Its not really a rumor that AMD is developing this standard, they announced the partnership with Hynix almost a year ago.
Whats up for debate is whether HBM is commercial ready.
Sounds like AMD has some huge plans in the works.
20nm should be huge for them, but increasing the memory speed on top of that will give them a huge advantage over the 900 series.
Which is why I believe the R300's is really facing against Nvidias 1000 series or whatever they call it. They'll move Maxwell to 20nm, maybe with the full GM200 chip. Nvidia just pulled a smart move and managed to push out a 10% gain on 28nm that also reduced power. Q1 2015 we'll see the big guns come out for both companies. In the meantime team green will be bringing in the cash with their success of the 900's.
Why? They are already the market leader. As a PC gamer interested in tech and decent pricing, it would be better if AMD trumped Nvidia again for a change.
390X is supposed to water cooled? http://wccftech.com/amd-radeon-r9-390x-arrives-1h-2015-feature-hydra-liquid-cooling/
Is this site reliable?
Ive heard this quite a few times in different areas if technology and have seen it not be the case. Basically, I'll believe it when the OEM announce it in a product.There is no debate, HBM is ready for production.
There are rumours that Nvidia won't be going 20nm and instead are looking at 16nm.
I think you misread my post. If NVidia's new architecture didn't massively outperform AMD's two and a half year old one, would that not be incredibly disappointing?
I bet someone's seen a prototype that just uses whatever highend cooling solution the engineers had available.A 400 card shouldn't be so complicated to install and a single chip card shouldn't actually need that much coolingassuming the rumour isn't completely made up
How would a closed loop cooler be hard to install? The 295X2 was extremely simple to install, more so than a CPU cooler.
I really don't see what you are talking about. This is a memory type. NVlink is an interconnect (and one primarily designed for HPC at that). These are entirely orthogonal.If they can get this to work with Standard PCI Express, they've won. Nvidia has been going around talking up their "NV Link" bullshit. AMD, if they can get their driver's shit together, should be able to offer some super compelling stuff going forward.
I really don't see what you are talking about. This is a memory type. NVlink is an interconnect (and one primarily designed for HPC at that). These are entirely orthogonal.
Didn't GlobalFoundries recently abandon their 14nm R&D and license Samsung's design?AMD to go to 14nm from Samsung?
Here's another interesting article. It could just be for AMD's CPU side, but why would they go to TMSC for 16finfet and Samsung for 14finfet.
I'm in.
The article reads that AMD will have a full year of HBM over Nvidia GDDR5... that is going to be very very nice for AMD.
This is why you don't rush out and buy overpriced Nvidia when it is announced.
This is why you don't rush out and buy overpriced Nvidia when it is announced.
Where does 9x faster come from?
My 970 has 224 GB/s bandwidth. 256 GB/s is not 5x that.
First generation of HBM is only 4.5x. It's 1GB/s I don't know where you're getting 256 GB/s. 2nd gen is 2 GB/s.Where does 9x faster come from?
My 970 has 224 GB/s bandwidth. 256 GB/s is not 5x that.
Where does 9x faster come from?
My 970 has 224 GB/s bandwidth. 256 GB/s is not 5x that.
First generation of HBM is only 4.5x. It's 1GB/s I don't know where you're getting 256 GB/s. 2nd gen is 2 GB/s.
It's per device/chip. They're comparing a 7.0Gbps GDDR5 chip (28 GB/sec total) to 1 HMB device that's going at 128-256 GB/sec. A GPU would have several of these allowing it to reach ~1TB/sec in BW.
The 1st Generation of HBM should have 4 GB and 128 GB/s per Stack. (4Hi)Where does 9x faster come from?
My 970 has 224 GB/s bandwidth. 256 GB/s is not 5x that.
It's per device/chip. They're comparing a 7.0Gbps GDDR5 chip (28 GB/sec total) to 1 HMB device that's going at 128-256 GB/sec. A GPU would have several of these allowing it to reach ~1TB/sec in BW.
It's per device/chip. They're comparing a 7.0Gbps GDDR5 chip (28 GB/sec total) to 1 HMB device that's going at 128-256 GB/sec. A GPU would have several of these allowing it to reach ~1TB/sec in BW.
At which point, it would saturate the PCI express bus.
I'm not sure I buy that this is gonna be used in the 390 and 380 gpus. One, AMD would have announced it by now if they made such a big breakthrough. Two, AMD would probably be focused on integrating this with their HSA APUs as PCI express would be a gigantic bottleneck since transferring the memory from system RAM to VRAM would be drastically slow.
At which point, it would saturate the PCI express bus.
I'm not sure I buy that this is gonna be used in the 390 and 380 gpus. One, AMD would have announced it by now if they made such a big breakthrough. Two, AMD would probably be focused on integrating this with their HSA APUs as PCI express would be a gigantic bottleneck since transferring the memory from system RAM to VRAM would be drastically slow.
I think it's pretty fair to assume that 512 GB/s is the target.
The interessting point will be, how much GB Fiji has to offer?
Intels Knights Landing will come with an MCDRAM solution (custom HMC) with 512 GB/s and 16 GB space.
So he says, Theres enough processing power on every laptop on the planet today, when AMD's best mobile CPUs are literally as slow as intel's i3 mobile parts.
Absurd.
At which point, it would saturate the PCI express bus.
I'm not sure I buy that this is gonna be used in the 390 and 380 gpus. One, AMD would have announced it by now if they made such a big breakthrough. Two, AMD would probably be focused on integrating this with their HSA APUs as PCI express would be a gigantic bottleneck since transferring the memory from system RAM to VRAM would be drastically slow.