The Vega RX Thread (Rumors and Info)

[Stacc]

Ouch, some cost must have gotten way out of control. Maybe HBM2 costs. Even if this had been able to compete with the GTX 1080 Ti and they sold it for $600+ the margins are still too small.

Perhaps Vega+HBM should have been sold as a compute only card like GP100. They could have made a big Polaris chip, with double the amount of compute cores as Polaris 10 but with GDDR5X memory. Probably could have had it done a year ago with performance about the same as Vega is now at a much lower cost.

 

[throAU]

GDDR5 memory costs 4-6$ per stack. Even accounting for 5 times higher prices for HBM2, compared to GDDR5(which is incorrect), it makes it 20-30 per stack, and 60$ total.

And where do those prices come from? Not sure that i believe single-supplier, high performance HBM 2 modules are cheaper in terms of $/GB than DDR4 is (at least retail).

 

[koyoot]

And where do those prices come from? Not sure that i believe single-supplier, high performance HBM 2 modules are cheaper in terms of $/GB than DDR4 is (at least retail).

Then explain why HBM2 stacks on Vega come from BOTH SK Hynix and Samsung?

 

[Stacc]

And where do those prices come from? Not sure that i believe single-supplier, high performance HBM 2 modules are cheaper in terms of $/GB than DDR4 is (at least retail).

Graphics cards don't use DDR4. They use GDDR5.

 

[throAU]

Graphics cards don't use DDR4. They use GDDR5.

No of course they don't. But I know what DDR4 costs me to buy, and GDDR5 and HBM is both higher performance than that and lower volume, thus should be more expensive.

 

[h9826790]

No of course they don't. But I know what DDR4 costs me to buy, and GDDR5 and HBM is both higher performance than that and lower volume, thus should be more expensive.

It's not that simple, they are in different market. e.g. Jet fuel (Jet A1) and car fuel, Jet A1 has higher performance, use in higher grade machines, but it's cheaper for airline to buy Jet A1, than you to refuel your car (per litter).

Since the retail market of HBM2 simply not exist (It's quite impossible to buy that in a local computer store). There is much much less packaging / logistic / marketing cost for HBM2 (then retail DDR4). End result, it's totally possible that factory HBM2 is cheaper than retail DDR4.

Also, there is no way to compare a retail price to a factory price. The difference can be huge. Depends on the items, the difference can be 100x (e.g. factory cost $0.2, retail price $20. Think about how much a screw cost in factory, and how much it cost you if you only buy one screw). One of the biggest difference I've seen is some electronics for BMW. Factory cost about $1, but they sell it at $500. After I met few businessmen who running factories in China, I know unless you really work inside a particular deal (or has real insider source), you can never know the correct price, any guess can be miles away from the truth.

Another businessman told me that the cost of his product (retail). If $1 from factory, then about $10 for logistic, and $100 for marketing. They sell it at $200-300. Of course, it's not a computer product, cannot say HBM2 will have the same ratio. However, what I want to point out is that a normal customer can hardly guess the real cost from factory. Also how much extra cost if they have to handle the retail market.

 

[throAU]

i'm aware of all that, i'm not stupid. i'm not simply comparing RRP of DDR 4 to HBM2 cost and going "herp derp it should be more than DDR4 RRP".

Of course there are margins involved. There is however zero economy of scale of manufacturing for HBM vs. DDR4. It is niche, high tech, brand new, low volume product vs. commodity memory that has been mass produced for 4 years or more.

 

[h9826790]

i'm aware of all that, i'm not stupid. i'm not simply comparing RRP of DDR 4 to HBM2 cost and going "herp derp it should be more than DDR4 RRP".

Of course there are margins involved. There is however zero economy of scale of manufacturing for HBM vs. DDR4. It is niche, high tech, brand new, low volume product vs. commodity memory that has been mass produced for 4 years or more.

I understand, most of the time this logic work, but not necessary always the truth.

New, high tech... means outsider no way to know the exact production cost (or even close). Sometimes, the newer, better, smaller volume product is cheaper. Because, occasionally, the some technology pops up and make the newer product easier + cheaper to produce. And they may use this small volume production as a test, in order to confirm that it can go large scale. Only the insider know what they are doing, not us.

I am not saying that I know HBM2 belongs to this case. In fact, it's the other way around, I have zero idea about their production cost. That's why I cannot assume HBM2 is more expensive (or cheaper) than the DDR4. I simply don't know.

 

[koyoot]

https://twitter.com/GFXChipTweeter/status/903109409739386880

It appears that the article on FUD is just that: FUD.

Another one to add to list of false articles revolving around the release of Vega.

 

[res0lve]

It's not that simple, they are in different market. e.g. Jet fuel (Jet A1) and car fuel, Jet A1 has higher performance, use in higher grade machines, but it's cheaper for airline to buy Jet A1, than you to refuel your car (per litter).

Since the retail market of HBM2 simply not exist (It's quite impossible to buy that in a local computer store). There is much much less packaging / logistic / marketing cost for HBM2 (then retail DDR4). End result, it's totally possible that factory HBM2 is cheaper than retail DDR4.

Also, there is no way to compare a retail price to a factory price. The difference can be huge. Depends on the items, the difference can be 100x (e.g. factory cost $0.2, retail price $20. Think about how much a screw cost in factory, and how much it cost you if you only buy one screw). One of the biggest difference I've seen is some electronics for BMW. Factory cost about $1, but they sell it at $500. After I met few businessmen who running factories in China, I know unless you really work inside a particular deal (or has real insider source), you can never know the correct price, any guess can be miles away from the truth.

Another businessman told me that the cost of his product (retail). If $1 from factory, then about $10 for logistic, and $100 for marketing. They sell it at $200-300. Of course, it's not a computer product, cannot say HBM2 will have the same ratio. However, what I want to point out is that a normal customer can hardly guess the real cost from factory. Also how much extra cost if they have to handle the retail market.

Can you stick to VEGA discussions, because your fuel analogy is laughable on chemistry and economics level

 

[h9826790]

Can you stick to VEGA discussions, because your fuel analogy is laughable on chemistry and economics level

I did, at least much more VEGA related than your post. I used that to explain why higher performance (and actually smaller volume as well) stuff not necessary more expensive.

How about your post? Apart from has the word "VEGA", it's totally VEGA unrelated.

Sure the fuel example won't be the best way to metaphor HBM2's situation. But I just want to point that out HBM2 is not necessary as expensive as we believe. Not sure how much you know about fuel. In my city, car fuel is tipically 2x the price of Jet fuel per litre. And I don't know any of those car fuel can burn better than the Jet fuel. If that's laughable. I am more than happy to learn from you.

 

[res0lve]

I did, at least much more VEGA related than your post. I used that to explain why higher performance (and actually smaller volume as well) stuff not necessary more expensive.

How about your post? Apart from has the word "VEGA", it's totally VEGA unrelated.

Sure the fuel example won't be the best way to metaphor HBM2's situation. But I just want to point that out HBM2 is not necessary as expensive as we believe. Not sure how much you know about fuel. In my city, car fuel is topically 2x the price of Jet fuel per litre. And I don't know any of those car fuel can burn better than the Jet fuel. If that's laughable. I am more than happy to learn from you.

Off topic discussion. Removed by Res0lve

 

[h9826790]

This waaay off topic and I'll get in trouble, but I'll bite:

1. Research how jet engine works and how IE engine works.
2. Research what fuel properties both type engines require.
2. Research what taxes are imposed on the Jet A1 and Gasoline/Diesel after they leave the refinery.

A few pointers: Octane rating, Nitromethane, Diesel, Kerosene, RP-1.

GPU can use GDDRX, HBMX (throughput /clock/voltage/latency/cost, availability).

Jet engine can't run on top fuel for ICE and ICE can't run on Jet A1 or RP-1(rocket fuel), maybe a diesel engine can run JetA1, as they are close

It's about how the GPU design. Same as engine. If the GPU is designed to use HBM2, you can't just connect it to the GDDR5 and make the graphic card works. They can use it, doesn't mean they you can change it at anytime. They can't use it now, doesn't mean they you can't work with it if re-design from the blue print.

Anyway, the idea is about higher performance stuff not always more expensive. Even though both HBM2 and GDDR5 are graphic memory, but they can be quite different in some fundamental point of view, same as different fuel. Also, wholesale vs retail also make the difference.

P.S. AFAIK, jet engine can burn almost any fuel (even alcohol), just a matter of efficiency. Obviously it doesn't need Octane, but Octane won't stop a jet engine working.

 

[BB8]

https://twitter.com/GFXChipTweeter/status/903109409739386880

It appears that the article on FUD is just that: FUD.

Another one to add to list of false articles revolving around the release of Vega.

lol, was that even a denial? pretty weak response if you ask me.

 

[Zarniwoop]

https://twitter.com/GFXChipTweeter/status/903109409739386880

It appears that the article on FUD is just that: FUD.

Another one to add to list of false articles revolving around the release of Vega.

Fake news are for sale... but I don't know who would benefit trashing AMD? Nvidia, Intel... someone looking to buy AMD stock or ATI cheap?

 

[koyoot]

Interposer - bigger on Fury than on Vega.

HBM1 4 stacks vs 2 stacks of HBM2.
Less TSV count.

Inteposer cost - 2.5$/100mm2.

If you know all this, how can magically manufacturing costs of Vega be higher than they are for Fury?

I have said: that manufacturing cost of Vega is around 120$ spot, because it is simpler design, than Fiji was.

Once again: If FUD article is correct: AMD is loosing 100$ on each Vega 64, and 200$(!) on each Vega 56. It is idiotic idea.

 

[3587]

I watched a couple blind tests on YouTube... where people would play games using a GTX 1080Ti and a Vega RX 64 without knowing which system was which... Vega RX 64 was reported to be smoother by most of the users. Charts and graphs only tell a part of the story...

 

[h9826790]

I watched a couple blind tests on YouTube... where people would play games using a GTX 1080Ti and a Vega RX 64 without knowing which system was which... Vega RX 64 was reported to be smoother by most of the users. Charts and graphs only tell a part of the story...

This is interesting, I have no doubt that Nvidia card can perform better on benchmarks, but I don't know the real world difference can be this small. May be the setting is not pushing hard enough? If both cards are overkill, then overkill 50% or overkill 100% of course won't make any significant difference. Users still feeling the game is 100% smooth.

 

[koyoot]

This is interesting, I have no doubt that Nvidia card can perform better on benchmarks, but I don't know the real world difference can be this small. May be the setting is not pushing hard enough? If both cards are overkill, then overkill 50% or overkill 100% of course won't make any significant difference. Users still feeling the game is 100% smooth.

If you are using FreeSync and G-Sync monitors - you will not see the difference between Vega 56 and GTX 1080 Ti.

 

[Asgorath]

Interposer - bigger on Fury than on Vega.

HBM1 4 stacks vs 2 stacks of HBM2.
Less TSV count.

Inteposer cost - 2.5$/100mm2.

If you know all this, how can magically manufacturing costs of Vega be higher than they are for Fury?

I have said: that manufacturing cost of Vega is around 120$ spot, because it is simpler design, than Fiji was.

Once again: If FUD article is correct: AMD is loosing 100$ on each Vega 64, and 200$(!) on each Vega 56. It is idiotic idea.

Right, because raw manufacturing costs are the only thing that matters. The chip design magically fell from the sky, and it was magically marketed for free all over the world.

 

[koyoot]

Right, because raw manufacturing costs are the only thing that matters. The chip design magically fell from the sky, and it was magically marketed for free all over the world.

In the article you quoted, there are words about manufacturing costs, and retail price of Vega, without design costs. Now you are moving the goalpost to design costs. Just FYI. Design costs for 14/16 nm process is between 70-80 mln $ per year, over 4 year period. Regardless of its size, and type of Chip. It cost AMD 80 mln $ per year to design Ryzen CPUs. it could cost AMD 70 mln $ per year to design Vega.

Just because you want to show AMD in worst possible way, does not mean you are right. More likely - you are not. Deal with it.

Edit: To make things worse for you, and your agenda, in this case. For AMD to lose 100$ on each Vega 64 GPU, HBM2 would have to cost 200$ per stack! A margin of 50 times higher price than GDDR5. Do you see how idiotic is this idea, and that article?

 

[Asgorath]

In the article you quoted, there are words about manufacturing costs, and retail price of Vega, without design costs. Now you are moving the goalpost to design costs. Just FYI. Design costs for 14/16 nm process is between 70-80 mln $ per year, over 4 year period. Regardless of its size, and type of Chip. It cost AMD 80 mln $ per year to design Ryzen CPUs. it could cost AMD 70 mln $ per year to design Vega.

Just because you want to show AMD in worst possible way, does not mean you are right. More likely - you are not. Deal with it.

Edit: To make things worse for you, and your agenda, in this case. For AMD to lose 100$ on each Vega 64 GPU, HBM2 would have to cost 200$ per stack! A margin of 50 times higher price than GDDR5. Do you see how idiotic is this idea, and that article?

Actually I don't want to show AMD in the worst possible way, not sure why you think that. I just think you're vastly underestimating the costs of producing Vega. We can agree to disagree, and I guess the proof will be in the coming quarters when AMD announces their profits/losses. There are tons of other costs that you seem to be ignoring (e.g. software engineers for drivers/compilers, QA etc). I'll be happy to be wrong when AMD shocks the world and announces that is it massively profitable for the first time in a very long time next quarter, because Vega was significantly cheaper to produce than I expected and they're making so much money on every one they sell. Not holding my breath though.

 

[Stacc]

I watched a couple blind tests on YouTube... where people would play games using a GTX 1080Ti and a Vega RX 64 without knowing which system was which... Vega RX 64 was reported to be smoother by most of the users. Charts and graphs only tell a part of the story...

It seems silly to put any weight towards these type of subjective tests when we can just look at one of the million reviews that does in depth frame time analysis. These type of tests were just a marketing ploy by AMD because Vega was going to struggle to beat Nvidia's competition so they had to fall back on the value proposition. In this case it wasn't just the value of one GPU vs another, they had to include the costs of the monitors as well.

Depending on the target frame rate or resolution it would be pretty easy to pick a case where both AMD and Nvidia max out the frame rate the monitor can actually display. That way they both appear similar in subjective tests. I don't think there is any question that if you want to game in 4k at 60 FPS the far and away best choice is the GTX 1080 Ti.

In the article you quoted, there are words about manufacturing costs, and retail price of Vega, without design costs. Now you are moving the goalpost to design costs. Just FYI. Design costs for 14/16 nm process is between 70-80 mln $ per year, over 4 year period. Regardless of its size, and type of Chip. It cost AMD 80 mln $ per year to design Ryzen CPUs. it could cost AMD 70 mln $ per year to design Vega.

Just because you want to show AMD in worst possible way, does not mean you are right. More likely - you are not. Deal with it.

Edit: To make things worse for you, and your agenda, in this case. For AMD to lose 100$ on each Vega 64 GPU, HBM2 would have to cost 200$ per stack! A margin of 50 times higher price than GDDR5. Do you see how idiotic is this idea, and that article?

You have no idea how much it costs to make Vega. Only AMD knows. What we do know is that Vega uses non-standard components such as an interposer and new, exotic memory that is provided by only a single manufacturer.

We have a bunch of anecdotal evidence that suggests that Vega is expensive. Like rumors from the supply chain that the margins are non-existent, or the fact that the only other graphics product that uses HBM2 is GP100, which costs upwards of $7000. There are rumors of HBM2 being in short supply. The lists goes on... Knowing these things its reasonable to assume that Vega costs more than an equivalent conventional GPU.

If using HBM(2) was cheaper than conventional memory we would see it used across the AMD's lineup, but we don't. Its only existed on the highest end components.

 

[wallysb01]

http://hexus.net/tech/news/graphics/109544-amd-radeon-rx-vega-64-makes-100-loss-msrp-says-report/

AMD is losing at least $100 on every Vega 64. Pretty sure I predicted this a while back.

I guess this is the growth at all costs model creeping into yet another market, like how Tesla loses money on every car they sell. Its a brave new world.

 

[koyoot]

You have no idea how much it costs to make Vega. Only AMD knows. What we do know is that Vega uses non-standard components such as an interposer and new, exotic memory that is provided by only a single manufacturer.

We have a bunch of anecdotal evidence that suggests that Vega is expensive. Like rumors from the supply chain that the margins are non-existent, or the fact that the only other graphics product that uses HBM2 is GP100, which costs upwards of $7000. There are rumors of HBM2 being in short supply. The lists goes on... Knowing these things its reasonable to assume that Vega costs more than an equivalent conventional GPU.

If using HBM(2) was cheaper than conventional memory we would see it used across the AMD's lineup, but we don't. Its only existed on the highest end components.

If I do not know how much Vega costs, the FUD even more so does not know how much Vega costs. And yet, we have a poster who claims that he predicted that AMD will be losing money on each GPU sold.

When provide evidence so it is not the case, its me who do not know how much the GPU costs to manufacture, and design.

Look at that graph. Look at the prices of the GPUs. Look at the article which claimed that AMD is selling 600$ for 499$. Then look at Asgorath's post.

Actually I don't want to show AMD in the worst possible way, not sure why you think that. I just think you're vastly underestimating the costs of producing Vega. We can agree to disagree, and I guess the proof will be in the coming quarters when AMD announces their profits/losses. There are tons of other costs that you seem to be ignoring (e.g. software engineers for drivers/compilers, QA etc). I'll be happy to be wrong when AMD shocks the world and announces that is it massively profitable for the first time in a very long time next quarter, because Vega was significantly cheaper to produce than I expected and they're making so much money on every one they sell. Not holding my breath though.

Its way easier to say: I was wrong.

Let me give you one example why. If AMD is selling each GPU, that cost the company 280 mln $ to design, 600$ to manufacture, for as low as 399$ - that means they will loose not 100$ on each GPU. AMD is going to lose almost half a billion US dollars.

The amount of stupidity in that post is beyond me. I cannot even comprehend that anyone of you can believe in it.