-
Hey Guest. Check out your NeoGAF Wrapped 2025 results here!
Radeon RX Vega thread
- Thread starter PFD
- Start date
StrongBlackVine
Banned
Thanks. It seems to be mostly speculation from some Chinese website, but nothing really solid. I won't count on it, but more options would be welcome.
opticalmace
Member
If I had to guess, I would say early next year. I doubt amd is going to force nvidia's hand, so I imagine nvidia is happy to continue selling pascal products for a while, maybe with a slight discount. It's always a waiting game with PC components... Just buy a new card when you need it. You'll never get away from the fact that something better is coming a bit down the line.
Battlefield 1 this time
https://www.youtube.com/watch?v=0PrSpv6F9aE
Fan speed and power limit maxxed out to keep the cards at max speeds
Performs better than stock 1080 but it does get pretty hot (up to 82°C) and that's with the fan @100%
https://www.youtube.com/watch?v=0PrSpv6F9aE
Fan speed and power limit maxxed out to keep the cards at max speeds
Performs better than stock 1080 but it does get pretty hot (up to 82°C) and that's with the fan @100%
If i get one its going to be a aib fan speeds at 100% nothanks
Locuza
Member
PCGH published yesterday several results in regards to the Vega Frontiers Edition, including the Beyond3D Suite:
http://www.pcgameshardware.de/Vega-Codename-265481/Tests/Benchmark-Preis-Release-AMD-Radeon-Frontier-Edition-1232684/3/
A summary what can be observed and might be the case:
- ALU: The latency and FMA throughput is absolutely identical between Vega (GCN Gen 5) and the Fury X (GCN Gen 3).
That's pretty much conforming that at least for basic operations AMD didn't changed anything crucial in regards to the design.
- Polygons: The geometry engines might be improved.
GCN Gen 4 improved the performance for triangles organized as strips, that's why GCN Gen 3 (Fury X) doesn't perform well there.
Vega (Gen 5) shows basically the same performance between strips 100% culled and lists 50% culled.
It's not better per clock for strips but for lists in comparison to Polaris (Gen 4).
With the DX11 SDK sample the performance with Tessellation isn't better per clock than it was with the Fury X.
It's better in comparison to Polaris but the geometry engines might not be the only relevant factor here.
With strips Nvidia is far ahead, something where I expected improvements from Vega but nothing to see there.
- The pixel fillrate is more or less the same, maybe a bit better with Vega.
Which might be interesting because AMD re-organized the ROPs as clients of the L2$.
Well at least nothing seems to be broken...
- The texel fillrate shows strange results.
Vega achieves ~20% less or Fury X 25% more throughput.
The number of TMUs should be the same, one could speculate that AMD maybe changed something internally to save die area or increase the effiency but the overall throuhput would be rather low in comparison to the past and the competition, so I don't think AMD opted for decreased performance.
- Effective Texture Bandwidth
Now here things are getting really interesting.
That's a bandwidth test where two different types of textures are tested.
One black texture and one with random colors.
Since the recent GPUs are using Delta Color Compression techniques you see a big difference between a black texture where no color deltas are found and the compression can be optimal and a random colored texture where the compression effectively doesn't work.
Nvidia is quite the king here, the difference between the best case and worst case is about105-130% in bandwidth.
GCN Gen 3 only manages 17%, GCN Gen 4 47%.
One possible speculation was that Nvidias Color Compression might not be that much better than from AMD but that the tiled based renderer is helping the color compression technique in addition.
But the DSBR seems currently to be inactive as the trianglebin test doesn't indicate any tiling.
Without the DSBR the results are 52-60% better for GCN Gen 5 with the black texture.
The range is maybe too small to call it a clear improvement over GCN Gen 4, maybe a few percent.
What's more interesting are the results with the random colored texture where the achieved bandwidth is actually 24% lower than from the Fury X, you would expect 6% (484 GB/s vs. 512 GB/s) but not 24%.
The bandwidth utilization is miserbel and explains the limited scaling seen with Vega.
Another topic are the clockrates, the voltages and the power consomption.
The Vega GPU works very coarse, there are only a few power states and measuring the power consumption it's evident that Vega is really jumping around big consumption numbers.
There is the notion from an user with AMD contacts that the Adaptive Voltage Frequency Scaling isn't yet fully optimized and that ACG (he didn't said what the acronym stands for) isn't active at all.
https://www.forum-3dcenter.org/vbulletin/showthread.php?p=11421717#post11421717
So what's the conclusion?
Either the hardware has severe design issues or/and AMD launched the product with very unoptimized software support.
The optimistic hope is, that it "just" lacks good software.
Either way, quite terrible execution from AMD.
http://www.pcgameshardware.de/Vega-Codename-265481/Tests/Benchmark-Preis-Release-AMD-Radeon-Frontier-Edition-1232684/3/
A summary what can be observed and might be the case:
- ALU: The latency and FMA throughput is absolutely identical between Vega (GCN Gen 5) and the Fury X (GCN Gen 3).
That's pretty much conforming that at least for basic operations AMD didn't changed anything crucial in regards to the design.
- Polygons: The geometry engines might be improved.
GCN Gen 4 improved the performance for triangles organized as strips, that's why GCN Gen 3 (Fury X) doesn't perform well there.
Vega (Gen 5) shows basically the same performance between strips 100% culled and lists 50% culled.
It's not better per clock for strips but for lists in comparison to Polaris (Gen 4).
With the DX11 SDK sample the performance with Tessellation isn't better per clock than it was with the Fury X.
It's better in comparison to Polaris but the geometry engines might not be the only relevant factor here.
With strips Nvidia is far ahead, something where I expected improvements from Vega but nothing to see there.
- The pixel fillrate is more or less the same, maybe a bit better with Vega.
Which might be interesting because AMD re-organized the ROPs as clients of the L2$.
Well at least nothing seems to be broken...
- The texel fillrate shows strange results.
Vega achieves ~20% less or Fury X 25% more throughput.
The number of TMUs should be the same, one could speculate that AMD maybe changed something internally to save die area or increase the effiency but the overall throuhput would be rather low in comparison to the past and the competition, so I don't think AMD opted for decreased performance.
- Effective Texture Bandwidth
Now here things are getting really interesting.
That's a bandwidth test where two different types of textures are tested.
One black texture and one with random colors.
Since the recent GPUs are using Delta Color Compression techniques you see a big difference between a black texture where no color deltas are found and the compression can be optimal and a random colored texture where the compression effectively doesn't work.
Nvidia is quite the king here, the difference between the best case and worst case is about105-130% in bandwidth.
GCN Gen 3 only manages 17%, GCN Gen 4 47%.
One possible speculation was that Nvidias Color Compression might not be that much better than from AMD but that the tiled based renderer is helping the color compression technique in addition.
But the DSBR seems currently to be inactive as the trianglebin test doesn't indicate any tiling.
Without the DSBR the results are 52-60% better for GCN Gen 5 with the black texture.
The range is maybe too small to call it a clear improvement over GCN Gen 4, maybe a few percent.
What's more interesting are the results with the random colored texture where the achieved bandwidth is actually 24% lower than from the Fury X, you would expect 6% (484 GB/s vs. 512 GB/s) but not 24%.
The bandwidth utilization is miserbel and explains the limited scaling seen with Vega.
Another topic are the clockrates, the voltages and the power consomption.
The Vega GPU works very coarse, there are only a few power states and measuring the power consumption it's evident that Vega is really jumping around big consumption numbers.
There is the notion from an user with AMD contacts that the Adaptive Voltage Frequency Scaling isn't yet fully optimized and that ACG (he didn't said what the acronym stands for) isn't active at all.
https://www.forum-3dcenter.org/vbulletin/showthread.php?p=11421717#post11421717
So what's the conclusion?
Either the hardware has severe design issues or/and AMD launched the product with very unoptimized software support.
The optimistic hope is, that it "just" lacks good software.
Either way, quite terrible execution from AMD.
D
Deleted member 465307
Unconfirmed Member
That's interesting because we're currently in Q3. If true, I imagine more rumors would have to start popping up soon.
The one weird thing about that website's speculation is that they entirely focused on the x70 and above. No love for a x60? I need AMD and Nvidia to release RX x80 and GTX x60 replacements to help with the stock/price situation of the current lineup.
Shin
Banned
The x70 and x80 are hogging the spotlight as they'll launch first. If the 1060 is any indication, though, the wait for the x60 won't be long.
Yeah, after some more poking around, it looks like MyDrivers received info on the V100. That's Volta-based and launching in Q3, but designed for data centres. AMD can breath a small sigh of relief.
StrongBlackVine
Banned
This puts me in a bit of a bind. I don't want to wait until 2018 to buy a GPU, but I will not be happy if something more powerful and cheaper than what I bought releases a few months later. Seems that is how this works though.
I'm just looking at the gaming benchmarks and, as of now, maxed out @1600MHz it's still on average behind a stock 1080.
@Auto it's closer to the stock 1070.
Fury X = 56.8
1070 (stock ~1683) = 64.8
Vega FE (Auto/stock) = 69.6
Vega FE (maxed out @1600MHz) = 75.1
1080 (stock ~1733) = 79.6
Titan Xp (stock ~1582) = 100
New comparison video
1080 / Vega FE / 1080 Ti Ghost Recon Wildland @1440p
https://www.youtube.com/watch?v=5RfYH7wJTFE
tioslash
Member
Yeah, the waiting game can be tricky. I´ll wait until around October/November, and if there is no solid news on Volta and Vega is priced accordingly, I´ll probably go with Vega.
thelastword
Banned
Nice, temps are not bad, not bad at all..
opticalmace
Member
400W is nuts.
It's worse, 400 through the plugs. They power the GPU, the HBM is powered by the PCI-E socket, so total draw will be even higher. PCI-E slot can provide up to 75W.
dgrdsv
Member
It's water cooled under hybrid mod.
Keep in mind it's still power throttled, temps are hardly the issue. If you've got the cooling, aftermarket design that removes any power delivery roadblocks and 1kw power supply, you can probably get past the GTX1080ti.
Though you have to wonder if they'd be better off just releasing a Fat Polaris version to the mass public if power delivery is this troublesome.
Though you have to wonder if they'd be better off just releasing a Fat Polaris version to the mass public if power delivery is this troublesome.
Booker.DeWitt
Member
I am a noob, so excuse the question . Which Vega model is similar to a geforce 1080?what about 1070? Do we have one ? Any idea of pricing ?
StrongBlackVine
Banned
Supposedly slightly better performance(15 percent maybe) than a 1080 without overclocking. Rumored price I saw was $499 and lower tier $399 card to compete with a 1070. I'm a complete noob as well so take this with a grain of salt.
We wont know this until the end of the month when AMD show RX Vega.
icecold1983
Member
From what we know now, top end vega is 1080 level. We dont know anything about lower level vegas. Pricing is sheer guesswork but i cant imagine amd being able to charge more than 500 for this level of performance
Unknown Soldier
Banned
HBM is useful for extremely bandwidth-intensive compute applications. The relatively low availability and high cost of HBM isn't particularly material for those uses, which is why Nvidia uses HBM on the highest-end compute-oriented models of a generation.
But for consumers, low cost and high volume availability are paramount which is why Nvidia is still using GDDR5 and 5X on consumer models. Nvidia doesn't really have any issues with it because they have been so effective in keeping GPU core power draw highly efficient while also integrating bandwidth conservation measures like color compression and tile-based rendering.
Booker.DeWitt
Member
Top end is 1080 ? I saw some specs of the top Vega and it might cost around $999, so it should be way more powerful than a 1080.
I was expecting something like 1080 for $400 to500 and something like 1070 for $250-300. I can only dream.
The Vega Frontier edition is $1000 or $1500 for water cooled variant. This is meant for professional creators and not good value for a gaming PC. Tests on it so far have shown it performs in line with GTX 1080.
Best to wait until July 30th when AMD actually release RX Vega.
Lagspike_exe
Member
You won't be getting 1070 equivalent for $250 or $300. That's barely above RX580 MSRP without including price hikes due to mining rush.
GTA V doesn't like Vega FE
https://www.youtube.com/watch?v=vIdEWzFePl4&t=5s
https://www.youtube.com/watch?v=vIdEWzFePl4&t=5s
Hesemonni
Banned
Nobody likes Vega FE.
lol pretty much
michaelius
Banned
Well lets see.
Pcper sample was running at 1440 MHz on average which gave it around 10% less performance than 1080 in most case.
From other sources we know that increasing power target makes it keep closer to advertised 1600Mhz at which point it matches 1080 but we are already at 360Watts or so.
1080ti is 30-40% faster than 1080 so to match it Vega would need to run around 2GHz since power consumption increases linearly with clock we are at 450+ Watts level which is close to maximizing 3x 8 pin connector. But that's only if Vega would do it on stock voltage - which is extremely unlikely. So even if gpu clocked 30% higher we would be looking at 500-600 watt - so we are looking at 4x8pin connector.
icecold1983
Member
Power doesnt increase linearly with clock. 2ghz vega would draw alot more than 450
Felix Lighter
Member
What the hell kind of cooling would you need for a card that consumes 600W?
opticalmace
Member
Oh that's soon. Link ? Neat.
th3sicknness
Member
I think people are dismissing the fact that FE technically isnt using the full driver package for its game mode in all the tests (that wont be released until RX is out). Also AMD has stated on multiple occasions RX Vega will outperform the FE.
If the FE is already sitting between a 1070 and 1080 I cant imagine RX Vega being slower than that.
If the FE is already sitting between a 1070 and 1080 I cant imagine RX Vega being slower than that.
I thought we had to wait until the end of the month. You have a link or something where they will reveal the news?
icecold1983
Member
There are no missing driver components that affect performance
Its just a tour. Dont expect any actual info. Purely a hype machine
My guess is it'll be a bunch of gaming setups with Vega RX in them, maybe even Crossfire, and no framerate counts enabled.
Gamers get to have an early look, see how awesome it is, with no frame of reference, and the real reveals will be at sigraph.
Oh, I definitely don't think your typical RX Vega is passing the GTX1080ti.
Really, I don't disagree with you. I was sort of being an ass, half joking that you probably need a 1kw PSU and an aftermarket power delivery design to have any chance of passing a GTX1080ti.
In other news, Gamers Nexus tested their 1700mhz Vega with the hybrid mod. It still gets wrecked by the reference GTX1080ti and isn't convincingly crushing the GTX1080.
Frankfurter
Member
PCGH did a bit of undervolting for the Radeon Vega Frontier Edition (see update from July 13th)
http://www.pcgameshardware.de/Vega-...ase-AMD-Radeon-Frontier-Edition-1232684/2/#a5
Interestingly enough they were able to undervolt it to 1.075V (default 1.2V) and it still ran rockstable with the 1600 Mhz clockspeed.
http://www.pcgameshardware.de/Vega-...ase-AMD-Radeon-Frontier-Edition-1232684/2/#a5
Interestingly enough they were able to undervolt it to 1.075V (default 1.2V) and it still ran rockstable with the 1600 Mhz clockspeed.
Yeah this happened with Polaris too. Basically AMD have significant variance in chips and a lot are poorer, needing more voltage. So they have decided to set all GPUs produced to use a fixed voltage of 1.2v. This means the potatoes can manage the clocks required and they can sell more units. Unfortunately it means that the better chips get hotter sooner than would otherwise be necessary, limiting their top end performance.
Heat causes voltage leakage and so a hotter chip requires more power to reach the same performance as a cooler-running chip. In Gamers Nexus' tests with their hybrid water Vega FE, they found the same card needed 30W less to hold the same clock when the improved cooler was installed. This is due to voltage leakage. That's 30W reduction you could enjoy as a cooler running system, or swallow up by pushing performance higher.
DemonCleaner
Member
Unknown Soldier
Banned
This is what happens when your fab is GlobalFlounderies
Locuza
Member
PCGH did a variety of tests and comparisons points:
http://www.neogaf.com/forum/showpost.php?p=243224713&postcount=1006
Still, Ryzen if I am correct is also made by GloFo. And those chips don't have the issues that there seems to be with Polaris and Vega.
Vega is also 14nm compared to Polaris' 28nm, so they're a different process, and I think 14m was licensed from Samsung.
I think there's more to Vega's issues than GloFo.
Manp
Member
Polaris is 14 nm