AMD Confirms RX 480 At $199 USD, Other APU & Polaris Announcements

[goMac]

Why would anyone pay $1000 to be on the "frontier" of AMD's unfinished technology?

Why do people pay a lot of money for unfinished VR headsets?

(Also, the drivers will be completed. It's just software. They're delivering compute performance today and gaming performance tomorrow. So you're getting the better performance in the end.)

 

[Stacc]

Question for the AMD faithful in this thread, are there any Vega machine learning benchmarks out there? I'm looking for Tensorflow or other commonly used tools. Oh, and AMD demos don't count.

 

[koyoot]

Looks like AMD was demoing Vega RX today at some sort of event in Budapest. It seems they are advertising Vega and a freesync monitor perform the same as a GTX 1080 and GSync monitor for $300 less. Thats a weak argument given that Gsync monitors seem to go for $100-$200 more than freesync monitors.

Maybe its just me, but AMD has not been hyping this nearly as much as they did with Ryzen. Probably not a good sign for how it will stack up gaming wise. Obviously this is all subjective, but it should only be a couple more weeks until they announce consumer/gaming Vega.

They compared to direct price point competitor.

AMD used 300$ Nvidia G-Sync Tax monitor over non-tax Freesync monitor.

So yeah - Vega will cost between 499 and 599$ just by looking at this...

 

[SoyCapitanSoyCapitan]

They compared to direct price point competitor.

AMD used 300$ Nvidia G-Sync Tax monitor over non-tax Freesync monitor.

So yeah - Vega will cost between 499 and 599$ just by looking at this...

That's not competitive if they offer Nvidia's 2016 performance.

If these companies were honest they would cut down on the bulky materials and large fans they ship these products with. They do it on purpose to profit a little more. We know from mobile full fat versions and single slot versions that all these desktop GPUs can be made slimmer and with lower TDP - without losing much performance!

Now if AMD did that they would have a good competitor. But no, their stock value has slowed and they need to make money from bits of extra plastic and metal stick to the cards.

 

[Stacc]

They compared to direct price point competitor.

AMD used 300$ Nvidia G-Sync Tax monitor over non-tax Freesync monitor.

So yeah - Vega will cost between 499 and 599$ just by looking at this...

Its been pretty clear all along Vega will have to positioned as an enthusiast product. It has a ~500 mm2 die, which is in the same ballpark as GP102, and uses exotic HBM2 memory. Anyone expecting GTX 1070 prices with > GTX 1080 performance is living in fantasy land.

If these companies were honest they would cut down on the bulky materials and large fans they ship these products with. They do it on purpose to profit a little more. We know from mobile full fat versions and single slot versions that all these desktop GPUs can be made slimmer and with lower TDP - without losing much performance!

Its not fans and "bulky materials" that make these cards expensive. It all has to do with the design and manufacture of a 500 mm2 chip with billions of transistors that can render graphics in realtime that exceed movie special effects of a few years ago.

Bigger coolers on GPUs actually make it easier and cheaper to produce. More restrictive cooling and lower TDP requires better chip binning instead of just throwing a lot of cooling and voltage at the problem and getting better yields. The R9 Nano was an expensive product for just this reason. Also, mobile GPUs carry significant premiums over their desktop counterparts.

 

[SoyCapitanSoyCapitan]

Its been pretty clear all along Vega will have to positioned as an enthusiast product. It has a ~500 mm2 die, which is in the same ballpark as GP102, and uses exotic HBM2 memory. Anyone expecting GTX 1070 prices with > GTX 1080 performance is living in fantasy land.



Its not fans and "bulky materials" that make these cards expensive. It all has to do with the design and manufacture of a 500 mm2 chip with billions of transistors that can render graphics in realtime that exceed movie special effects of a few years ago.

Bigger coolers on GPUs actually make it easier and cheaper to produce. More restrictive cooling and lower TDP requires better chip binning instead of just throwing a lot of cooling and voltage at the problem and getting better yields. The R9 Nano was an expensive product for just this reason. Also, mobile GPUs carry significant premiums over their desktop counterparts.

You misunderstand.

They supply us with larger cooling and higher TDPs than they need to. We have single slot GTX and full fat mobile versions of GTX 1070/1080 for example. Or Radeon Nano. The list is long.

Even the premium costs of Nano and mobile Pascal are inflated because of the tier structure in the product line ups. By reducing the tier structure, being efficient across the remaining line up, consumers can have lower TDP and more affordable devices.

The Nano fell in price fast. The single slow Pascal cards are the same price as the dual slot fat cards. Full fat Pascal cards are in affordable PC laptops. So we see that the inflated prices can easily be adjusted when desired.

By being inefficient the duopoly of AMD and Nvidia (and their partners) are happy to extract more money than we need to be paying. The same goes for the Intel-AMD duopoly. The former is happy to slow down development until threatened by the latter. And the latter are happy to spend months hyping a product that doesn't live up to promise.

 

[koyoot]

Manufacturing cost of Vega GPU is around 120$... without packaging and testing, and validation, so add another 10-15$ to this. This is including HBM2 stacks.

Yield is very high. We are talking about around 92-95% yields from GloFo/Samsung process. AMD can thank for those IBM engineers which are doing tremendous job over past months at GloFo.

As for price. I have written this before. I think Vega price will be in 599-699$ ballpark.

 

[Stacc]

You misunderstand.

They supply us with larger cooling and higher TDPs than they need to. We have single slot GTX and full fat mobile versions of GTX 1070/1080 for example. Or Radeon Nano. The list is long.

Even the premium costs of Nano and mobile Pascal are inflated because of the tier structure in the product line ups. By reducing the tier structure, being efficient across the remaining line up, consumers can have lower TDP and more affordable devices.

The Nano fell in price fast. The single slow Pascal cards are the same price as the dual slot fat cards. Full fat Pascal cards are in affordable PC laptops. So we see that the inflated prices can easily be adjusted when desired.

By being inefficient the duopoly of AMD and Nvidia (and their partners) are happy to extract more money than we need to be paying. The same goes for the Intel-AMD duopoly. The former is happy to slow down development until threatened by the latter. And the latter are happy to spend months hyping a product that doesn't live up to promise.

I did not misunderstand. The smaller your target TDP, the better binned and more resilient to lower voltages the chip has to be. This is why desktop GPUs always come out before mobile, Intel's desktop quad cores always come out before the mobile ones, etc.

If your logic was true, the cheapest enthusiast graphics cards would all have a single fan. This is not the case. There is a premium for ITX versions of cards and usually a small performance penalty.

 

[Asgorath]

Manufacturing cost of Vega GPU is around 120$... without packaging and testing, and validation, so add another 10-15$ to this. This is including HBM2 stacks.

Yield is very high. We are talking about around 92-95% yields from GloFo/Samsung process. AMD can thank for those IBM engineers which are doing tremendous job over past months at GloFo.

As for price. I have written this before. I think Vega price will be in 599-699$ ballpark.

Are you forgetting about the billions and billions of dollars of R&D investment that went into designing the chip in the first place? Final price isn't just about how much it costs to manufacture, assuming AMD actually wants to be a profitable company they have to make back all the money they've poured into designing this chip.

 

[koyoot]

Are you forgetting about the billions and billions of dollars of R&D investment that went into designing the chip in the first place? Final price isn't just about how much it costs to manufacture, assuming AMD actually wants to be a profitable company they have to make back all the money they've poured into designing this chip.

This is what every chip design has to "make for" in the first place, then it can "make for" itself.

I know perfectly well about the design costs. I am talking only about manufacturing chips, without design cost.
What those numbers mean is that actually the manufacturing costs are not bad, and AMD can earn "some" money on Vega.

 

[Stacc]

Are you forgetting about the billions and billions of dollars of R&D investment that went into designing the chip in the first place? Final price isn't just about how much it costs to manufacture, assuming AMD actually wants to be a profitable company they have to make back all the money they've poured into designing this chip.

Don't get too hung up on it. He has no idea how much it costs to design and manufacture Vega or any AMD chip. Only AMD knows unless koyoot can source his information.

 

[koyoot]

Don't get too hung up on it. He has no idea how much it costs to design and manufacture Vega or any AMD chip. Only AMD knows unless koyoot can source his information.

The designs for 14 nm FF process cost around 70 mln $ per year, with exception of Zen architecture which cost AMD 320 mln $ over 4 years period of time.

I can not give you source for this. As I have mentioned. The manufacturing costs per die, right now are not accounting design costs. They were just to show that manufacturing costs are not that bad, and that AMD has quite a lot of room for margin.


One more thing about Vega this time. It appears that 8-Hi stacks for this GPU are provided by... Samsung.

 

[Asgorath]

The designs for 14 nm FF process cost around 70 mln $ per year, with exception of Zen architecture which cost AMD 320 mln $ over 4 years period of time.

I can not give you source for this. As I have mentioned. The manufacturing costs per die, right now are not accounting design costs. They were just to show that manufacturing costs are not that bad, and that AMD has quite a lot of room for margin.


One more thing about Vega this time. It appears that 8-Hi stacks for this GPU are provided by... Samsung.

Again, you're kind of missing the point. Manufacturing costs are a tiny fraction of the overall amount of money spent to design and build a GPU as large and complex as Vega. $320M over 4 years might be 10% or less of the total amount of money AMD has already spent to design the Vega architecture. It's literally multiple billions (with a B) of dollars to design and build a modern GPU. Good luck making that money back if they turn around and sell it for $700.

Edit: Volta had an R&D budget of $3B.

http://www.pcworld.com/article/3196...th-21-billion-transistors-and-5120-cores.html

Volta is “at the limits of photolithography,” Huang said with a smirk, created using an R&D budget of over $3 billion.

 

[koyoot]

Again, you're kind of missing the point. Manufacturing costs are a tiny fraction of the overall amount of money spent to design and build a GPU as large and complex as Vega. $320M over 4 years might be 10% or less of the total amount of money AMD has already spent to design the Vega architecture. It's literally multiple billions (with a B) of dollars to design and build a modern GPU. Good luck making that money back if they turn around and sell it for $700.

Billions of dollars each year costs AMD to run company, not design the GPUs, only...

It appears you do not understand what company is AMD, and what they targeted. Why is the number so low? Let me give you an example. AMD is able to offer Ryzen CPUs for such low price, because the design costs were low, because the manufacturing costs are low.

Badum tss.

AMD always targeted lowest manufacturing costs, and highest possible TAM. That is actually historical. Accidentally they were able to design best product we have had in CPUs since Nehalem.

 

[cube]

Excavator was engineered with low cost design techniques. And shrinking Bristol/Stoney to 22FDX should cost little.

I imagine the 14nm products are not so automatically designed.

 

[Asgorath]

Billions of dollars each year costs AMD to run company, not design the GPUs, only...

It appears you do not understand what company is AMD, and what they targeted. Why is the number so low? Let me give you an example. AMD is able to offer Ryzen CPUs for such low price, because the design costs were low, because the manufacturing costs are low.

Badum tss.

AMD always targeted lowest manufacturing costs, and highest possible TAM. That is actually historical. Accidentally they were able to design best product we have had in CPUs since Nehalem.

Get back to me when AMD has a profitable quarter then, sure. A modern GPU involves multiple years worth of design effort before the first silicon is produced. I'd also argue that a modern GPU is massively more complicated than a modern CPU. As I said, manufacturing costs are a tiny fraction of the overall investment required to produce a new GPU architecture like Vega.

Or, maybe you're correct and AMD isn't investing much in R&D for their GPUs, which is why they can't produce a competitive product on time. Similarly, they're not investing in their driver teams, which is why the Vega drivers still aren't done.

 

[koyoot]

Get back to me when AMD has a profitable quarter then, sure. A modern GPU involves multiple years worth of design effort before the first silicon is produced. I'd also argue that a modern GPU is massively more complicated than a modern CPU. As I said, manufacturing costs are a tiny fraction of the overall investment required to produce a new GPU architecture like Vega.

Or, maybe you're correct and AMD isn't investing much in R&D for their GPUs, which is why they can't produce a competitive product on time. Similarly, they're not investing in their driver teams, which is why the Vega drivers still aren't done.

Your neglecting efforts of what is obvious are laughable.

Just because Nvidia has to spend 3 bln USD on GPU design does not mean everybody else has to do the same thing.

If AMD is spending 10% of what Intel and Nvidia are spending on their designs and still is competitive, in some cases extremely competitive, providing better product than their direct competitor - that is pretty great achievement.

As have been said by cube before. Here on this forum people are more interested in hating AMD than discussing the technology.

Its funny that we came from spending less on design of chips being better, to spending more on designing of the chips being better, just because it fits narration.

 

[Asgorath]

Your neglecting efforts of what is obvious are laughable.

Just because Nvidia has to spend 3 bln USD on GPU design does not mean everybody else has to do the same thing.

If AMD is spending 10% of what Intel and Nvidia are spending on their designs and still is competitive, in some cases extremely competitive, providing better product than their direct competitor - that is pretty great achievement.

As have been said by cube before. Here on this forum people are more interested in hating AMD than discussing the technology.

Its funny that we came from spending less on design of chips being better, to spending more on designing of the chips being better, just because it fits narration.

Okay, get back to me when AMD has a profitable quarter and has a product that can compete with Volta.

For reference, I'm not hating on AMD here, I'm just pointing out that I think it's extremely naive to expect RX Vega to be priced at $600 and for AMD to be making a ton of money at that price. They may end up being forced to price it that low given how bad their competitive performance is, but that doesn't mean they're going to make money at that price. We'll see in a few more weeks when RX Vega is announced.

 

[koyoot]

600$ is low? Its very high, actually...

But that is my opinion only.

Another film from AdoredTV:

Whaaaat? Freesync Monitor 799$, G-Sync Monitor 1399$, price difference between them 500$, price difference between Vega+monitor and Nvidia GTX + G-Sync Monitor 300$.

It means that Vega will actually cost 699$, if they compared to GTX 1080@499$...

 

[SoyCapitanSoyCapitan]

I did not misunderstand. The smaller your target TDP, the better binned and more resilient to lower voltages the chip has to be. This is why desktop GPUs always come out before mobile, Intel's desktop quad cores always come out before the mobile ones, etc.

.

You are correct but note they stick to the same inefficient bulk on desktop models until EOL. No efficient updated designs except for a small number of board partners who are brave enough to buck the trend.

Look at Polaris. Out for a year and the updated cards become even more inefficient and bulkier designs with very little increase in performance. They could have shipped 580s in Nano form at 110w TDP (many people have undervolted the cards). Nope. Customers forced to spend money on big cooling designs that aren't actually needed.

 

[res0lve]

Your neglecting efforts of what is obvious are laughable.

Just because Nvidia has to spend 3 bln USD on GPU design does not mean everybody else has to do the same thing.

If AMD is spending 10% of what Intel and Nvidia are spending on their designs and still is competitive, in some cases extremely competitive, providing better product than their direct competitor - that is pretty great achievement.

As have been said by cube before. Here on this forum people are more interested in hating AMD than discussing the technology.

Its funny that we came from spending less on design of chips being better, to spending more on designing of the chips being better, just because it fits narration.

AMD R&D spendings
https://ycharts.com/companies/AMD/r_and_d_expense

NVDA:

https://ycharts.com/companies/NVDA/r_and_d_expense

Intel:
This includes all kinds of stuff cpu, broadband wired + wireless, silicon fabrication etc
https://ycharts.com/companies/INTC/r_and_d_expense

Apple:https://ycharts.com/companies/AAPL/r_and_d_expense


I wrote it once, I'll do it again. You don't have a clue.

 

[koyoot]

AMD R&D spendings
https://ycharts.com/companies/AMD/r_and_d_expense

NVDA:

https://ycharts.com/companies/NVDA/r_and_d_expense

Intel:
This includes all kinds of stuff cpu, broadband wired + wireless, silicon fabrication etc
https://ycharts.com/companies/INTC/r_and_d_expense

Apple:https://ycharts.com/companies/AAPL/r_and_d_expense


I wrote it once, I'll do it again. You don't have a clue.

Yes I did not have a clue about Nvidia spendings. Which is quite funny considering the context of using a hyperbole, to picture something, and taking it 100% correct, but anyway... Did I had not a clue about AMD and Intel R&D Budget? Nope, I had very much clue.

P.S. Those are quarterly spendings, not yearly.

 

[Asgorath]

Yes I did not have a clue about Nvidia spendings. Which is quite funny considering the context of using a hyperbole, to picture something, and taking it 100% correct, but anyway... Did I had not a clue about AMD and Intel R&D Budget? Nope, I had very much clue.

P.S. Those are quarterly spendings, not yearly.

So you can't multiply by 4? Or add 4 numbers together? I don't understand the point you're trying to make.

 

[Stacc]

You are correct but note they stick to the same inefficient bulk on desktop models until EOL. No efficient updated designs except for a small number of board partners who are brave enough to buck the trend.

Look at Polaris. Out for a year and the updated cards become even more inefficient and bulkier designs with very little increase in performance. They could have shipped 580s in Nano form at 110w TDP (many people have undervolted the cards). Nope. Customers forced to spend money on big cooling designs that aren't actually needed.

Faster and more power hungry is driven by market demands. Look at all the factory overclocked, gamer oriented cards. All of these are less efficient than the reference designs but these are able to command higher prices due to higher demand. AMD could ship the 580 with a 110 W TDP, but it would lose every benchmark to its direct competitor, the GTX 1060.

 

[cube]

Faster and more power hungry is driven by market demands. Look at all the factory overclocked, gamer oriented cards. All of these are less efficient than the reference designs but these are able to command higher prices due to higher demand. AMD could ship the 580 with a 110 W TDP, but it would lose every benchmark to its direct competitor, the GTX 1060.

The 1060 3GiB lacks VRAM for some things.