Return to

The VEGA 56 / 64 Cards Thread! General Discussion



With 13TFlops of single precision compute, it should be the top dog of anything short of Tesla V100. But AMD's driver team has repeatedly shown, they can't get as much as is possible out of AMD gpus...


Well the Fury X is a 8.3Tflop card but gets beaten by a Rx580 in quite a few tests (which is a 6.3Tflop card I believe).

In saying all that I don't understand why the Vega isn't getting close to 60MH/s in Ethereum mining, just weird that it only gets around 31..... I JUST DON'T GET IT


I guess that is because of 4GB vs 8GB? Because otherwise the fury should still be faster.


Got nothing to do with the memory really, most benchmarks don´t use over 4GB for this reason.


Woah wait what? I've never seen Polaris beat Fiji in anything, even power efficiency.


580 does in Tomb Raider and BF1 at 4k I believe. (FuryX)


I'm about to have a 580 to play with, I'll need to check that out for myself. Not that I don't believe you, this is just the first I've heard of it and I really want to investigate the claim.


Please report back. Would love to find out too.


Fury X gained up to 50% extra perf from going to Vulkan from OpenGL in doom, so I really don't see why VEGA can't gain a healthy amount from drivers that actually use the hardware features.
The idea that VEGA has less throughput clock for clock than Fiji does whilst being way larger is absolutely preposterous, VEGA was made to take the inefficiencies of the old GCN implementation away and take a somewhat similar approach to what maxwell did, though the shader pipeline on VEGA is more akin to something like Fermi with it's programmable floating point units.

Also I don't see RX 580s beating Fiji anytime soon, and if it does it's just due to driver related issues.


People must remember that gpu manufacturers are focusing more on the professional marked first.

I really want to see the performance of these vega cards in machine learning and how are they against nvidia cards.

And i think that right now amd need first to make ryzen, epyc, thread-ripper succeed and make them som money.


That is most likely because of video memory. But it might be an interesting video, @wendell? Fiji vs Polaris at settings that are actually used, like 1080p high?


Gamersnexus posted their video and article review, here are links if anyone is interested:

They tested a couple more things that PcPer didn't test, but the conclusions are basically the same.


YES OMG YES finally someone backs up what the fuck ive been saying for months now.. I am not crazy :smiley: seriously.. i have literally been saying that its possible AMD isnt making this a big performance card... i feel like they have something else in stock




I like the idea of cutting out one stack of HBM and improving the cooling (triple fan air cooler). Priced correctly it would make for a nice upgrade from the Fury line.


The Vega FE is a bit of a let down to me but all is not lost. At the moment you can't overclock is because there is a bug that drops HBM too 500MHz and it running slower. So we dont know how it overclocks.

It is running between a 1070 and 1080. If the custom gaming cards can come out with higher CPU / Memory / not a fucking blower cooler and solidly beat a 1080 and priced well I would seriously look at it for a purchase recommendation.

I have a RX480 so I most likely will skip the first gen because I am gaming well already. It will come down to price to performance gaming wise on the Vega RX cards.

And there is the looming RX mining card crash perhaps early next year. A cheap 2nd hand crossfire setup might put a dent in vega.


I suspect that the 500Mhz HBM is probably not a bug. More like a hobble


It only downclocks when they overclock though, that doesn't make sense. My guess is that there is some kind of software bug.


already confirmed to be a wattman issue, yes.


Pretty obvious bug.

Something that should have been pretty obvious within a minute or two of testing. It would seem to me that maybe the "bug" was left in accidentally on purpose.