kirankara78
Member
The power draw would be nuts.
Plus Amd been heard saying xfire is dead recently.
Not impossible but sounds unlikely
-
Hey Guest. Check out your NeoGAF Wrapped 2025 results here!
Radeon RX Vega thread
- Thread starter PFD
- Start date
The power draw would be nuts.
Plus Amd been heard saying xfire is dead recently.
Not impossible but sounds unlikely
They gave Vega a card that they designed ages ago, that was the better part of a year late because of hbm problems, good mining capabilities because they have a time machine and saw that etherium mining would explode? That's an...interesting take on things.
llien
Banned
Talking about x2 cards, 295x2 vs 1080.
Vega FE at Fury X clock performs nearly identical. It doesn't scale well with higher clocks, perhaps (likely mem starved), but "GCN not scaling well with more CUs" is a myth, based on wrong assumptions.
Vega FE at Fury X clock performs nearly identical. It doesn't scale well with higher clocks, perhaps (likely mem starved), but "GCN not scaling well with more CUs" is a myth, based on wrong assumptions.
icecold1983
Member
Didnt they reduce it to 32 SPs per SM or am i remembering wrong?
Unknown Soldier
Banned
This statement doesn't make sense either in the present or in the past because Bitcoin started out years ago being mined by GPUs and was only moved to ASICs when they proved cheaper and better at it. It only happens now that Ethereum is still mined by GPUs because it was specifically designed to defeat ASIC use.
None of what you wrote counters anything I wrote. I have no clue what you're talking about or who you're talking to.
dgrdsv
Member
+50% flops to current ones, spring 2018.
Basically, expect 1070 performance in 1060 price range. If we're lucky with the architecture then we may be looking at 1080 performance even.
thelastword
Banned
Wasn't the PCB for dual Vega leaked ages ago, I even think I heard AMD made mention of this before.... Inifinty fabric Vega here we come....
In any case, I don't think AMD needs dual vega to beat 1080ti, I think Vega 64 will come pretty close or match it with 1700+ clocks... Dual Vega is simply overkill at 25+ Teraflops. Last I read, Volta was 15TF....., so if Dual Vega lands later this year, AMD should be in the drivers seat..
RadioactiveLobster
Member
How exactly would AMD needing dual GPUs to beat a single GPU put them in the drivers seat?
frontieruk
Member
I think Vega's dual card still uses PLX to split pcie lanes, Navi is the introduction of infinity fabric to gpus, which might explain why they've killed crossfire as multigpu setups will be on card now which the PC will see as a single gpu.
Maintenance
Member
llien
Banned
With "infinity fabric" and some wishful thinking.
I would cautiously expect that to work for Navi, however.
Could you explain "they killed crossfire" bit, please?
Polaris does support it and so does Vega FE. Could find any word on Navi not supporting it.
PS
Wow, tweaktowns article from Jan 17 mentiones 300W TDP for Vega 10.
That is because of the Instinct slides showing the TDP for the Vega version would be 300W. IIRC those got release/leaked? in December last year.
frontieruk
Member
AMD giving up on crossfire
No the device would be seen as a single gpu like Ryzen cpus are seen as a single device.
Not a big loss if both companies move away from it.
thelastword
Banned
It's a single card pumping out 25 TFLOPS. If the process supercedes crossfire's implementation and of course 295x2, then it will not just only be the fastest card in name, but in execution and benchmarks....
It's the way of the future anyway, for more perf and power..Just take a look at how intel declared Threadripper as two ryzen's glued together, but look at it's performance...
Yeah, perhaps Navi will take it to a next level, but whatever implementation they have here with dual vega will be miles better than crossfire and 295x2...Also, I think AMD planned for dual vega's from the offset, so it will behave closer to a single GPU in execution.
It's crazy that such an old dual GPU card can still go toe to toe with a GTX 1080, beats it in certain titles and generally performs better at 4k...For what it's worth, the card only has 4GB of Vram, much lower core clockspeeds than the NV part, has the limitations of older dual GPU architecture and yet, the fact remains that games like GTA are all built around and favor Nvidia. I think it still pushes out great performance despite all those odds....
Coulomb_Barrier
Member
Not any more than an 8-core 1800X is recognized as two 4-core CPUs i.e it isn't.
I wonder if AMD has quite a jump on Nvidia with multi GPU on-die chip tech because of their Infinity Fabric. I wonder what solution Nvidia has for fast communication between separate GPU complexes.
I suspect that Nvidia will drag out the standard single monolithic GPU design whilst AMD will be first to market with multi-GPU cards with Navi because, as we all know, Nvidia like to play it safe whilst maximising profits and AMD enjoy risking their own annihilation with company direction.
The problem with dual GPUs is that they have awful frametimes, so while they might beat a single GPU is average framerate you'll often have to deal with microstutter.
But from what I gather to properly use a Ryzen CPU in games developers still need to program certain activities together within a single CCX.
Doesn't Nvidia have NVLINK? Or is that something different from infinity fabric?.
scoobs
Member
Ya.... I really don't think so
thelastword
Banned
Surely the design of this new GPU won't use older technology with these limitations.....
napata said:
NVLink is not necessarily for Dual GPU's, not saying there's no GPU to GPU connectivity or improvements there, but it's more so to bolster the link between CPU and GPU where they heighten the throughput there. NVlink is hardware though, yet, I don't see massive adoption for it.....
Well. we'll see...on the 14th..
kirankara78
Member
As far as I'm aware both are working on the technology but neither have perfected this and problems still remain in getting two chips to behave as one
scoobs
Member
All the early benchs have it competing with the 1070 and 1080, not even close to the 1080ti. Or are we talking about different things here? The 64 is priced at the 1080s price point, there is no world where it even approaches the 1080ti.
Not sure if posted, but GN contacted its industry sources, and they confirmed (but not yet tested themselves) that one of the Vega cards will do ~70 MH/s for mining. Which I believe is a good thing... for miners.
https://www.youtube.com/watch?v=qrnkbmabK4U
https://www.youtube.com/watch?v=qrnkbmabK4U
Coulomb_Barrier
Member
That's over double the 1080 Ti (31 mh/s). Kiss goodbye to hopes of getting one of these near launch.
Tallshortman
Member
There's a dual Asus card coming.
Hesemonni
Banned
Pretty much. I wonder if HBM2 production is still severely lagging behind the demand? I mean the way I understood that HBM2 is pretty much the production bottleneck at the moment and thus AMD can't even fully enjoy the demand.
Well, as the 64 WC uses the very same cooling solution as the FE WC, we have a pretty good idea of how much overclocking headroom there'll be. PCper managed to hit a stable 1,712MHz on the latter, but the card had trouble maintaining that speed consistently and the needle just moved from ~1070 performance to ~1080 performance. That actually aligns nicely with AMD's own insistence that the 64 will be competing with the 1080 as the boost clock of the 64 WC is only a stone's throw away: 1,677MHz.
Edit: Oh, hey, the 4K benchmarks include the 1080 Ti:
I excluded GTA5 and The Witcher 3 out of fairness as they overwhelmingly favoured Nvidia.
Shin
Banned
Samsung got them covered with their new fab, whenever it's ready for mass production.
Well, as the 64 WC uses the very same cooling solution as the FE WC, we have a pretty good idea of how much overclocking headroom there'll be. PCper managed to hit a stable 1,712MHz on the latter, but the card had trouble maintaining that speed consistently and the needle just moved from ~1070 performance to ~1080 performance. That actually aligns nicely with AMD's own insistence that the 64 will be competing with the 1080 as the boost clock of the 64 WC is only a stone's throw away: 1,677MHz.
Edit: Oh, hey, the 4K benchmarks include the 1080 Ti:
I excluded GTA5 and The Witcher 3 out of fairness as they overwhelmingly favoured Nvidia.
Yeah, the 1080Ti is in a league of its own sadly
Maintenance
Member
kirankara78
Member
Let's wait and see.
AMD were claiming cross fire was dead a week ago.
Would be a really dumb move in my opinion.
You would have a 600w toaster that doesn't work in so many games as support for multi gpu is dying
Starbuck2907
Member
To be clear, nvidia announced they were scaling back dual card support and AMD are simply following their lead. Neither company has officially cancelled support, but just said it will receive less attention going forward.
Unknown Soldier
Banned
Considering how little attention it already received, we're probably moving towards a world where SLI and Crossfire work for 3DMark and not much else. If you feel that's worth spending the money on 2 video cards then go for it I guess.
llien
Banned
Thanks.
That's quite different 'giving up on' from nvidia, 1060 outright does not support it.
Reasons for it might be totally different (i.e. it's developer's problem now, with DX12/Vulkan, I am not really into DX12/Vulkan details on the subject, though, just making things up)
Juggernaut chips are more expensive to develop and manufacture, it is the only feasible way for a R&D budget starved company, not really a choice, in my opinion.
If it is at 1080 levels at 1600Mhz (and that extra 550Mhz vs Fury X (remember, they are nearly identical at 1050Mhz, so that's a 52% higher clock) only gives it about 35% of extra performance), why would another 100Mhz on top of it suddenly push it to Ti levels?
That's over double the Vega FE as well, which is rather strange.
dgrdsv
Member
Both companies are following the market here, as with D3D11 and especially 12 and VK their ability to provide a good mGPU solution with drivers have highly diminished. 12/VK offloads this to the developers basically while giving them much better tools to implement it than whatever was available in 11. Most developers won't bother though.
The mGPU via Infinity Fabric thing so far smells a lot like another of these fan fiction videos. It may be viable beyond the process scaling point but until there's still a couple of scaling steps left a monolithic GPU will always win over any kind of "glued" approach. Ryzen example is irrelevant as CPUs are not GPUs.
Sir Abacus
Member
Ryzen also requires the scheduler to be aware of the CCX layout so that it doesn't send performance through the floor.
thelastword
Banned
We have not had any official benches yet, what we had were perception tests and leaks....The leaked benches...indicate that Vega 56 is at least 15-25% stronger than the 1070, that puts it within GTX 1080 range, so the extra CU's on the Vega 64 coupled with higher clocks on the WC version should pull it's weight way beyond 1080.
There are two things you need to realize;
1) The FE card does not have the proper drivers to take advantage of it's gaming capability, it's gaming driver is severely limited since they try to fuse it with the pro driver atm. It is just like Nvidia not having proper pro drivers for the Titans, but they got an uplift of 3x performance when they did. I have a feeling AMD will eventually release proper gaming drivers for the FE, especially since NV is boasting 3x pro performance and great gaming performance on Titans atm...
2.) There's still some kinks AMD has to resolve with overclocking et al,,,They wanted to rush the FE out the door to launch within the quarter and sort some issues after....Let's be fair here, Even Gamersnexus overclocked the FE to 1700Mhz, but in games the FE never reaches 1600Mhz core clock, it hovers around 1350-1400MHz. That coupled with non-gaming drivers for the FE can't be doing it any favors...
As has been said many times, the RX Vega is a gaming card, it's the card we should look to under proper benches to get proper perspective. So far, the leaks are promising, so there's that. So I'm not sure how pegging FE's performance to cards with proper gaming drivers is a good indicator at all, not with all I've specified above.
Of course Vega may come out and not be as good as the leaks specified and by extension the 64's, but the information out there is more indicative of the gaming cards being good than underperforming. I especially shun the idea that everything is conclusive with Gaming Vega's performance based on FE or against NV. We need to give everything a fair shake...The only way we could say conclusively that NV is better is when we have benches come the 14th....I'm leaning towards AMD on this one....
icecold1983
Member
the information out their is overwhelmingly indicative that it trades blows with a 1080
Sir Abacus
Member
So what you're saying is that Vega may have a blast of processing that is hitherto unknown?
I finally got around to updating the OP with news from SIGGRAPH
I think you're setting yourself up for a huge disappointment. Sure, we cannot say conclusively that the 1080Ti is better, but it is most likely going to be the case. Even AMD's own numbers seem to point in this direction
I think you're setting yourself up for a huge disappointment. Sure, we cannot say conclusively that the 1080Ti is better, but it is most likely going to be the case. Even AMD's own numbers seem to point in this direction
Yea, AMD's has in no uncertain terms said the Vega 64 is a 1080 competitor. At best expect slightly better. If they were holding back a magical power boost that put it above 1080ti levels their marketing would be all over it.
That's nothing for AMD to be ashamed of though it's a product competitive on the high end at both price and performance with a working implementation of HBM2. The water cooled variant (all their official numbers are based on the air cooled one) will probably nip at the ti's heels in optimized games.
The problem amd has is that people will just buy a 1070 and OC it 1080 levels, and these Vega chips are all gonna be scooped by miners anyway. Or people will just wait on Volta if they have been waiting for Vega this long onky to find out they could have just bought a 1070 eons ago.
franklyshankly
Member
Why is it a problem for AMD if miners buy up every Vega they make?
llien
Banned
That would be Fury X vs 980Ti season two, but this time AMD's product is expected to have quite a perf lead.
Also, adaptive sync, cough. Those $200 extra.
If they can make a lot of them, good, but if they are supply constrained and prices rise above msrp, that's a problem. Amd could be selling a lot more 580s right now if they had the supply to meet the mining demand. They don't. I don't see how they will have the supply for Vega either.
From everything we've seen it really doesn't look you'll get much better than an ocd 1070, even with an ocd non wc 64. So yeah, they are trying to sell a $500 card that isn't meaningfully better than a $400 card, and it consumes literally twice the power.
Delusibeta
Banned
Plus when the mining bubble properly crashes, the market will be saturated by cheap used cards, and prices will fall.