Samsung has just announced that they have begun mass production of 4GB HBM2 memory modules. When used on GPUs, four packages can provide 16GB of Video RAM with very high performance. They do this with a very wide data bus, which trade off frequency for transferring huge chunks. Samsung's offering is rated at 256 GB/s per package, which is twice what the Fury X could do with HBM1.
They also expect to mass produce 8GB HBM2 packages within this calendar year. I'm guessing that this means we'll see 32GB GPUs in the late-2016 or early-2017 time frame unless "within this year" means very, very soon (versus Q3/Q4). They will likely be for workstation or professional cards, but, in NVIDIA's case, those are usually based on architectures that are marketed to high-end gaming enthusiasts through some Titan offering. There's a lot of ways this could go, but a 32GB Titan seems like a bit much; I wouldn't expect that this affects the enthusiast gamer segment. It might mean that professionals looking to upgrade from the Kepler-based Tesla K-series might be waiting a little longer, maybe even GTC 2017. Alternatively, they might get new cards, just with a 16GB maximum until a refresh next year. There's not enough information to know one way or the other, but it's something to think about when more of it starts rolling in.
Samsung's HBM2 are compatible with ECC, although I believe that was also true for at least some HBM1 modules from SK Hynix.
HBM/HBM2 is about to become
HBM/HBM2 is about to become less expensive at a faster pace, now if only some other market players will get their HBM processes going more devices could be using HBM. Nvidia still had to get an Interposer subcontractor for its interposer needs, so it may be more of the supply of interposers that keeps HBM from being more quickly adopted! Samsung needs to get an Interposer manufacturing line also. A lot of that HBM chip stacking and solder bump technology/IP will have to be licensed from third parties and Interposers are just bigger slabs of silicon, but the manufacturing capacity for more interposers will have to keep up or HBM will have no way to be wired up to the GPUs/APUs that will make use of HBM.
“HBM/HBM2 is about to become
“HBM/HBM2 is about to become less expensive at a faster pace, now if only some other market players will get their HBM processes going more devices could be using HBM.”
SK Hynix and Samsung are both producing HBM, Intel/Micron are working on HMC instead so likely never will. Gen 1 HBM was short-run, gen 2 may or may not be depending on process yields and customer demand. I’m not going to bet on price dropping anywhere close to GDDR5 levels anytime soon though.
“Nvidia still had to get an Interposer subcontractor for its interposer needs, so it may be more of the supply of interposers that keeps HBM from being more quickly adopted!”
It’s more that fabbing interposers on low-nm fab lines is throwing money away (because those fabs are expensive, and you really need very little detail for an interposer), so you want to find someone with an old high-nm fab with amortised costs and low margins to fab as physically large a chip as you can (reticule size limited) for as cheap as you can. Nvidia contracts out all it’s fab work anyway, so that’s not unusual.
I am not sure HBM and HMC
I am not sure HBM and HMC really compete in the same market. HMC will probably be a bit more expensive than HBM and will have higher power consumption. HMC seems more like an HPC product, not a consumer graphics card product. It would take quite a few links to reach graphics card levels of bandwidth. On the other hand, HMC can allow a much larger amount of memory than HBM since each package is a stack of memory die and the bottom logic die has multiple serialized links which allow them to be chained. The Wikipedia article indicates HMC with 4 or more links per stack and the possibility of chaining up to 8 stacks. The memory size made possible with HMC is much larger than what is required for the consumer market and the number of links required to compete directly with HBM on bandwidth would be too expensive for a consumer product also.
I am curious as to what Intel will use for their CPUs with integrated graphics processors. I don’t know if an on package cache (like Xbox One design) like they currently have with crystalwell will compete with an HBM based APU, if AMD actually delivers that. Some one posted about an Intel packaging tech called EMIB (embedded multi-die interconnect bridge) which may allow them to use HBM type memory without a silicon interposer. I don’t know if they actually will use HBM though. HMC signaling would be unsuitable to using with EMIB or silicon interposers. Hopefully they don’t invent their own type of memory instead.
Some one posted about the
Some one posted about the interposers before. They are currently made on a 65 nm process. I don’t remember whether that was Global Foundries or some other company though. The 65 nm node isn’t exactly in high demand at the moment.
Gimme some HBM APU`s that
Gimme some HBM APU`s that dont require system RAM!
And then put those into freesync IPS screen laptops and Zboxes with NO dedicated gpus.
So who’s doing the final work
So who’s doing the final work of putting this together. There is an interposer, the HBM1/2 memory and the processor. Does it all come together at TSMC/GF or does someone have the job of bringing it all together and doing the final verification testing for each chip?
Apu’so with this tech would
Apu’so with this tech would benefit greatly. In my opinion, AMD is getting a serious push from Samsung in a unwritten partnership… This is the time for AMD to gain ground in the marketplace
What product would Samsung be
What product would Samsung be interested in silicon interposers for? Samsung is working on Wide I/O, which is a similar JEDEC standard to HBM, but it is meant for much lower power and it is meant to stack on top of other devices. It may also be used on an interposer at some point, which makes it similar to HBM. If Samsung uses such devices in their phones, even if it is on an interposer, it will probably be with Samsungs own processors and their own Wide I/O memory.
I would expect most laptop makers to be interested in an HBM APU though. If you combine such an APU with m.2 for storage, you can make a very powerful system with an absolutely tiny foot print. Also, if you make full use of the interposer, you could add other chips onto the device besides memory and processing which could shrink the size even further. Such devices would fit very well with Apple’s designs. Apple switched to using all AMD GPUs a while ago, but this doesn’t necessarily indicate anything. Apple has switched back and forth several times in the past. Apple’s only choice for CPUs right now is Intel. I doubt that they like that.