- Status
I would had though they would only implement HBM2 on 980 equivalent and above, and considering the trend is similar between NVIDIA and AMD (although other poster does not want to acknowledge this) in that they both only seem to have low power/efficient/smaller GPUs.
All the rumour about the 1080 seems to have some anomalous details *shrugs*.
Kinda sucks that this may end up going like Intel and their recent CPU releases (for both GPU manufacturers); disappointing and require much more waiting.
Cheers
Jawed
Legend
980 performance or 980 replacement?
Blazkowicz
Legend
HBM2 needs be dual-sourced, so could it be that we're waiting for high volume/availability for both manufacturers?
It needs to dual sourced to get prices down. I don't think they'll mix and match vendors during production.
They sell full interposer/memory combo. That adds a significant additional qualification burden for a relatively new technology (with all kinds of problems that might result.) For a relatively small volume part, it may not be worth doing that twice.
Jawed
Legend
Who is "they"? Fury cards come with a UMC interposer and Hynix memory.
https://beta.finance.yahoo.com/news/umc-enters-volume-production-tsv-100000122.html
Why have a JEDEC standard if chips aren't interchangeable?
https://beta.finance.yahoo.com/news/umc-enters-volume-production-tsv-100000122.html
Why have a JEDEC standard if chips aren't interchangeable?
Spoken like somebody who has never experienced RMAs due to some assembly house swapping one JEDEC compliant DRAM for another to save a few cents.
(And that's not for state of the art fastest speed possible DRAMs.)'They' is AMD or Nvidia.
There is a reason why every major silicon vendor has an approved component vendor list: minimum specs like JEDEC alone can't possibly cover all the issues that can come up something as complex as a high speed component. But at least you have the power to throw the PCB back to the integrator with a curt message saying "use our approved components."
Something like HBM probably solves some of that by handing over control from the system assembly house back to AMD/Nvidia (they own the interposer assembly), but with that the (financial) responsibility as well. There is absolutely no way they'll swap stacks without strenuous testing. An HBM from Hynix will (not 'may') have subtle differences from an HBM chip from Samsung.
Maybe there are certain operating frequencies where the stars align in a bad way such that you get occasional errors. Maybe those only happen for AMD and not for Nvidia due to some power supply quirks etc. Maybe the one vendor has some corner case bugs that the other doesn't have.
None of those will fatal, but they'll required slightly different settings one way or the other.
There was a time when Nvidia sold the DRAM alongside the GPU silicon. That hurt their margins, but it was one way to control these kind of issues.
Blazkowicz
Legend
Hopefully you can make a spin of the GPU (i.e. GP-xxx A2 or A3) that works with the different set of interposer/memory in addition to the first one?
Or failing that, a revision ( GP-xxx B1)
I agree it may be a high expense for a low volume gaming part, but when we get to high volume upper-mid range surely that may get justified. (could be later, such as Volta GV104)
If HBM2 is on GP100 only.. That's the high margin part with perhaps highish demand (HPC, engineering work etc.), even if it's just for Tesla and GTX Titan. Large memory sizes too (32GB for Tesla, 16GB for Titan) and a lot of validation will be spent on it anyway.
If the "staged release" idea has any value I think they would likely do it here. Over the course of an unspecified number of monthes if needed.
It sucks but GTX 960 has shown quite amazing performance with only 128 bit gddr5 bus. Could be that you can get about GTX 980 / Radeon Nano levels of performance with even a 128 bit gddr5X interface, at $299.99. GP-104 / GTX 1080 has 256bit, 8GB, more like 980 Ti / Fury X at $499.99? Boring, but not a bad kind of boring. If you already have a big GPU from the last few years I can understand there's nothing to see here move along..
Heck there is some sort of plateauing with game content too. A push to smaller PC may happen or continue, even today you can get a 65W CPU and a 120W GPU, or a 35W CPU (up to 6700T) and 75W GPU (watt-reduced GTX 950 variant) which go quite a long way.
Or failing that, a revision ( GP-xxx B1)
I agree it may be a high expense for a low volume gaming part, but when we get to high volume upper-mid range surely that may get justified. (could be later, such as Volta GV104)
If HBM2 is on GP100 only.. That's the high margin part with perhaps highish demand (HPC, engineering work etc.), even if it's just for Tesla and GTX Titan. Large memory sizes too (32GB for Tesla, 16GB for Titan) and a lot of validation will be spent on it anyway.
If the "staged release" idea has any value I think they would likely do it here. Over the course of an unspecified number of monthes if needed.
It sucks but GTX 960 has shown quite amazing performance with only 128 bit gddr5 bus. Could be that you can get about GTX 980 / Radeon Nano levels of performance with even a 128 bit gddr5X interface, at $299.99. GP-104 / GTX 1080 has 256bit, 8GB, more like 980 Ti / Fury X at $499.99? Boring, but not a bad kind of boring. If you already have a big GPU from the last few years I can understand there's nothing to see here move along..
Heck there is some sort of plateauing with game content too. A push to smaller PC may happen or continue, even today you can get a 65W CPU and a 120W GPU, or a 35W CPU (up to 6700T) and 75W GPU (watt-reduced GTX 950 variant) which go quite a long way.
I'm not suggesting at all that silicon spins are necessary.
Just programmable settings in the low level bowels of the memory controller that manages timings, sampler trimmers etc.
It's just that changes to these values may requires a drawn out qualification process that can be overly burdensome. Testing X number of samples from various speed lots, at different temperatures, voltages, for different workloads etc. These things can take a couple of weeks.
And when one of those settings fails, you need to isolate the problem and start all over again.
Jawed
Legend
I think you're right, generally.
But this should be in the design of the MC. Otherwise you've consigned a chip, which could be in the market for 2-3 years to a single supplier for a crucial component.
Blazkowicz
Legend
Most of my guesses have been horribly wrong on these threads but I believe they'd do that grueling work.
In fact with things slowing down at wide we've seen AMD and Apple fabbing integrated CPUs at competing fabs, one whole different version of the otherwise same thing twice.
In fact with things slowing down at wide we've seen AMD and Apple fabbing integrated CPUs at competing fabs, one whole different version of the otherwise same thing twice.
For a relatively low volume product, that's really not a big issue: your supplier will happily keep on delivering the necessary parts (because low volume is pretty high if you're a major GPU vendor), and if there's ever a real issue, there's always the possibility of switching suppliers after doing the qual work. Or threaten to do that.
What replaces the 980, so whatever card it is also has a performance increase rather than say just power/efficiency.
I do wonder if both AMD and NVIDIA will be able to fudge this due to the blurred ground of performance/efficiency, meaning they could use a lower tier replacement card and name that for say 980/390X - fingers cross they do not.
Cheers
At least earlier, while the bulk of the interposer came indeed from UMC, process of getting the interposer ready for chips and chips on to the interposer needed Amkor and ASE, too
Jawed
Legend
Well, HBM is exclusively graphics right now, isn't it? So the three companies that are making HBM have two companies to sell their stuff to.For a relatively low volume product, that's really not a big issue: your supplier will happily keep on delivering the necessary parts (because low volume is pretty high if you're a major GPU vendor), and if there's ever a real issue, there's always the possibility of switching suppliers after doing the qual work. Or threaten to do that.
The way I see it is, for example: if Hynix gets exclusivity on Vega 11 that doesn't preclude AMD's MC being able to work with Samsung or Micron for Vega 10. etc.. And similar for NVidia's chips.... At the very least.
It's too easy for both IHVs to play these companies off against each other.
Last edited:
3? Micron does HMC aka Hybrid Memory Cube, not HBM
Jawed
Legend
You're right. I was thinking of GDDR5 suppliers. It's a thin market, isn't it...
Yeah, it is
I was going to say there's more than 3 GDDR5 suppliers, but then double checked and it seems all the others have been aquired by others (Elpida is owned by Micron) or went bankrupt (Qimonda)
- Status