This year is shaping up to be a good year for memory with the promise of 3D XPoint (Intel/Micron), HBM2 (SK Hynix and Samsung), and now GDDR6 graphics memory from Micron launching this year. While GDDR6 was originally planned to be launched next year, Micron recently announced its intentions to start producing the memory chips by the later half of 2017 which would put it much earlier than previously expected.
Computer World reports that Micron is citing the rise of e-sports and gaming driving the computer market that now sees three year upgrade cycles rather than five year cycles (I am not sure how accurate that is, however as it seems like PCs are actually lasting longer between upgrade as far as relevance but i digress) as the primary reason for shifting GDDR6 production into high gear and moving up the launch window. The company expects the e-sports market to grow to 500 million fans by 2020, and it is a growing market that Micron wants to stay relevant in.
If you missed our previous coverage, GDDR6 is the successor to GDDR5 and offers twice the bandwidth at 16 Gb/s (gigabits per second) per die. It is also faster than GDDR5X (12 Gb/s) and uses 20% less power which the gaming laptop market will appreciate. HBM2 still holds the bandwidth crown though as it offers 256 GB/s per stack and up to 1TB/s with four stacks connected to a GPU on package.
As such, High Bandwidth Memory (HBM2 and then HBM3) will power the high end gaming and professional graphics cards while GDDR6 will become the memory used for mid range cards and GDDR5X (which is actually capable of going faster but will likely not be pushed much past 12 Gbps after all if GDDR6 does come out this soon) will replace GDDR5 on most if not all of the lower end products.
I am not sure if Micron’s reasoning of e-sports, faster upgrade cycles, and VR being the motivating factor(s) to ramping up production early is sound or not, but I will certainly take the faster memory coming out sooner rather than later! Depending on exactly when in 2017 the chips start rolling off the fabs, we could see graphics cards using the new memory technology as soon as early 2018 (just in time for CES announcements? oh boy I can see the PR flooding in already! hehe).
Will Samsung change course as well and try for a 2017 release for its GDDR6 memory as well?
Are you ready for GDDR6?
Sorry for general
Sorry for general question/comment but I don’t understand why HBM is a “necessity” to hardcore gamers. I realise having fast memory is important for gaming but when the R9 Fury had HBM 1 the benefits of the extra bandwidth (to my knowledge) were not there to see with the 980 Ti beating out the Fury in alot (if not all) of situations. I own a Titan XP and as an experiment decided to underclock as well as overclock the G5X. When overclocked I saw minimal if any performance improvement but when underclocking only starting seeing degradation once the memory goes down approx 9Ghz. I guess in summation I would think a product like G6 would be more than we need for gaming with most of the superior bandwidth of HBM2 being lost just as with the R9 Fury X. I would very much love to hear all of the PC Per team’s opinion on this as I have never seen this discussed although it may have already been. P.S love your work guys. P.P.S Get Josh a new light 😉
Josh will get a new light the
Josh will get a new light the same day I actually have a proper room to stream from (i.e. not my basement).
one problem with HBM is they
one problem with HBM is they were very expensive because of it’s complexity. even AMD they only put HBM on their high end because it is not feasible cost wise with mid range and lower. personally i don’t agree with the notion about GDDR5/GDDR5X is more for mid range card and lower and HBM is needed for high end. as been proven by 980ti and titan xp both GDDR5 and GDDR5X is more than capable to keep up with the gpu. i can only see that benefit of that massive bandwidth in HPC class of application. sure AMD try to make something new with HBM in regards how HBM will handle memory in the future but it will not going to solve the problem on HBM high cost.
HBM2 will come down in price
HBM2 will come down in price once more companies begin using it in their products and Intel has just announced a new AI co-processor SKU that will use 32GB of HBM2. IF Intel begins using HBM2 then maybe even Intel will begin making its own HBM2 from one of its mothballed Chip Fabs. The more HBM2 units sold the more of the initial costs of HBM2 R&D and plant costs can be amortized at a faster pace to allow for the HBM2 price to come down.
Supply and Demand issues will always be a problem, just look at DRAM pricing currently with the switch to a new 20nm node production creating an initial shortage in the DRAM market.
HBM2 would probably be of better use for APUs built on an interposer with a Zen/Vega APU using at minimum a 4GB single stack of HBM2 as a last level Cache for its integrated graphics texture store to allow the APU’s graphics to work from HBM2 mostly. For a mobile laptop APU HBM2 can provide the high effective bandwidth that integrated APU graphics has greatly lacked in the past and HBM2 used as cache can help if it’s used in addition to a larger pool of regular DDR4 DRAM(Even single channel DRAM).
Cache memory systems where created with the purpose of creating a smaller sized but faster and lower latency pool of memory to hide any lower bandwidth/higher latency issues in a larger pool of slower memory. So HBM2 may be of better use for APUs/SOCs with HBM2 used as a last level cache to a larger pool of DDR4 DRAM(Single or Dual Channel).
That High Bandwidth Cache Controller/memory controller IP in Vega will be able to manage any texture staging above a 4GB store of HBM2/cache to the larger pool of slower DDR4 DRAM in the background and keep the integrated graphics working mostly from the higher bandwidth HBM2.
Kawalski,
A lot of hardware
Kawalski,
A lot of hardware changes like HBM get introduced before the benefits can be properly utilized. The same thing is happening with the new memory changes VEGA.
New API’s like DX12 and Vulkan with game code to properly swap data into the appropriate memory space for best utilization (size vs speed) is on a learning curve.
I think we need another year or two of software development to get a more clear benefit of how all this is going to shake out.
(and it’s worth noting that a high-speed buffer like eSRAM in the XBOX ONE has both a pro and a con. It can reduce the overall COST by mixing a high-speed buffer with cheaper, higher-capacity memory but it’s more hassle to PROGRAM. You need to ensure the data being processed is actually WITHIN that high-speed buffer.
With varying memory configurations it might be too difficult to optimize something like a VEGA design so we might settle on just managing the VRAM to system memory data swapping.
The advantages of HBM2 within this framework or uncertain. Will the latency and bandwidth benefits make more than a small difference?)
Beg only one question: Is it
Beg only one question: Is it pin compatible with GDDR5X?
most likely not. AFAIK not
most likely not. AFAIK not even GDDR5 and GDDR5X are pin compatible with each other.
As far as I know: no.
As far as I know: no.
GDDR5X also uses 190 pins per
GDDR5X also uses 190 pins per chip; (190 BGA).[4] By comparison, standard GDDR5 has 170 pins per chip; (170 BGA).[5] It therefore requires a modified PCB.
Initially gddr6 will be more
Initially gddr6 will be more competitive in terms of pricing vs hbm, it seems hbm is limited to enterprise/professional and enthusiast market initially, it’s rumored that AMD’s Vega with hbm 2.0 will still be inferior to Titan x Pascal based on gddr5x but will likely give a run for its money in low level APIs.
* rumored to inferior to
* rumored to inferior to 1080ti as well which is not based on hbm!
Yeah, I can believe that, I
Yeah, I can believe that, I have a feeling that Vega is not going to steal the performance crown but there will still be plenty of good cards that will come out and the HPC market will be happy with them as well as gamers since we will finally have more price competition at the upper mid to high range.
re VEGA:
AMD is in a weird
re VEGA:
AMD is in a weird spot where their architecture seems more likely to improve relative to NVidia’s.
Part of the issue is that AMD has NOT had the best drivers, but with DX12/Vulkan arriving and the game CONSOLES having similar AMD GPU’s they are really pushing the creation of low-level code (i.e. intrinsic shaders etc) and of course just coding properly for ASync Compute etc will gain some of the untapped performance.
NVidia conversely isn’t “bad” with DX12. That’s a misconception based on benchmarks seeing relatively poor or no gains in DX12 vs DX11. Part of the reason for that was early drivers, but also the fact that NVidia was doing a good job of ensuring their GPU was utilized well.
FREQUENCY:
AMD’s Polaris (RX-480 etc) was started several years ago and meant to be a MOBILE GPU. Plans changed, but the architecture was optimized for lower frequencies.
Suddenly AMD had to shift gears to compete with NVidia so they had to take Polaris and produce it for desktop. That’s the main reason the frequencies are a lot lower than NVidia’s. I say this because VEGA was likely intended for desktop, plus they’ll have had more time to optimize for the higher frequencies.
If we increased the GPU frequency of an RX-480 by 25% to be similar to what NVidia can do they would get about the same performance as the Fury X (and roughly 80% of a GTX1070, and even got close to matching it when titles optimize really well for AMD).
Now NVidia can obviously find ways to improve performance as well so we’ll see.
GSYNC/GSYNC 2 vs Freesync/Freesync 2 is unfortunately a divisive area too and frankly GSYNC 2 will IMO be the superior product since NVidia can optimize at a hardware level. Pixel overdrive, and HDR latency on high-res/high-refresh monitors will be very challenging to do and I think starting from SCRATCH with a dedicated module is really the right decision.
AMD is trying to push VEGA’s memory optimizations however I think it will take some time before games can truly benefit from this. I still think game developers will focus on the middle ground of simply optimizing swapping between system and video memory without fine tuning it for a potentially superior AMD solution.
Increasingly smaller, but faster memory pools happen all the way up from bulk storage, to system memory, to video memory, then buffers/caches on the GPU.
The XBOX Scorpio is actually dropping the eSRAM buffer. It’s a high-speed buffer to compensate for the slower (cheaper) shared 8GB of system memory of the XBOX ONE. The problem is that game developers need to ensure the working data is sitting in that buffer and that proved to be challenging. In fact, many games had stutters when the buffer filled up so game code was read direct from the slower DDR3 memory.
My point is that CHANGE IS SLOW, though there’s nothing about VEGA’s memory system that looks problematic, just that I don’t know if it will be easy to take full advantage of.
Vega’s memory optimizations
Vega’s memory optimizations take place below the game/game level of software and likely below the level of graphics APIs. These GPU processor memory optimizations work on the same level that the memory/Virtual memory and cache memory systems work for CPU processors and OS(Ring 0). The Vega IP(HBCC/memory controller) is very well able to treat HBM2 as just another level of cache as well as handle its own virtual memory paging files.
Do not kid yourself a GPU processor’s memory/cache memory subsystems are way more robust than a CPU’s systems that handle the same types of tasks. There are current and previous generation AMD PRO WX GPU SKUs that can handle their own virtual memory and have the hardware ability in the GPU itself to divide a single physical GPU into several independently operating logical GPUs that can be assigned to work servicing individual applications or service virtual OS instances operating under a hypervisor. That memory and virtualization IP is not totally new to AMD’s line of GPUs but the Vega IP takes things to a new level of performance regarding the way that the Vega will handle levels of Cache and Memory all the way down the tiers from top level GPU cache to HBM2/video memory and system DRAM on down to memory paged to SSD or hard drive. RTG knows what they are doing and so does AMD.
“GSYNC/GSYNC 2 vs
“GSYNC/GSYNC 2 vs Freesync/Freesync 2 is unfortunately a divisive area too and frankly GSYNC 2 will IMO be the superior product since NVidia can optimize at a hardware level. Pixel overdrive, and HDR latency on high-res/high-refresh monitors will be very challenging to do and I think starting from SCRATCH with a dedicated module is really the right decision.”
Maintaining a standard connection is the right decision. The display controller can handle such things perfectly fine. It will just take a bit of time for the display controllers to catch up with the new features. Proprietary display connections just needs to go away. Nvidia spent some research dollars to make it work without really changing the standard when the right thing to do is just change the standard.