AMD RX Vega’s pricing was “not just for launch, but ongoing”

[ltguy]

15 minutes ago, LAwLz said:

Oh look, some reddit post from a user who is constantly attacking Nvidia and defending AMD claims to have inside information which just so happens makes AMD seem like the good guys, and retailers/miners seem like the bad guys.

What a surprise, right?

 

I mean, who wouldn't trust someone who said Vega would perform better than a 1080, and cost 400 dollars.

A person who said that he had already sold his current GPU and since he was heavily invested in FreeSync he would be ordering a Vega.

A person who, once benchmarks for the Vega FE were released, parroted the same bullshit things about "it's the drivers! It doesn't have tile based rasterization" and all the other things he probably doesn't understand the first things about, just to keep the engine in the hype train going for a bit longer.

 

I think it is sad that people just leech onto any positive comments they can find on Reddit. It's as if all you need to do on Reddit these days is to claim to have insider information, write a really long post that is positive towards AMD, and then it will get reposted over and over and over again until people become certain that it is true.

Breathe in Breath out.... 3. 2. 1... 1. 2. 3..

 

I think the only reason people were paying attention was because they had a picture of an invoice. I personally dont care which of the 5 people in the supply chain is responsible for the price hikes.

[ltguy]

1 minute ago, Notional said:

I doubt they can do much. The bottleneck is called HBM2, and the fact that the manufacturers (SK Hynix + Samsung) seems to have issues manufacturing them.

They also have no influence on actual prices at retailers. Not only do they not sell them directly to the retailers (the vendors like Asus, MSI, etc. do), but they can't dictate prices either, as that would break cartel laws in most countries.

Mining sucks, HBM2 sucks and the resulting supply and prices suck.

This! HBM was what ruined Fury cards. 4gb of vram was never enough for the intended purpose of that card. So to see them double down on stupid and launch another gaming card with hbm its frustrating for a long term amd fan like myself. RX 480/580 were good cards, they should have just scaled those up untill vega and hbm2 was more feasable.

[Agosto]

Why is everyone pointing at HBM2? The big issue are internal bottlenecks, just like Fiji, and the lack of fully exploiting the new added features.

AMD has to solve SP count scaling, since it's clearly there from Hawaii era. Vega 56 and 64 perform almost the same at same frequencies... and oddly sape IPC as Fiji in games, but much higher on other applications.

[ltguy]

1 minute ago, Agost said:

Why is everyone pointing at HBM2? The big issue are internal bottlenecks, just like Fiji, and the lack of fully exploiting the new added features.

AMD has to solve SP count scaling, since it's clearly there from Hawaii era. Vega 56 and 64 perform almost the same at same frequencies...

Were not blaming hbm2 for performance problems... were blaming it for availability and pricing (bom cost) problems

[Agosto]

1 minute ago, ltguy said:

Were not blaming hbm2 for performance problems... were blaming it for availability and pricing (bom cost) problems

If Vega performed as it should (1080Ti+ levels), price would not have been a problem.

[Notional]

4 minutes ago, Agost said:

Why is everyone pointing at HBM2? The big issue are internal bottlenecks, just like Fiji, and the lack of fully exploiting the new added features.

AMD has to solve SP count scaling, since it's clearly there from Hawaii era. Vega 56 and 64 perform almost the same at same frequencies... and oddly sape IPC as Fiji in games, but much higher on other applications.

ROPS are very much tied to HBM. On 1080ti, you can see it's missing a VRAM module, as the correlating vram controller part is disabled. You can't do that with HBM2, as you only have 2 modules. That makes Vega 64 performance limited. The entire architecture is bandwidth starved. Having 4HBM modules would have been better, but impossible due to HBM yields an availability. But we are mostly talking supply issues, not performance.

1 minute ago, Agost said:

If Vega performed as it should (1080Ti+ levels), price would not have been a problem.

Yes, because demand would be a LOT higher compared to the available supply of HBM modules. As a result the price would skyrocket. This is basic supply and demand market functions.

[Agosto]

7 minutes ago, Notional said:

Yes, because demand would be a LOT higher compared to the available supply of HBM modules. As a result the price would skyrocket. This is basic supply and demand market functions.

Vega shouldn't need 800 GB/s to perform well, that's the point. There must be some internal bottleneck, and I'm not talking about ram bandwidth itself but what causes GCN/NCU to need so much bandwidth to perform.

[Dylanc1500]

7 minutes ago, Agost said:

Vega shouldn't need 800 GB/s to perform well, that's the point. There must be some internal bottleneck, and I'm not talking about ram bandwidth itself but what causes GCN/NCU to need so much bandwidth to perform.

If AMD had public documents detailing how their architecture worked it would help explain a lot, but like with IF they keep the lower workings very quiet.

[Coaxialgamer]

2 hours ago, ltguy said:

This! HBM was what ruined Fury cards. 4gb of vram was never enough for the intended purpose of that card. So to see them double down on stupid and launch another gaming card with hbm its frustrating for a long term amd fan like myself. RX 480/580 were good cards, they should have just scaled those up untill vega and hbm2 was more feasable.

HBM is also what allowed AMD to reach the massive memory bandwith they had ( impossible on gddr5 ) . That same bandwith meant that they needed less ROP .If AMD had gone gddr for the fury lineup , two things could have happened :

-64 ROPS would not have been anywhere near enough , and as such , fiji's backend would have been even more of a bottleneck .

-AMD would have needed to add more ROPS and a larger memory controller , meaning they would have had to cut shader count , given that tsmc's 28 nm reticule limit is roughly 600-ish mm² .

Performance would have gone down overall .

Sure , 4GB of VRAM when high end cards had 6 wasn't a great selling point , but ultimately it was the better choice.

[Coaxialgamer]

1 hour ago, Agost said:

Vega shouldn't need 800 GB/s to perform well, that's the point. There must be some internal bottleneck, and I'm not talking about ram bandwidth itself but what causes GCN/NCU to need so much bandwidth to perform.

if i had to guess , it's likely to be a backend bottleneck . Vega 10 uses 64 ROPs , just like hawaii and fiji before it . Gp102 has 50% more . 

The backend of a gpu uses as much bandwith as it can get . A bottleneck in this area can be alleviate by either adding ROP's or adding bandwith , which is probably why hbm2 overclocking helps so much .

 

If i had to guess , AMD was planning on having much faster HBM2 by the time the gpu was being manufactured ( Back then , hynix had 1ghz hbm2 planned for late 2016). Unfortunately , neither hynix nor samsung could keep that roadmap , and only have ~800mhz hbm2 atm .

But it was too late to change the gpu design ( RTG taped out vega 10 in june last year ) , and as such Vega has an undersized backend ( ROP-starved )

[Agosto]

3 minutes ago, Coaxialgamer said:

if i had to guess , it's likely to be a backend bottleneck . Vega 10 uses 64 ROPs , just like hawaii and fiji before it . Gp102 has 50% more . 

The backend of a gpu uses as much bandwith as it can get . A bottleneck in this area can be alleviate by either adding ROP's or adding bandwith , which is probably why hbm2 overclocking helps so much .

 

If i had to guess , AMD was planning on having much faster HBM2 by the time the gpu was being manufactured ( Back then , hynix had 1ghz hbm2 planned for late 2016). Unfortunately , neither hynix nor samsung could keep that roadmap , and only have ~800mhz hbm2 atm .

But it was too late to change the gpu design ( RTG taped out vega 10 in june last year ) , and as such Vega has an undersized backend ( ROP-starved )

Always said that Fiji was ROP starved, it's even worse on Vega.

HBM OC to 1-1.1GHz helps a bit but not as much as more ROPs could do...

[Coaxialgamer]

Just now, Agost said:

Always said that Fiji was ROP starved, it's even worse on Vega.

HBM OC to 1-1.1GHz helps a bit but not as much as more ROPs could do...

I think AMD was just caught off guard honestly . Sure , adding more ROPs would have helped , but it would have made Vega larger and even more power hungry . During the gpu design phase , there was no reason to do so because memory roadmaps were on track.

I mean , it worked well enough on fiji : they used the same number of ROPs as hawaii ( with 20 extra CU's ) but were largely able to mitigate that by using hbm . I think they wanted to pull that again because they though HBM2 would allow even more bandwith , but they were wrong i guess . 

I'm checking out AMD's GCN and Vega whitepaper atm to see if they have a section about the back-end on their gpu's .I'll tell you if i find anything.

[Coaxialgamer]

44 minutes ago, Agost said:

Always said that Fiji was ROP starved, it's even worse on Vega.

HBM OC to 1-1.1GHz helps a bit but not as much as more ROPs could do...

I checked out the whitepaper . It appears my previous assumptions are at least partially incorrect . While VEGA did keep the same number of ROPs , it's peak render back end throughput has doubled per clock vs fiji ( 67GP/s for fiji vs 107GP/s on vega ) .

Which means the actual number of ROP's is maybe not the problem . 3 possible solutions:

-Either Pixel throughput demands have substantially increased , meaning that 62% increase wasn't enough .

-Because of only being labelled "peak throughput" , it might be that there still is a memory bottleneck , and thus this peak rate is never achieved.

- ROP's aren't the bottleneck at all , and i was just wrong

[Agosto]

2 minutes ago, Coaxialgamer said:

I checked out the whitepaper . It appears my previous assumptions are at least partially incorrect . While VEGA did keep the same number of ROPs , it's peak render back end throughput has doubled per clock vs fiji ( 67GP/s for fiji vs 107GP/s on vega ) .

Which means the actual number of ROP's is maybe not the problem . 3 possible solutions:

-Either Pixel throughput demands have substantially increased , meaning that 62% increase wasn't enough .

-Because of only being labelled "peak throughput" , it might be that there still is a memory bottleneck , and thus this peak rate is never achieved.

- ROP's aren't the bottleneck at all , and i was just wrong

The 50% increase in throughput (GPU-Z gives 99 GP/s) could just be given by higher clocks

Fiji has very similar performance to Polaris 10 and Hawaii in non-optimized scenarios, the same kind of bottleneck could be happening on Vega

Moreover, what about those 11 polygons per clock vs 4 on Fiji, but only with 4 geometry engines?

Vega is still a mistery.

[Sakkura]

4 hours ago, Agost said:

I seriously can't tell why it's happening... /s


BTW, has anyone already posted this somewhere in the forum?

https://np.reddit.com/r/hardware/comments/6vp1b1/retailers_are_buying_amd_rx_vega_64_at_675_each/dm20et2/

tl;dr: AMD supposedly sells Vega to retailers at the "right price" only if they also buy X packs for every standalone card

I posted that 6 posts above yours...

[Notional]

2 hours ago, Agost said:

Vega shouldn't need 800 GB/s to perform well, that's the point. There must be some internal bottleneck, and I'm not talking about ram bandwidth itself but what causes GCN/NCU to need so much bandwidth to perform.

800 GB/s, no. More than 400 GB/s, absolutely. 1080ti has 484 GB/s, which is a bit more. Bandwidth means a lot in gaming. Kinda the entire point of HBM in general.

1 hour ago, Coaxialgamer said:

If i had to guess , AMD was planning on having much faster HBM2 by the time the gpu was being manufactured ( Back then , hynix had 1ghz hbm2 planned for late 2016). Unfortunately , neither hynix nor samsung could keep that roadmap , and only have ~800mhz hbm2 atm .

But it was too late to change the gpu design ( RTG taped out vega 10 in june last year ) , and as such Vega has an undersized backend ( ROP-starved )

14

Well, it's already confirmed to be the case. SK Hynix could not deliver full speed HBM2, so AMD had to (also) to go Samsung, who never claimed to make such clock speeds. So the HBM2 runs at a lower speed than what AMD thought it would. That certainly doesn't help.

But generally, for performance, it seems a lot of features in the architecture is still disabled.

[Doobeedoo]

I just want to see custom cards in stock retail. Hopefully priced well by then. 

[Jito463]

4 hours ago, LAwLz said:

I think it is sad that people just leech onto any positive comments they can find on Reddit

I think it's sad that people use Reddit (and Twitter, and Facebook, and.....)  

[Coaxialgamer]

1 hour ago, Notional said:

800 GB/s, no. More than 400 GB/s, absolutely. 1080ti has 484 GB/s, which is a bit more. Bandwidth means a lot in gaming. Kinda the entire point of HBM in general.

Well, it's already confirmed to be the case. SK Hynix could not deliver full speed HBM2, so AMD had to (also) to go Samsung, who never claimed to make such clock speeds. So the HBM2 runs at a lower speed than what AMD thought it would. That certainly doesn't help.

But generally, for performance, it seems a lot of features in the architecture is still disabled.

Not doubting or anything , but do you have a source that some vega features are still disabled ? I knew some were during vega FE's launch , but i searched and couldn't find any info about features that are not enabled now that RX vega is launched.

[XenosTech]

33 minutes ago, Jito463 said:

I think it's sad that people use Reddit (and Twitter, and Facebook, and.....)  

No wonder I have terrible depression. *weeps in corner of shame*

[Sauron]

On 24/8/2017 at 5:52 PM, Coaxialgamer said:

At 100$ more , i can't see rx vega selling at all for anything but niche markets.

Meanwhile AMD has actually gained marketshare because miners bought them in bulk.

[Agosto]

1 hour ago, Sakkura said:

I posted that 6 posts above yours...

Sorry, didn't see it, that's why I asked.
 

1 hour ago, Notional said:

800 GB/s, no. More than 400 GB/s, absolutely. 1080ti has 484 GB/s, which is a bit more. Bandwidth means a lot in gaming. Kinda the entire point of HBM in general.

Well, it's already confirmed to be the case. SK Hynix could not deliver full speed HBM2, so AMD had to (also) to go Samsung, who never claimed to make such clock speeds. So the HBM2 runs at a lower speed than what AMD thought it would. That certainly doesn't help.

But generally, for performance, it seems a lot of features in the architecture is still disabled.

At 512GB/s Fiji was still limited by bandwidth (high scaling with HBM OC), I wonder what speeds Vega would need.

AMD should really enable those features and improve scaling on next gpus. I hope Navi's "scalability" will also mean better scaling when adding more SPs. However if Navi doesn't use HBM2 but a "next gen" memory (supposedly HBM3) I'm afraid we won't see anything different from Vega until late 2019.

 

3 minutes ago, Coaxialgamer said:

Not doubting or anything , but do you have a source that some vega features are still disabled ? I knew some were during vega FE's launch , but i searched and couldn't find any info about features that are not enabled now that RX vega is launched.

Primitive shaders should work without developer input AFAIK, but performance per clock is the same as Fiji. 11vs4 polygons per clock geometry engines too, I suppose. Still same performance per clock as Fiji.

We'll see rapid packed math results on Far Cry 5 and Wolfenstein II, but there are some feature missing or not fully working in the drivers, I'm pretty sure.

[Coaxialgamer]

Just now, Sauron said:

Meanwhile AMD has actually gained marketshare because miners bought them in bulk.

AMD lost 2% market share actually .

Q1 2017 , AMD lost 2% market share...

https://jonpeddie.com/press-releases/details/add-in-board-market-decreased-in-q117-from-last-quarter-with-nvidia-gaining

 

Sure , they made money , (more than expect even ) because miners bought a lot of cards . But that's not healthy income .

Vega itself isn't even a great miner , averaging 40Mh/s in ethereum , with a high power draw.

[Sauron]

10 minutes ago, Coaxialgamer said:

AMD lost 2% market share actually .

Q1 2017 , AMD lost 2% market share...

https://jonpeddie.com/press-releases/details/add-in-board-market-decreased-in-q117-from-last-quarter-with-nvidia-gaining

 

Sure , they made money , (more than expect even ) because miners bought a lot of cards . But that's not healthy income .

Vega itself isn't even a great miner , averaging 40Mh/s in ethereum , with a high power draw.

Q1 was before vega...

[Coaxialgamer]

Just now, Sauron said:

Q1 was before vega...

Q2 results are not out , and neither are q3 results . So i don't know how you could know either way .

And as i said , Vega isn't a particularly good miner  for the dollar or for the watt.