![[H]ard|Forum](/styles/hardforum/xenforo/logo_dark.png){kind=logo}
On which titles? All I've seen was that it worked on the Specview Energy test which is somewhat meaningless and still doesn't mean they are optimally setting tile and bin sizes.
Not a tirade, I just repeatedly pointed out that a GCN core could provide similar functionality as the tensor cores on Volta. The change being a 32/64 wide wave running as a single thread with 16 wide vectors for the tensor. More of a software change than anything in hardware, but there are ways to make it more efficient.
Not bias, I just take a position outside that of various marketing narratives. I never said 480 as we know it would be up in 1070/80 territory. I said that a potential bonded version with HBM would be competing. Fiji does quite well in certain titles at higher resolution.
I didn't complain about the community, I said discussions about AMD have a habit of running off course with pro-Nvidia marketing lines from a handful of posters making it difficult to follow. As I said, a few ignores and the SNR goes up substantially so I've dealt with it. Still sucks for everyone else though.
Not sure it's even possible to increase primitive discard from Polaris. Now making the process more efficient as Mantor laid out in that video is another matter entirely. Primitive shaders would be a method to discard, or more appropriately not create, the triangles in the first place. That's what devs have been presenting SIGGRAPH papers on for the past few years. Not sure what the percentage of discarded triangles has to do with geometry throughput. In Linux drivers at the very least AMD has already merged the initial stages to facilitate some of those savings. Wouldn't be surprising if the driver is executing a limited "primitive shader" for all titles to achieve those savings and handle the binning with DSBR.
Has AMD actually said what ram is on RX? The 8-Hi stacks we're doing 2200MHz overclocks and 4-Hi could come faster stock.
I doubt it's the controller, but a thermal issue. Some tests of FE showed the memory running at thermal limits which would cause throttling. What isn't mentioned is that simply running hot under the limit will require more refresh cycles on the ram. Vega might be able to manage more memory bandwidth just by turning up the fans or water cooling at the same clocks, but nobody checked that.Or even if it is a side-effect of the HBCC<--->Memory Controller with some quirks in real world but then again that would apply to all.
I guess more waiting for review benchmarks
Cheers
AMD stated it's usefulness will be limited to resource limited GPUs, don't overhype something the manufacturer himself is not.
Yep just like you are saying a potential driver/features will improve Vega dramatically, we can safely assume this will go down the same way your prediction for 480 did, into thin air.
It is not just memory BW but issue with the cache, it is having issues in both cache and BW tests, I think it is more of a stretch to say it is a thermal issue rather than the cache<---->HBCC<---->Memory controller; the issue may be the implementation and structure with the HBCC.
Not sampling to review sites would be a pretty bad move, but certainly would show a considerable lack of confidence. I think Vega looks like a good 2016 card and that is all.
Same with not sampling to any site. You send the message you don't want consumers to make an informed purchasing choice. It's not uncommon to not sample Pro cards but pretty odd to not sample consumer GPUs.
Sell the Radeon division and dump all the R&D into their CPU which actually look to have a rosy picture right now. Radeon has to be a real drag on the R&D budget.
TFLOPs mean absolutely fuck all. This is a forum mostly dedicated to gamers, not HPC use, and personally I couldn't care less about any theoretical performance numbers, only how much they deliver in the real world.
Ok old guy lol. The 285 is referencing the Nvidia 285 not AMDs 285.
My bad, I'm apparently retarded.
Look, I understand what you are saying. I get it. And I apologize for misreading the TFLOPS post. What I have been saying is that in practical terms, AMD cards haven't really challenged NVIDIA in performance - in games - the past two generations (1000 series, 900 series), and certainly not in perf/watt. And in the generation priors to that, they were mostly "almost as fast but not quite", with the exception of the 7000 series on release. Meanwile, the common perception of AMD's drivers was that they were buggy and games were not as compatible (more or less).
no biggie. I was afraid you might be offended my old man jab. I am almost old, fighting the daylights out of it, not yet...
I asked around some friends I know, nobody that I know has Vega, despite them receiving ThreadRipper samples 10 days in advance for reviews. And the clock is ticking, 5 days to go now.
They could be dropping them late so that reviewers "had them" but not really enough time to post a review ahead of release. Dunno?
Reviewers are free to inform people they have the cards, they often do so frequently, for example all reviewers announced they have ThreadRippers in their hands way ahead of reviews. we are already seeing many leaks for ThreadRipper as well. But Vega is complete radio silence, no one has cards.
Could be, they could be planning to sample them just before launch, and then we have to wait for reviews several days later.
