for high end gaming notebook graphics chips. On the other hand, Polaris 11 will reportedly be a smaller chip aimed at thin-and-light notebooks and mainstream laptops.

Now, for the juicy bits of the leak: the rumored specifications!

AMD's "Polaris 10" GPU will feature 32 compute units (CUs) which TPU estimates – based on the assumption that each CU still contains 64 shaders on Polaris – works out to 2,048 shaders. The GPU further features a 256-bit memory interface along with a memory controller supporting GDDR5 and GDDR5X (though not at the same time heh). This would leave room for cheaper Polaris 10 derived products with less than 32 CUs and/or cheaper GDDR5 memory. Graphics cards would have as much as 8GB of memory initially clocked at 7 Gbps. Reportedly, the full 32 CU GPU is rated at 5.5 TFLOPS of single precision compute power and runs at a TDP of no more than 150 watts.

Compared to the existing Hawaii-based R9 390X, the upcoming R9 400 Polaris 10 series GPU has fewer shaders and less memory bandwidth. The memory is clocked 1 GHz higher, but the GDDR5X memory bus is half that of the 390X's 512-bit GDDR5 bus which results in 224 GB/s memory bandwidth for Polaris 10 versus 384 GB/s on Hawaii. The R9 390X has a slight edge in compute performance at 5.9 TFLOPS versus Polaris 10's 5.5 TFLOPS however the Polaris 10 GPU is using much less power and easily wins at performance per watt! It almost reaches the same level of single precision compute performance at nearly half the power which is impressive if it holds true!

  R9 390X R9 390 R9 380 R9 400-Series "Polaris 10"
GPU Code name Grenada (Hawaii) Grenada (Hawaii) Antigua (Tonga) Polaris 10
GPU Cores 2816 2560 1792 2048
Rated Clock 1050 MHz 1000 MHz 970 MHz ~1343 MHz
Texture Units 176 160 112 ?
ROP Units 64 64 32 ?
Memory 8GB 8GB 4GB 8GB
Memory Clock 6000 MHz 6000 MHz 5700 MHz 7000 MHz
Memory Interface 512-bit 512-bit 256-bit 256-bit
Memory Bandwidth 384 GB/s 384 GB/s 182.4 GB/s 224 GB/s
TDP 275 watts 275 watts 190 watts 150 watts (or less)
Peak Compute 5.9 TFLOPS 5.1 TFLOPS 3.48 TFLOPS 5.5 TFLOPS
MSRP (current) ~$400 ~$310 ~$199 $ unknown

Note: Polaris GPU clocks esitmated using assumption of 5.5 TFLOPS being peak compute and accurate number of shaders. (Thanks Scott.)

Another comparison that can be made is to the Radeon R9 380 which is a Tonga-based GPU with similar TDP. In this matchup, the Polaris 10 based chip will – at a slightly lower TDP – pack in more shaders, twice the amount of faster clocked memory with 23% more bandwidth, and provide a 58% increase in single precision compute horsepower. Not too shabby!

Likely, a good portion of these increases are made possible by the move to a smaller process node and utilizing FinFET "tri-gate" like transistors on the Samsung/Globalfoundries 14LPP FinFET manufacturing process, though AMD has also made some architecture tweaks and hardware additions to the GCN 4.0 based processors. A brief high level introduction is said to be made today in a webinar for their partners (though AMD has come out and said preemptively that no technical nitty-gritty details will be divulged yet). (Update: Tech Altar summarized the partner webinar. Unfortunately there was no major reveals other than that AMD will not be limiting AIB partners from pushing for the highest factory overclocks they can get).

Moving on from Polaris 10 for a bit, Polaris 11 is rumored to be a smaller GCN 4.0 chip that will top out at 14 CUs (estimated 896 shaders/stream processors) and 2.5 TFLOPS of single precision compute power. These chips aimed at mainstream and thin-and-light laptops will have 50W TDPs and will be paired with up to 4GB of GDDR5 memory. There is apparently no GDDR5X option for these, which makes sense at this price point and performance level. The 128-bit bus is a bit limiting, but this is a low end mobile chip we are talking about here...

  R7 370 R7 400 Series "Polaris 11"
GPU Code name Trinidad (Pitcairn) Polaris 11
GPU Cores 1024 896
Rated Clock

925 MHz base (975 MHz boost)

~1395 MHz
Texture Units 64 ?
ROP Units 32 ?
Memory 2 or 4GB 4GB
Memory Clock 5600 MHz ? MHz
Memory Interface 256-bit 128-bit
Memory Bandwidth 179.2 GB/s ? GB/s
TDP 110 watts 50 watts
Peak Compute 1.89 TFLOPS 2.5 TFLOPS
MSRP (current) ~$140 (less after rebates and sales) $?

Note: Polaris GPU clocks esitmated using assumption of 2.5 TFLOPS being peak compute and accurate number of shaders. (Thanks Scott.)

Fewer details were unveiled concerning Polaris 11, as you can see from the chart above. From what we know so far, it should be a promising successor to the R7 370 series even with the memory bus limitation and lower shader count as the GPU should be clocked higher, (it also might have more shaders in M series mobile variants versus of the 370 and lower mobile series) and a much lower TDP for at least equivalent if not a decent increase in performance. The lower power usage in particular will be hugely welcomed in mobile devices as it will result in longer battery life under the same workloads, ideally. I picked the R7 370 as the comparison as it has 4 gigabytes of memory and not that many more shaders and being a desktop chip readers may be more widely familiar with it. It also appears to sit between the R7 360 and R7 370 in terms of shader count and other features but is allegedly going to be faster than both of them while using at least (on paper) less than half the power.

Of course these are still rumors until AMD makes Polaris officially, well, official with a product launch. The claimed specifications appear reasonable though, and based on that there are a few important takeaways and thoughts I have.

amd-2016-polaris-blocks.jpg

The first thing on my mind is that AMD is taking an interesting direction here. While NVIDIA has chosen to start out its new generation at the top by announcing "big Pascal" GP100 and actually launching the GP104 GTX 1080 (one of its highest end consumer chips/cards) yesterday and then over the course of the year introducing lower end products AMD has opted for the opposite approach. AMD will be starting closer to the lower end with a mainstream notebook chip and high end notebook/mainstream desktop GPU (Polaris 11 and 10 respectively) and then over a year fleshing out its product stack (remember Raja Kudori stated Polaris and GCN 4 would be used across the entire product stack) and building up with bigger and higher end GPUs over time finally topping off with its highest end consumer (and professional) GPUs based on "Vega" in 2017.

This means, and I'm not sure if this was planned by either Nvidia or AMD or just how it happened to work out based on them following their own GPU philosophies (but I'm thinking the latter), that for some time after both architectures are launched AMD and NVIDIA's newest architectures and GPUs will not be directly competing with each other. Eventually they should meet in the middle (maybe late this year?) with a mid-range desktop graphics card and it will be interesting to see how they stack up at similar price points and hardware levels. Then, of course once "Vega" based GPUs hit (sadly probably in time for NV's big Pascal to launch heh. I'm not sure if Vega is Fury X replacement only or even beyond that to 1080Ti or even GP100 competitor) we should see GCN 4 on the new smaller process node square up against NVIDIA and it's 16nm Pascal products across the board (entire lineup). Which will have the better performance, which will win out in power usage and performance/watt and performance/$? All questions I wish I knew the answers to, but sadly do not!!

Speaking of price and performance/$... Polaris is actually looking pretty good so far at hitting much lower TDPs and power usage targets while delivering at least similar performance if not a good bit more. Both AMD and NVIDIA appear to be bringing out GPUs better than I expected to see as far as technological improvements in performance and power usage (these die shrinks have really helped even though from here on out that trend isn't really going to continue...). I hope that AMD can at least match NV in these areas at the mid range even if they do not have a high end GPU coming out soon (not until sometime after these cards launch and not really until Vega, the high end GCN GPU successor). At least on paper based on the leaked information the GPUs so far look good. My only worry is going to be pricing which I think is going to make or break these cards. AMD will need to price them competitively and aggressively to ensure their adoption and success.  

I hope that doing the rollout this way (starting with lower end chips) helps AMD to iron out the new smaller process node and that they are able to get good yields so that they can be aggressive with pricing here and eventually at the hgh end!

I am looking forward to more information on AMD's Polaris architecture and the graphics cards based on it!

Also read:

I will admit that I am not 100% up on all the rumors and I apologize for that. With that said, I would love to hear what your thoughts are on AMD's upcoming GPUs and what you think about these latest rumors!

Report: NVIDIA GP104 Die Pictured; GTX 1080 Does Not Use HBM

Subject: Graphics Cards | April 22, 2016 - 10:16 AM |
Tagged: rumor, report, pascal, nvidia, leak, graphics card, gpu, gddr5x, GDDR5

According to a report from VideoCardz (via Overclock.net/Chip Hell) high quality images have leaked of the upcoming GP104 die, which is expected to power the GeForce GTX 1070 graphics card.

NVIDIA-GP104-GPU.jpg

Image credit: VideoCardz.com

"This GP104-200 variant is supposedly planned for GeForce GTX 1070. Although it is a cut-down version of GP104-400, both GPUs will look exactly the same. The only difference being modified GPU configuration. The high quality picture is perfect material for comparison."

A couple of interesting things have emerged with this die shot, with the relatively small size of the GPU (die size estimated at 333 mm2), and the assumption that this will be using conventional GDDR5 memory - based on a previously leaked photo of the die on PCB.

NVIDIA-Pascal-GP104-200-on-PCB.jpg

Alleged photo of GP104 using GDDR5 memory (Image credit: VideoCardz via ChipHell)

"Leaker also says that GTX 1080 will feature GDDR5X memory, while GTX 1070 will stick to GDDR5 standard, both using 256-bit memory bus. Cards based on GP104 GPU are to be equipped with three DisplayPorts, HDMI and DVI."

While this is no doubt disappointing to those anticipating HBM with the upcoming Pascal consumer GPUs, the move isn't all that surprising considering the consistent rumors that GTX 1080 would use GDDR5X.

Is the lack of HBM (or HBM2) enough to make you skip this generation of GeForce GPU? This author points out that AMD's Fury X - the first GPU to use HBM - was still unable to beat a GTX 980 Ti in many tests, even though the 980 Ti uses conventional GDDR5. Memory is obviously important, but the core defines the performance of the GPU.

If NVIDIA has made improvements to performance and efficiency we should see impressive numbers, but this might be a more iterative update than originally expected - which only gives AMD more of a chance to win marketshare with their upcoming Radeon 400-series GPUs. It should be an interesting summer.

Source: VideoCardz

Rumor: NVIDIA's Next GPU Called GTX 1080, Uses GDDR5X

Subject: Graphics Cards | March 11, 2016 - 05:03 PM |
Tagged: rumor, report, pascal, nvidia, HBM2, gtx1080, GTX 1080, gtx, GP104, geforce, gddr5x

We are expecting news of the next NVIDIA graphics card this spring, and as usual whenever an announcement is imminent we have started seeing some rumors about the next GeForce card.

GeForce-GTX.jpg

(Image credit: NVIDIA)

Pascal is the name we've all being hearing about, and along with this next-gen core we've been expecting HBM2 (second-gen High Bandwidth Memory). This makes today's rumor all the more interesting, as VideoCardz is reporting (via BenchLife) that a card called either the GTX 1080 or GTX 1800 will be announced, using the GP104 GPU core with 8GB of GDDR5X - and not HBM2.

The report also claims that NVIDIA CEO Jen-Hsun Huang will have an announcement for Pascal in April, which leads us to believe a shipping product based on Pascal is finally in the works. Taking in all of the information from the BenchLife report, VideoCardz has created this list to summarize the rumors (taken directly from the source link):

  • Pascal launch in April
  • GTX 1080/1800 launch in May 27th
  • GTX 1080/1800 has GP104 Pascal GPU
  • GTX 1080/1800 has 8GB GDDR5X memory
  • GTX 1080/1800 has one 8pin power connector
  • GTX 1080/1800 has 1x DVI, 1x HDMI, 2x DisplayPort
  • First Pascal board with HBM would be GP100 (Big Pascal)

VideoCardz_Chart.png

Rumored GTX 1080 Specs (Credit: VideoCardz)

The alleged single 8-pin power connector with this GTX 1080 would place the power limit at 225W, though it could very well require less power. The GTX 980 is only a 165W part, with the GTX 980 Ti rated at 250W.

As always, only time will tell how accurate these rumors are; though VideoCardz points out "BenchLife stories are usually correct", though they are skeptical of the report based on the name GTX 1080 (though this would follow the current naming scheme of GeForce cards).

Source: VideoCardz

Podcast #384 - Corsair Carbide 600Q, GDDR5X, a Dual Fiji Graphics card and more!

Subject: General Tech | January 28, 2016 - 01:38 PM |
Tagged: podcast, video, corsair, carbide, 600q, 600c, gddr5x, jdec, amd, Fiji, fury x, fury x2, scythe, Ninja 4, logitech, g502 spectrum, Intel, Tigerlake, nzxt, Manta

PC Perspective Podcast #384 - 01/28/2016

Join us this week as we discuss the Corsair Carbide 600Q, GDDR5X, a Dual Fiji Graphics card and more!

You can subscribe to us through iTunes and you can still access it directly through the RSS page HERE.

The URL for the podcast is: http://pcper.com/podcast - Share with your friends!

Hosts: Ryan Shrout, Jeremy Hellstrom, Josh Walrath, and Allyn Malventano

Subscribe to the PC Perspective YouTube Channel for more videos, reviews and podcasts!!

GDDR5X Memory Standard Gets Official with JEDEC

Subject: Graphics Cards, Memory | January 22, 2016 - 11:08 AM |
Tagged: Polaris, pascal, nvidia, jedec, gddr5x, GDDR5, amd

Though information about the technology has been making rounds over the last several weeks, GDDR5X technology finally gets official with an announcement from JEDEC this morning. The JEDEC Solid State Foundation is, as Wikipedia tells us, an "independent semiconductor engineering trade organization and standardization body" that is responsible for creating memory standards. Getting the official nod from the org means we are likely to see implementations of GDDR5X in the near future.

The press release is short and sweet. Take a look.

ARLINGTON, Va., USA – JANUARY 21, 2016 –JEDEC Solid State Technology Association, the global leader in the development of standards for the microelectronics industry, today announced the publication of JESD232 Graphics Double Data Rate (GDDR5X) SGRAM.  Available for free download from the JEDEC website, the new memory standard is designed to satisfy the increasing need for more memory bandwidth in graphics, gaming, compute, and networking applications.

Derived from the widely adopted GDDR5 SGRAM JEDEC standard, GDDR5X specifies key elements related to the design and operability of memory chips for applications requiring very high memory bandwidth.  With the intent to address the needs of high-performance applications demanding ever higher data rates, GDDR5X  is targeting data rates of 10 to 14 Gb/s, a 2X increase over GDDR5.  In order to allow a smooth transition from GDDR5, GDDR5X utilizes the same, proven pseudo open drain (POD) signaling as GDDR5.

“GDDR5X represents a significant leap forward for high end GPU design,” said Mian Quddus, JEDEC Board of Directors Chairman.  “Its performance improvements over the prior standard will help enable the next generation of graphics and other high-performance applications.”

JEDEC claims that by using the same signaling type as GDDR5 but it is able to double the per-pin data rate to 10-14 Gb/s. In fact, based on leaked slides about GDDR5X from October, JEDEC actually calls GDDR5X an extension to GDDR5, not a new standard. How does GDDR5X reach these new speeds? By doubling the prefech from 32 bytes to 64 bytes. This will require a redesign of the memory controller for any processor that wants to integrate it. 

gddr5x.jpg

Image source: VR-Zone.com

As for usable bandwidth, though information isn't quoted directly, it would likely see a much lower increase than we are seeing in the per-pin statements from the press release. Because the memory bus width would remain unchanged, and GDDR5X just grabs twice the chunk sizes in prefetch, we should expect an incremental change. No mention of power efficiency is mentioned either and that was one of the driving factors in the development of HBM.

07-bwperwatt.jpg

Performance efficiency graph from AMD's HBM presentation

I am excited about any improvement in memory technology that will increase GPU performance, but I can tell you that from my conversations with both AMD and NVIDIA, no one appears to be jumping at the chance to integrate GDDR5X into upcoming graphics cards. That doesn't mean it won't happen with some version of Polaris or Pascal, but it seems that there may be concerns other than bandwidth that keep it from taking hold. 

Source: JEDEC