Please drop the HBM2 stuff for Raven Ridge. Would it be great? Yeah, I'd be all over it. Mark Papermaster, you know the CTO of AMD has said a few times that the cost structure isn't there. They have the designs, but not the price. Will we have Ryzen cores, Vega Gpu and HBM2 this year. No. No way, no how. 2019? Possible. Give that a 50/50. 2020, more likely.Does anyone think Raven Ridge with 2GB of HBM + dual DDR4 controllers would be a good idea? The Vega iGPU would use HBCC to control caching between main memory and faster HBM.
I think HBCC is already working in Vega, but you don't see a benefit unless the workset is bigger than VRAM.
Large-scale rolluout are always slow, besides EPYC currently lacks racks from major OEMs.
The memory bandwidth is definitely going to be a bottleneck. RX 550 has a similar shader throughput with ~3x the memory bandwidth and the 50W TDP includes 4GB GDDR5 so it's not as apples to oranges as it might seem at first glance.
As for HBM, using it instead of DDR for system RAM would actually be preferable for designs where the RAM chips are soldered anyways, if not for the fact that it costs an arm an a leg right now.
That's a bit worrying about the lack of LPDDR support. Although how much power does this really save? And does LPDDR come in higher speeds (2400) which is kinda necessary to get decent performance out of the GPU. I thought LPDDR's real advantage is during sleep mode, not really during idle use so worst case we may have to go back to hibernation for long periods of inactivity which kind of sucks, although not as bad since SSD's become standard. My Zenbook can easily sleep for a week and still a few hours left of active use, this almost instant sleep/on mode would be missed.
This has been repeated so often as to become fact. The modern world in microcosm.You could be right about the memory bottleneck although it seems hard to know for sure until you see how far it throttles under-load. I'd assume that the 550 would be able to sustain something closer to max clocks longer but who knows. As for HBM vs DDR as main system memory, regardless of price I don't think you'd ever want that. HBM is high bandwidth, high latency, which would be devastating for CPU performance. If you have a laptop with an embedded GPU, you've already made a major design decision that indicates that GPU performance is at best the third most important attribute (behind at least CPU and size/thermals/battery life).
This has been repeated so often as to become fact. The modern world in microcosm.
A Wccftech article using official stats.
https://wccftech.com/jedec-publishes-hbm2-specifications-scale-8hi-32gb-stacks-1-tbs-bandwidth/
tRC:
DDR3 = 4x-5x ns
GDDR5 = 40ns(1.5V) - 48ns(1.35V)
HBM1 = 48ns
HBM2 = 45ns
Comparable to DDR3, so latency is not one of the problems in using HBM2 as main memory.
As far as I know tRC isn't really a good/proper measure of latency. Again, I'm not hardware/EE type but my understanding is that tRC is more or less a physical property of the DRAM cell which is more or less the same in any memory technology made with the same process (which I think is why they are comparing to DDR3 as it's the most similar to GDDR5?). A more relevant number for DDR4 latency is CAS latency (although not perfect) and for DDR4-2400 that seems to range from 12.5-15ns (this is actually slightly higher than DDR3). I don't know what the comparable number would be for HBM. The big advantage of HBM is that it is physically close to the die which means you can do things that are impractical or impossible with less power than you could otherwise, in this case a REALLY wide memory interface. HBC (Hybrid Memory Cube) is another memory technology that is similar to HBM that is more compute focused targeting HPC/data center use. All of this to say that it probably is possible to use a different memory technology that has both lower latency and higher bandwidth. That technology may even be HBM, although I'd still bet against it, and since the Infinity Fabric is tied to memory clocks it probably won't be soon.
Since HBM2 and interposer are expensive, why isn't AMD just copying intel and putting on some edram to cache the most common access. Seems to bring some huge benefits on intels iGPU.
aka "It's also expensive but in a different way"Because Intel can manufacture eDRAM chips, and AMD can't.
I never thought I would see an Intel CPU with HBM powered graphics before an AMD CPU with HBM...
In a way, AMD contributed to Intel stealing their thunder, by teaming up with Intel to deliever an Intel HBM-GPU solution.
Since HBM2 and interposer are expensive, why isn't AMD just copying intel and putting on some edram to cache the most common access. Seems to bring some huge benefits on intels iGPU.
I hope that AMD eventually does announce a higher-end APU with HBM2.
Please drop the HBM2 stuff for Raven Ridge.
I wish I could find that white paper again, but supposedly the memory controller can handle HBM already. Mark Papermaster has said it's something they want to do, and will do when the cost is there. It just isn't there yet.I've heard the die doesn't feature an HBM2 controlller. Not having an interface for communicating with HBM2 is a very good reason for why HBM2 won't be used for Raven Ridge.
Also, HBM2 has been only featured in very high end products. Further, none of Nvidia's consumer line has them. The GPU portion would have to be several times larger to justify using HBM.