⋮    ⋮  

Micron To Commence Mass Production of GDDR5X Memory in Summer – Early Samples Achieve 13 GB/s Data Rates


Micron has officially released an update on the status of their GDDR5X memory solution. Last month, JEDEC published the specifications for the upcoming GDDR5X memory which will double the rate of data flow over the existing GDDR5 memory at around the same power envelope. Being a cheaper solution compared to the HBM (High-Bandwidth Memory) design, we might be looking at several next generation graphics cards that make use of the updated memory standard.

Micron Commences GDDR5X Mass Production in Summer 2016 - Early Samples With 13 GB/s Data Rates

Micron has officially stated that GDDR5X memory will be going under mass production in summer, this year. This suggests that the yields of the new DRAM will be high and early samples are already achieving speeds of 13 GB/s compared to the 14 GB/s or higher that it is expected to hit when the memory launches.

Micron is currently ramping GDDR5X to mass production, and will be announcing sample dates later this spring. We plan to be in full volume production this summer. I also hope to give you an update later this year on performance specifications. via Micron

Feature Benefit
• QDR mode • Highest data rate in the industry – >13Gbps achieved to date
• VDD/VDDQ 1.35V • Improved power per bit
• 190-ball FBGA Package, 0.65mm pitch • Smaller outline – improved electrical performance

For those who like to know what difference is between GDDR5 and GDDR5X, we know from the official specifications published by JEDEC the both memory standards are not a whole lot different from each other but they aren't the same thing either. The GDDR5X solution is built upon the DNA of GDDR5 and has been updated to deliver twice the data rate. There are a lot of design changes that went in developing GDDR5X to achieve the faster transfer speeds, higher bandwidth and in a package that consumers just around the same power or even lower. Micron states that GDDR5X uses around 70% of the energy that is consumed by GDDR5.

The GDDR5X memory supports a 64-bit wide memory access that has double the prefetch against GDDR5 which supports a 32-bit wide memory access. This allows the memory to offer double the bandwidth and at speeds of 1.75 GHz, the GDDR5X memory will achieve data rates of 14 GB/s that will offer bandwidth of 448 GB/s across a 256-bit memory interface. Currently, the fastest GDDR5 solutions like the GTX 980 Ti can achieve a theoretical bandwidth of 336 GB/s while having access to a large 384-bit bus interface clocked at speeds of 1.75 GHz.

Micron’s GDDR5X program is in full swing and first components have already completed manufacturing. We plan to hit mass production this summer. The team at our Graphics DRAM Design Center in Munich, Germany is doing a fantastic job, too. Not only do we have functional devices earlier than expected, these early components are performing at data rates of more than 13Gb/s! Memory components mature as they move through the development and manufacturing process, so to see first silicon performing at nearly full performance specs was a pleasant surprise—these early results are incredibly promising. Our first generation GDDR5X is an 8Gb (1GB) density manufactured on our 20-nanometer process technology. via Micron

Compared to GDDR5 which can push speeds of up to 8 GB/s data rate per DRAM, the GDDR5X has speeds of up to 14 GB/s. Currently, Micron has achieved their first breakthrough of 13 GB/s on the 20nm process technology which was used to design a 8 Gb (1 GB) density package. Micron officials however believe that due to such good speeds delivered on early samples, their GDDR5X solution has the potential to achieve speeds even beyond 14 GB/s which is currently the maximum data rate suggested by them.

Based on the results so far, we believe that GDDR5X has the clear potential to achieve speeds of 14Gbps and potentially beyond. via Micron

Will Discrete Graphics Cards From NVIDIA/AMD Utilize GDDR5X?

Now the main question which everyone has in their minds is that who will utilize the new GDDR5X memory solution. The most obvious contender are discrete graphics cards built by both AMD and NVIDIA. Both companies are on the verge of launching 16/14nm based graphics cards codenamed Polaris and Pascal. Of all the samples which have been showcased, we haven't seen a HBM based solution as of yet but we can't really be sure since the samples are in early engineering phase.

In fact, NVIDIA and AMD both have confirmed in their slides that HBM2 is going to be incorporated in their next generation cards and knowing that NVIDIA is going to have 1 TB/s bandwidth on their Pascal chips which is something only possible with HBM2, we will definitely be looking at HBM2 cards. The HBM standard remains a vital solution for the high-end cards, offering up to 1 TB/s of bandwidth. Even the current HBM1 solution offers more bandwidth than GDDR5X (512 GB/s vs 448 GB/s). But, GDDR5X isn't necessarily meant to compete against HBM but both solutions are meant to exist for a range of discrete solutions.

Image Credits: NVIDIA @ Flickr

Since the last two generations, NVIDIA has offered a 256-bit wide bus solutions on their G**04 GPUs (GK104 and GM204). If cost of the GDDR5X DRAM is marginal, then we would see NVIDIA make a switch to GDDR5X and the switch would be easy since the command protocols for the memory is same. The GDDR5X solution is also a valid option for entry to mid-range offerings so that the higher-end could make use of the HBM standard which makes more sense. Cards such as the Radeon R9 390 and R9 390X offer great value and offer up to 8 GB of GDDR5 VRAM with bandwidth of 384 GB/s. Compared to the flagship Fiji based cards, the R9 390 series are often sighted as more better value. This would be the edge that GDDR5X might have over HBM1 and HBM2 however if mass production is expected to begin in Summer 2016, then we might get to see actual GDDR5X products in the second half of 2016.

At SC15, NVIDIA pointed out that HBM is a great memory architecture which will be implemented across Pascal and Volta chips but those chips have max bandwidth of 1.2 TB/s (Volta GPU). Moving forward, there exists a looming memory power crisis. HBM2 at 1.2 TB/s sure is great but it adds 60W to the power envelope on a standard GPU. The current implementation of HBM1 on Fiji chips adds around 25W to the chip. Moving onwards, chips with access of 2 TB/s bandwidth will increase the overall power limit on chips which will go from worse to breaking point. A chip with 2.5 TB/s HBM (2nd generation) memory will reach a 120W TDP for the memory architecture alone, a 1.5 times efficient HBM 2 architecture that outputs over 3 TB/s bandwidth will need 160W to feed the memory alone. GDDR5 and GDDR5X themselves face increased I/O power when compared to HBM.

This is not the power of the whole chip mentioned but just the memory layout, typically, these chips will be considered non-efficient for the consumer and HPC sectors but NVIDIA is trying to change that and is exploring new means to solve the memory power crisis that exists ahead with HBM and higher bandwidth. In the near future, Pascal and Volta don’t see a major consumption increase from HBM but moving onward in 2020, when new GPUs are expected to arrive, we will see introduction of a much more efficient memory interface for the discrete graphics solutions.

While HBM in comparison to GDDR5 saves space and consumes less power, an interposer design is needed to house the memory stacks and GPU die. A lot more development is required to make a fully functional HBM product and that could be one main reason why we would see cards beneath the high-end stuff based on GDDR5 designs. The HBM2 designs are also entering mass production in early 2016 as announced by Samsung and will allow GPU makers to develop their flagship cards for HPC and Consumer markets. It will be interesting to see the new standard being incorporated in new graphics cards when they launch later in 2016.