[Anandtech] Discrete Q2 GPU Marketshare - AMD rises to 30%

[tajoh111]

Wait,what even in the UK websites like Overclockers UK were selling the RX470 and RX480 as successors to the R9 270/R9 270X and the R9 380/R9 380X. If the pound had not been devalued due to BREXIT and BREXIT fears,they would have slotted in around £200 and under. The same goes with the two versions of the GTX1060.

Plus,look at the die sizes for the chips - Polaris 10 is around 200MM2 to 250MM2 ,which is close to the same size as previous GPUs in cards like the HD6870 and HD7870. The HD6870 launched at the same price as the RX480.

Also,using the HD7850 and HD7870 as comparisons is also forgetting history on purpose. The GTX660 was released six months after the HD7870 and that was the true competitor to the HD7870.

Nvidia made the cut down GTX660TI using a much larger chip which forced the HD7870 price down,and then AMD repositioned the HD7950 as the main competitor for the GTX660TI.

However,the GTX660TI was released in August 2012,and that was still months after the March 2012 launched of the HD7870.

That is why AMD priced it so highly,since it was competing with GTX570 cards at launch.

Prior to the gtx 1080 launch, if Nvidia didn't launch anything, we could see history repeat itself with AMD having higher pricing prior to Nvidia launching kepler.

The rx480 was performing along the lines up gtx 970 range and slightly better and also similar to the r9 390, another 300-330 dollar card. Both these cards were great value for their generation. So it would be no surprise if AMD was not aggresive with pricing vs them. A 250-260 dollar card would definitely be midrange(4gb model) and it would have provided AMD with the buffer it needs to prevent canibalization.

A rx 480 at 250-259 for the 4gb and 299 for the 8gb, would have been par for the course, if AMD previous behavior is anything to go by. In the last 5 years, the only aggressively priced card from AMD considering it's competition was Hawaii. Hawaii and arguably the fury none x has been the only card on 28nm that has forced Nvidia to lower their prices.

The 7970, 7870, r9 285, fury x, nano came priced high and it was Nvidia cards, that forced them to do price cuts(gk104, gm104 and the gtx 980 ti).

What AMD fans forget is Nvidia puts pressure on AMD too. It's not a one way street and its one o the reasons why Nvidia was able to pick up so much marketshare during 28nm.

Your revising the bold part or telling a lie. What caused the 7870 to drop in price was the gtx 670.
It was 50 dollars more than a 7870 and performed better than a 7950. Hence 7950 prices had to fall and 7870 prices had to fall accordingly. The gtx 660 ti came out well after AMD's price drops including the 7870. Particularly when street prices were belows AMD official msrp.

And it adds more evidence to my case that AMD cut prices to polaris actual launch. Nvidia gx104 series forced price cuts everywhere in AMD lineup including their pitcairns cards.

The thing with the 6870 is that it was made on a very mature node that was dirt cheap. And because it performed worse than a 5870 it had to be priced below it. 40nm and 14/16nm finfet production cost is vastly difference which adds more evidence for my theory. For the die size Polaris is the cheapest on a new node ever. As a node matures, it becomes progressively cheaper to make larger and largers chips on it. the 6870 was made at the peak of maturity on 40nm's. During this generation Nvidia could get away with selling 5xxmm2 chips at 499 and 350. They could't do this anymore. And these wafers being around 3000 dollars and the yields being at the best they are going to get, makes putting larger chips profitable at lower price ranges.

Polaris is the complete opposite. It is made on a brand spanking new 14nm LPP finfet design. Finfet wafers costs between 7000-8500 dollars which made cost per transistor not decrease vs last generation. What his means is if yields were the same, a 400mm2 chip 28nm is the same cost as a 200mm2 finfet chip. The only savings come in the form of yields which is part of every smaller chip. This is well documented. Prior to finfet, wafers might go 15%-20% but because transistor density doubled there was a net savings.

This brings me back to my point. The problem is AMD is charging historically low prices for their 2xxmm2 die(they haven't been this low since the 3870(the 6870 1gb is more equivalent to the 4gb rx480), at a time when production cost is at record high levels. It's a contradiction meaning AMD is not making much money on their cards.

Add in the other wrinkles in their pricing and you can tell AMD is hitting near the floor of where they begin to lose money and its why its price to performance is so screwed up.

The rx 480 4gb is 199 and the rx470 4gb is 179. The rx 480 requires a perfect die which makes it a far rarer die than a rx470. The higher the price of the rx 480, the lower the demand for the card is. What is going to happen when their is only a 20 dollars difference between the two cards?

Basically the rx480 takes demand and sales away from the rx470 and because there is not enough difference in price between the two cards. And because the rx480 is much more supply limited than the rx470, there is a huge loss in potential revenue.

Ideally what you want is there to be enough of a price gap between the two cards that rx480 sales meet the supply and the same with the rx470. What you have in this situation is because the rx 480 is so closely priced to the rx470, most people prefer it and skip the rx470 to get the rx480. What happens as a result is the rx480 supply runs out, and you lose a tremendous amount of rx470 sales, because those people wait for the rx480 to become in stock again

The situation is even worse for the rx460 4gb. The rx 470 4gb is 70-80% faster than an rx 460 4gb, but there is only a 40 dollar difference in price. As a result no one besides the OEMS bite on the rx 460 because it's a no brainer to spend a little more to get vastly better performance. And the cannibalization of sales continue.

This very concept is why there has been a minimum of 50 difference between the x870 and x850 cards and sometimes 100 dollars. Absolutely never has there been just 20 dollars difference between a cut die and a full die. Even memory upgrades cost more than this. The only explanation for this is the rx470 hits floor pricing.

To all the fanboys questioning this, why has AMD set up their pricing so closely space when there will undoubtedly be cannibalizing their own sales?

It's basic marketing, economics and business.

Warning issued for inflammatory language.
-- stahlhart

 

[Bacon1]

Polaris is the complete opposite. It is made on a brand spanking new 14nm LPP finfet design. Finfet wafers costs between 7000-8500 dollars which made cost per transistor not decrease vs last generation. What his means is if yields were the same, a 400mm2 chip 28nm is the same cost as a 200mm2 finfet chip. The only savings come in the form of yields which is part of every smaller chip. This is well documented. Prior to finfet, wafers might go 15%-20% but because transistor density doubled there was a net savings.

This brings me back to my point. The problem is AMD is charging historically low prices for their 2xxmm2 die(they haven't been this low since the 3870(the 6870 1gb is more equivalent to the 4gb rx480), at a time when production cost is at record high levels. It's a contradiction meaning AMD is not making much money on their cards.

Add in the other wrinkles in their pricing and you can tell AMD is hitting near the floor of where they begin to lose money and its why its price to performance is so screwed up.

Source on pricing?

AMD said a long time ago when announcing Polaris that they were "Bringing VR performance to the masses at lower price points". Considering VR was 970/290 @ $300, why would you expect Polaris to launch near $300? They said that the jump to 14nm saved them a lot per die compared to 28nm which means they were able to charge less. You are contradicting them without any proof.

 

[iiiankiii]

Source on pricing?

AMD said a long time ago when announcing Polaris that they were "Bringing VR performance to the masses at lower price points". Considering VR was 970/290 @ $300, why would you expect Polaris to launch near $300? They said that the jump to 14nm saved them a lot per die compared to 28nm which means they were able to charge less. You are contradicting them without any proof.

IMO, the plan was to bring VR to the masses. However, I didn't think they wanted the RX480 to do that. They probably wanted the RX470 and/or the RX460 to do that. Based on the die size, Polaris should have been a much better performer. I think the plan was to have the RX480 perform closer to the R9 Fury and the RX470 a bit better than the GTX 970. As such, if Polaris was a better performer, we would have the RX480 @ $299-$349 and the RX470 @ $199-$249. But, since they found out that Polaris sucks, they had to move the entire stack down.

 

[Bacon1]

IMO, the plan was to bring VR to the masses. However, I didn't think they wanted the RX480 to do that. They probably wanted the RX470 and/or the RX460 to do that. Based on the die size, Polaris should have been a much better performer. I think the plan was to have the RX480 perform closer to the R9 Fury and the RX470 a bit better than the GTX 970. As such, if Polaris was a better performer, we would have the RX480 @ $299-$349 and the RX470 @ $199-$249. But, since they found out that Polaris sucks, they had to move the entire stack down.

They never said they were going to have a fury replacement until Vega. Their whole Polaris push was pure mass-market cards. What source do you have to show they wanted Polaris to perform at Fury levels?

 

[tajoh111]

Source on pricing?

AMD said a long time ago when announcing Polaris that they were "Bringing VR performance to the masses at lower price points". Considering VR was 970/290 @ $300, why would you expect Polaris to launch near $300? They said that the jump to 14nm saved them a lot per die compared to 28nm which means they were able to charge less. You are contradicting them without any proof.

http://caxapa.ru/thumbs/598000/WP_handel-jones.pdf

Page 1, table 1 shows the wafer cost for TSMC and Samsung. The bar graph below this shows the cost per transistor not going down.

Look at page 7 and 8 or table 6 and 7. This shows the marked up cost for AMD and Nvidia. This doesn't include packaging costs for the die. The 45% is the mark up cost so TSMC and samsung make money.

Also FD SOI isn't used on polaris. It's finfet.

The price dramatically went up. More than doubling as table 1 shows.

The 6870 1gb was 40nm card and it was 240 dollars on a mature node.

The rx480 4gb is 200 dollars somehow and the chip is made on a vastly more expensive manufacturing node.

If AMD, had any time to raise GPU prices for the die size, 16/14nm finfet was the time to raise the price. Nvidia did and their margins didn't rise since last quarter. But the the rx 480 is priced the same and if we look at AMD's competion from the gtx 1070, its pricing makes sense.

 

[Piroko]

http://caxapa.ru/thumbs/598000/WP_handel-jones.pdf

Page 1, table 1 shows the wafer cost for TSMC and Samsung. The bar graph below this shows the cost per transistor not going down.

Look at page 7 and 8 or table 6 and 7. This shows the marked up cost for AMD and Nvidia. This doesn't include packaging costs for the die. The 45% is the mark up cost so TSMC and samsung make money.

Also FD SOI isn't used on polaris. It's finfet.

The price dramatically went up. More than doubling as table 1 shows.

That slide is from January '14, since then a lot has happened. A good tellsign that this may not be accurate anymore is that lower priced SOCs have started to appear in volume (for example Snapdragon 625). Meanwhile, TSMC has also launched a lower cost 16nm process and there are some analysts that point towards sinking cost per transistor at TSMC 10nm: http://www.eetimes.com/author.asp?section_id=36&doc_id=1326864
(take the source with a grain of salt though)

Price per mm² is certainly going up by a lot, but maybe not as much as was predicted in 2014.

 

[tajoh111]

That slide is from January '14, since then a lot has happened. A good tellsign that this may not be accurate anymore is that lower priced SOCs have started to appear in volume (for example Snapdragon 625). Meanwhile, TSMC has also launched a lower cost 16nm process and there are some analysts that point towards sinking cost per transistor at TSMC 10nm: http://www.eetimes.com/author.asp?section_id=36&doc_id=1326864
(take the source with a grain of salt though)

Price per mm² is certainly going up by a lot, but maybe not as much as was predicted in 2014.

The single biggest cost to the wafer by far is depreciation expense, which is the depreciation cost of writing off the fab and equipment. Aka 4 billion dollars fabs equipment being written off in a few years. And them needing to buy new stuff for the new process.

This is generally set in stone well before everything is set up and thus is easy and accurate to predict.

The reason 10nm isn't as expensive per transistor as 16nm/14nm is because it is still finfet and much of the same equipment can see be carried over as a result. The big jump occurs at 16/14nm because that's when completely new equipment needed to be purchased.

However because of the 4x increase in transistor density of 10nm vs 28nm, there is still a 4x increase per die area since the cost per transistor of 10nm and 28nm is the same.

Also, going by the new a10 cost in the new iphone 7, which is $27, add the packaging cost which is 4 dollars or so, the costs are pretty much bang on considering the a10 is 124mm2.

 

[Pariah]

Source on pricing?

AMD said a long time ago when announcing Polaris that they were "Bringing VR performance to the masses at lower price points". Considering VR was 970/290 @ $300, why would you expect Polaris to launch near $300? They said that the jump to 14nm saved them a lot per die compared to 28nm which means they were able to charge less. You are contradicting them without any proof.

Stop repeating yourself over and over again with either irrelevant information that only wastes people's time. In the first video YOU linked in a previous post here, the part YOU linked to starts with AMD showing a slide with the price of 970/290x listed at $350. That is what AMD was targeting to beat. Despite you droning on and on about TAM and $200, and how those 2 videos somehow proved that, there is NO mention of any other price in either video; only the $350 970/290x I just noted.

If you look at this thread about the announcement of Polaris. The first 7 pages or so are people guessing the price before the announcement. Pretty much every poster guessed $250-$300. Including this gem:

If the cheapest good 1070 is $399, then Polaris 10 can be $299 and still be perfectly viable even if it's 25% slower...This is critical because $249-299 P10 will target key market segments that no $380-450 card is aimed at.

Remember that one Russian? Then you post this in this thread:

^ The entire analysis above would have made a lot more sense except for the fact that Polaris 10 is an R9 380/X market segment replacement. For the 100th time, we have 4 designating a new generation, and the 2nd number is the class of card it is replacing:

R9 380/X -> RX 480

R9 380 was a $199 videocard and R9 380X was a $229 videocard. These were not $299-349 videocards:

The other major flaw in the argument is that AMD already had a mid-range $329 R9 390 8GB a long time ago before Polaris 10 launched. How would AMD aim to expand TAM if they released a $299-$349 RX 480 and $249-279 RX 470 when they already were selling $250 R9 290 and $330 R9 390? The logic makes no sense.

The logic makes no sense, huh? Odd considering it was your own logic just before the launch of Polaris.

The consensus among people here was that the RX 480 would launch at a price of $250 to $300, which was acceptably below the $350 price of the 970 at the time. Prices ended up tanking to clear out inventory, but that is a temporary market condition that was irrelevant as far as new product launch prices go. Car makers don't drop the list price of 2017 models because 2016's are being sold well below MSRP to clear out inventory.

 

[krumme]

I cant see the flaw in rs logic.

What is pretty aparent imo is that if the rx480 was 300 the big die vega made absolutely no sense from a business perspective as amd sells very few cards at 350-400+. No damn sense. I argue that back then. And guess what. The 480 is like what 240. Now the lineup makes a little more sense.

I still very much question the business reason for big vega but at least there is a small market for amd at 300 to 450 now.

 

[krumme]

There ia another reason. 460 and 470/480 die simply have trouble covering the price segment from 120 to 240 already. Look how different 460 is to 470. 750ti have sold like crazy and ofcource amd wanted that segment to. It meant low 460 price. They simply didnt have anything to compete there. Then you just dont go from 750ti level to say 250 for a rx470. Its a huge gab where a lot of cards is sold.

And revenue and moving gf capacity is very much a target for amd because of the wsa. 120 to 240 segment does that in spades.

 

[Enigmoid]

I cant see the flaw in rs logic.

What is pretty aparent imo is that if the rx480 was 300 the big die vega made absolutely no sense from a business perspective as amd sells very few cards at 350-400+. No damn sense. I argue that back then. And guess what. The 480 is like what 240. Now the lineup makes a little more sense.

I still very much question the business reason for big vega but at least there is a small market for amd at 300 to 450 now.

By the time big die vega launches (probably almost a year after Polaris) its basically rebrand and price drop time.

Hawaii was sold completely above $400, so was Fiji. Vega is launching well after Polaris so price drops are always possible.

There ia another reason. 460 and 470/480 die simply have trouble covering the price segment from 120 to 240 already. Look how different 460 is to 470. 750ti have sold like crazy and ofcource amd wanted that segment to. It meant low 460 price. They simply didnt have anything to compete there. Then you just dont go from 750ti level to say 250 for a rx470. Its a huge gab where a lot of cards is sold.

And revenue and moving gf capacity is very much a target for amd because of the wsa. 120 to 240 segment does that in spades.

I'm not sure what AMD was trying to do with the 460. It uses too much power to cover the main mobile market. Not to mention its to inefficient to even compete on mobile but its not in a power range that most OEMs care about for the mass market (<45W). On the desktop the 460 is launching too close to the 950 as well and hovers around where Pitcarin is as well.

 

[Piroko]

Also, going by the new a10 cost in the new iphone 7, which is $27, add the packaging cost which is 4 dollars or so, the costs are pretty much bang on considering the a10 is 124mm2.

I have always assumed that this estimation was made with your link as basis, but I'll take your word for it. Still, even at a worst case scenario we should be talking about a total manufacturing cost increase of less than 30$ on a 200mm² chip (16ff vs. 28planar) for the first generation on that process. And AMD probably plans with a long product lifecycle of 24 months or more (I doubt they had planned for >24 months with Pitcairn) with all their talk about fewer numbers of SKUs per launch.

 

[Bacon1]

Stop repeating yourself over and over again with either irrelevant information that only wastes people's time. In the first video YOU linked in a previous post here, the part YOU linked to starts with AMD showing a slide with the price of 970/290x listed at $350. That is what AMD was targeting to beat. Despite you droning on and on about TAM and $200, and how those 2 videos somehow proved that, there is NO mention of any other price in either video; only the $350 970/290x I just noted.

If you look at this thread about the announcement of Polaris. The first 7 pages or so are people guessing the price before the announcement. Pretty much every poster guessed $250-$300. Including this gem:

Thanks for finding that link, I couldn't find it the other day.

Did you bother to watch it though?

@17:10 he talks about Polaris and how its cheaper to produce than their current 28nm chips.

@18:10 - "We can produce GPUs that run the minimum spec of VR (meaning 480 / 970 / 1060 specs) at a lower cost, in larger volume, consuming less power, and running faster."

 

[tajoh111]

I have always assumed that this estimation was made with your link as basis, but I'll take your word for it. Still, even at a worst case scenario we should be talking about a total manufacturing cost increase of less than 30$ on a 200mm² chip (16ff vs. 28planar) for the first generation on that process. And AMD probably plans with a long product lifecycle of 24 months or more (I doubt they had planned for >24 months with Pitcairn) with all their talk about fewer numbers of SKUs per launch.

The thing is 30 dollars is a tremendous deal because it means the difference between 15% margin and 50% margin.

From what I can gather, and considering the cost of apples chip, a full polaris chip cost somewhere along the lines of 70 dollars. A doubling in chip size make the price of a chip go up more than the size because the chance for a defect is exponentially greater. Add in apples engineers have a better budget and better process engineers because of the sheer volume and budget and 70 dollars is a fair estimate.

From previous BOM studies. The rest of the components to make the videocard in the same class is 66 dollars, with inflation lets call this 70.

Add in packaging costs for the card lets say is 7 dollars(same as iphone). This is 147 dollars, and this is the cost for the partner. To make a 15% profit to make money and cover their warranty service, this is 169.00. Add in retailer margin, and this adds another 10%. Add in logistics to get the cards where they need to go and AMD is making 10 dollars on this card.

(80-70/80)*100 = 12.5% margin.

Now if AMD was selling the same card or 250, it would be making 60 dollars on each card. This would mean AMD is making

130-70/130* 100= 46.1% profit margin.

This is the type of margin you need to make money or atleast not lose it. Nvidia with a 46.1% profit margin would basically just break even.

The problem with low teen profit margins is if prices fall beyond this threshold and the card has the potential to be sold at a loss.

Using a similar estimate and estimating the potential savings between using a cut wafer, Basically the same data but the cost for the rx470 chip is 58, AMD is making 6 dollars on a rx470.

This explains why the rx470 is not priced lower. There is simply no room for it to be priced lower and why AMD has set up pricing which cannibalizes its own sales.

The rx 470 prices at 200 would leave AMD with a 35% margin, rather than a 9% margin at 180.

If AMD could add another 50 to 20 dollars to polaris it would do wonders for their profitability and greatly add to the segmentation of their products.
$299 rx 480 8gb
$259 rx 480 4gb
$229 rx 470 8gb
$209 rx 470 4gb

139 rx 460 4gb

120 rx460 2gb

Is a much more sounds product stack from a business perspective than what AMD has right now.

If AMD picked their current pricing with out any reaction from Nvidia, they would simply have to be morons.

A 70 dollar price difference between the rx460 and 470 makes the rx 460 a somewhat viable option.

The problem right now iis the rx470 is 70% faster than a rx460 and it only costs 40 dollars more or 28% more money. This basically makes the rx 460 unviable. It's one of the poorest rated cards in history and this is because it is priced to close to the rx 470 series. The same problem occurs with the rx470 to a lesser degree vs the rx480 4gb.

Thanks for finding that link, I couldn't find it the other day.

Did you bother to watch it though?

@17:10 he talks about Polaris and how its cheaper to produce than their current 28nm chips.

@18:10 - "We can produce GPUs that run the minimum spec of VR (meaning 480 / 970 / 1060 specs) at a lower cost, in larger volume, consuming less power, and running faster."

Cheaper and lower cost than gm204 and hawaii. This doesn't mean there is enough savings to cut their costs in half. These cards sold for 330 -399 for their cut chips and 550 for their full chips during their first run.

Amd is selling their full chip at 199 now and 179 for their cut chip. That's a 350 to around a 200 dollar cut. Using a smaller die, brings some savings due to yields and a cheaper PCB/cooler because of less heat but not enought to cut the costs that much.

 

[raghu78]

The problem right now iis the rx470 is 70% faster than a rx460 and it only costs 40 dollars more or 28% more money. This basically makes the rx 460 unviable. It's one of the poorest rated cards in history and this is because it is priced to close to the rx 470 series. The same problem occurs with the rx470 to a lesser degree vs the rx480 4gb.

AMD's Rx 460 is obviously poor in terms of perf/$. But thats primarily because Nvidia has not filled out its Pascal stack. once the GTX 1050 and 1050 Ti launches AMD will be forced to move the pricing of Rx 460 to $99 for 2GB and $119 for 4GB. I feel AMD's Polaris might do better once a second revision launches in H1 2017 with clock/power issues fixed. AMD's Rx series naming has already hinted at second revisions of the Polaris chips.

http://videocardz.com/61721/amd-radeon-rx-400-series-naming-scheme-explained

Similarly if Vega is really good we could see price cuts on GTX 1080 and faster models like GTX 1080 Ti launch at USD 700. Thats why its essential that AMD have very competitive products as it benefits the consumer.

 

[Magee_MC]

I'm probably missing something obvious, but when you talk about the 470 cannibalizing 460 sales and the 480 eating 470 sales, I don't see why that is a problem. People are effectively being upsold which is usually considered a good thing. I could see it being a huge problem if the 460's perf and price were such that it was stealing away 470 sales, but going the other direction seems like it would be ok from AMD's perspective.

The thing is 30 dollars is a tremendous deal because it means the difference between 15% margin and 50% margin.

From what I can gather, and considering the cost of apples chip, a full polaris chip cost somewhere along the lines of 70 dollars. A doubling in chip size make the price of a chip go up more than the size because the chance for a defect is exponentially greater. Add in apples engineers have a better budget and better process engineers because of the sheer volume and budget and 70 dollars is a fair estimate.

From previous BOM studies. The rest of the components to make the videocard in the same class is 66 dollars, with inflation lets call this 70.

Add in packaging costs for the card lets say is 7 dollars(same as iphone). This is 147 dollars, and this is the cost for the partner. To make a 15% profit to make money and cover their warranty service, this is 169.00. Add in retailer margin, and this adds another 10%. Add in logistics to get the cards where they need to go and AMD is making 10 dollars on this card.

(80-70/80)*100 = 12.5% margin.

Now if AMD was selling the same card or 250, it would be making 60 dollars on each card. This would mean AMD is making

130-70/130* 100= 46.1% profit margin.

This is the type of margin you need to make money or atleast not lose it. Nvidia with a 46.1% profit margin would basically just break even.

The problem with low teen profit margins is if prices fall beyond this threshold and the card has the potential to be sold at a loss.

Using a similar estimate and estimating the potential savings between using a cut wafer, Basically the same data but the cost for the rx470 chip is 58, AMD is making 6 dollars on a rx470.

This explains why the rx470 is not priced lower. There is simply no room for it to be priced lower and why AMD has set up pricing which cannibalizes its own sales.

The rx 470 prices at 200 would leave AMD with a 35% margin, rather than a 9% margin at 180.

If AMD could add another 50 to 20 dollars to polaris it would do wonders for their profitability and greatly add to the segmentation of their products.
$299 rx 480 8gb
$259 rx 480 4gb
$229 rx 470 8gb
$209 rx 470 4gb

139 rx 460 4gb

120 rx460 2gb

Is a much more sounds product stack from a business perspective than what AMD has right now.

If AMD picked their current pricing with out any reaction from Nvidia, they would simply have to be morons.

A 70 dollar price difference between the rx460 and 470 makes the rx 460 a somewhat viable option.

The problem right now iis the rx470 is 70% faster than a rx460 and it only costs 40 dollars more or 28% more money. This basically makes the rx 460 unviable. It's one of the poorest rated cards in history and this is because it is priced to close to the rx 470 series. The same problem occurs with the rx470 to a lesser degree vs the rx480 4gb.



Cheaper and lower cost than gm204 and hawaii. This doesn't mean there is enough savings to cut their costs in half. These cards sold for 330 -399 for their cut chips and 550 for their full chips during their first run.

Amd is selling their full chip at 199 now and 179 for their cut chip. That's a 350 to around a 200 dollar cut. Using a smaller die, brings some savings due to yields and a cheaper PCB/cooler because of less heat but not enought to cut the costs that much.

The thing is 30 dollars is a tremendous deal because it means the difference between 15% margin and 50% margin.

From what I can gather, and considering the cost of apples chip, a full polaris chip cost somewhere along the lines of 70 dollars. A doubling in chip size make the price of a chip go up more than the size because the chance for a defect is exponentially greater. Add in apples engineers have a better budget and better process engineers because of the sheer volume and budget and 70 dollars is a fair estimate.

From previous BOM studies. The rest of the components to make the videocard in the same class is 66 dollars, with inflation lets call this 70.

Add in packaging costs for the card lets say is 7 dollars(same as iphone). This is 147 dollars, and this is the cost for the partner. To make a 15% profit to make money and cover their warranty service, this is 169.00. Add in retailer margin, and this adds another 10%. Add in logistics to get the cards where they need to go and AMD is making 10 dollars on this card.

(80-70/80)*100 = 12.5% margin.

Now if AMD was selling the same card or 250, it would be making 60 dollars on each card. This would mean AMD is making

130-70/130* 100= 46.1% profit margin.

This is the type of margin you need to make money or atleast not lose it. Nvidia with a 46.1% profit margin would basically just break even.

The problem with low teen profit margins is if prices fall beyond this threshold and the card has the potential to be sold at a loss.

Using a similar estimate and estimating the potential savings between using a cut wafer, Basically the same data but the cost for the rx470 chip is 58, AMD is making 6 dollars on a rx470.

This explains why the rx470 is not priced lower. There is simply no room for it to be priced lower and why AMD has set up pricing which cannibalizes its own sales.

The rx 470 prices at 200 would leave AMD with a 35% margin, rather than a 9% margin at 180.

If AMD could add another 50 to 20 dollars to polaris it would do wonders for their profitability and greatly add to the segmentation of their products.
$299 rx 480 8gb
$259 rx 480 4gb
$229 rx 470 8gb
$209 rx 470 4gb

139 rx 460 4gb

120 rx460 2gb

Is a much more sounds product stack from a business perspective than what AMD has right now.

If AMD picked their current pricing with out any reaction from Nvidia, they would simply have to be morons.

A 70 dollar price difference between the rx460 and 470 makes the rx 460 a somewhat viable option.

The problem right now iis the rx470 is 70% faster than a rx460 and it only costs 40 dollars more or 28% more money. This basically makes the rx 460 unviable. It's one of the poorest rated cards in history and this is because it is priced to close to the rx 470 series. The same problem occurs with the rx470 to a lesser degree vs the rx480 4gb.



Cheaper and lower cost than gm204 and hawaii. This doesn't mean there is enough savings to cut their costs in half. These cards sold for 330 -399 for their cut chips and 550 for their full chips during their first run.

Amd is selling their full chip at 199 now and 179 for their cut chip. That's a 350 to around a 200 dollar cut. Using a smaller die, brings some savings due to yields and a cheaper PCB/cooler because of less heat but not enought to cut the costs that much.

 

[Piroko]

If AMD could add another 50 to 20 dollars to polaris it would do wonders for their profitability and greatly add to the segmentation of their products.

Well that's a given. However, do take note that I still consider all of your theorizing to rely on a single speculative source from 2014 on a fabrication process that isn't used for the product you theorize over. Even a small variation will lead to a huge error in your viability assessment.
Couple of additional thoughts:
1) It is known that Samsungs 14ff is a little bit denser than TSMCs 16ff (see Apple A9). Polaris 10 on TSMCs process would theoretically yield around ~228 chips (est. 15.8 * 16.0 for 253mm²) per wafer whereas it yields ~250 chips (est. 15.0 * 15.5 for 232mm²) on 14ff. That would be a 10% cost advantage for GloFo at equal wafer prices (assumption).
2) AMD has to fill its WSA with GloFo. That doesn't change the cost structure per chip directly, but it does change the viability of selling a product at low margins. At least until another product with the potential for higher margins can fill its place (Zen? Vega?).
3) Polaris is a replacement for the 360mm² Tonga GPU in name and for the 440mm² Hawaii GPU in performance. I think we can both agree that Polaris is a more viable product to compete in the 200$ market than either of these in 2016.

 

[tajoh111]

I'm probably missing something obvious, but when you talk about the 470 cannibalizing 460 sales and the 480 eating 470 sales, I don't see why that is a problem. People are effectively being upsold which is usually considered a good thing. I could see it being a huge problem if the 460's perf and price were such that it was stealing away 470 sales, but going the other direction seems like it would be ok from AMD's perspective.

The problem is supply. Basically those cards have a much larger supply of cards and those cards need to be sold to maximize revenue.

Lets demonstrate this with a hypothetical math example. We have a pool of 17k buyers.

And the supply of the merchandise is as follows.

We have 10k rx 460 4gb 140

We have 5k rx 470 180

and 2k rx480s. 200.

Revenue = volume * price.

To maximize revenue we want to supply to equal demand under all three products. With AMD product stack being so bad, what happens is 1k out of that 10k rx 460's get sold because the rx460 is such poor value proposition card. This leaves about 3k who absolutely cannot afford the step up to the rx480, and the entire stock of rx 480 sells out because demand outstrips supply.

AMD's revenue is usually just the chip sale and not the whole value of the card sale. However, for simplicity sake lets just take the msrp as the revenue.

At those volumes and selling points, AMD is generating $140,000 from rx 460 sales, $540,000 and $400,000 from rx480 sales. In total this is $1,080,000. The problem with this is the price AMD wants to sell at is a point where supply and demand match up so its can sell its entire supply of goods. With the desirability of rx 480 over it's siblings, it takes sales away from those line and they just sit on the shelf not generating AMD revenue, while the rx 480 sells out. The problem is the supply of the rx480 is the most finite. And there is a huge loss in potential revenue.

The potential revenue if the entire supply was sold is $1,400,000 from rx 460s, $900,000 from rx470's and $400,000. This is $2,700,000. What is happening is AMD is taking a loss of revenue of $1,620,000. This is termed a deadweight loss. This happens when supply does not meet demands due to externalities. This can be caused by outside factors like tax, but in this case AMD is doing it to themselves with their bad product pricing. In this case, the rx480 decrease the percieved value of the rx 460 and rx 470. This results in a huge decrease in sales. This results in a loss in sales and thus an overall decrease in revenue which is the deadweight loss of $1,620,000.

If AMD did this consciously and it wasn't a forced decision, they are literally morons. I don't think they are.

Why this is happening is more than likely a combination of floor pricing and competition from Nvidia. Basically AMD can't price their products at a level where supply meets demands on all its product lines because if they price too high, they lose to Nvidia because of superior performance per dollar and the good will and value of the Nvidia brand. If they price too low of course, they can lose money on their products.

And this is why the rx 480 is priced the best. The rx 480 is priced at a point where it offers better price to performance than Nvidia's product because it needs to because of the inferiority of the AMD brand to the general public. However the rx 470 needs to be priced a certain distance away so demand and supply meet an equilibrium. This has to be a significant amount, other wise cannibalization of sales occur. This is exactly why there is a 200 dollar difference between the gtx 1080 and gtx 1070. What is happening and why AMD has only a 20 dollar difference between the two lines is the rx470 is running to close to it's floor pricing, which is the minimum price AMD needs to sell at to make money. This is preventing AMD from reaching equilibrium pricing.

Same thing is happening to the rx460. This is why Polaris needs to be priced higher. A higher price for flagship polaris would allow AMD to price their products at points where supply meets demands across all lines and revenue is maximized. Having closely price products, particularly when the most desireable products, is the most expensive, is a no no, in business and economics. It violates the basic principles of supply and demand.

Well that's a given. However, do take note that I still consider all of your theorizing to rely on a single speculative source from 2014 on a fabrication process that isn't used for the product you theorize over. Even a small variation will lead to a huge error in your viability assessment.
Couple of additional thoughts:
1) It is known that Samsungs 14ff is a little bit denser than TSMCs 16ff (see Apple A9). Polaris 10 on TSMCs process would theoretically yield around ~228 chips (est. 15.8 * 16.0 for 253mm²) per wafer whereas it yields ~250 chips (est. 15.0 * 15.5 for 232mm²) on 14ff. That would be a 10% cost advantage for GloFo at equal wafer prices (assumption).
2) AMD has to fill its WSA with GloFo. That doesn't change the cost structure per chip directly, but it does change the viability of selling a product at low margins. At least until another product with the potential for higher margins can fill its place (Zen? Vega?).
3) Polaris is a replacement for the 360mm² Tonga GPU in name and for the 440mm² Hawaii GPU in performance. I think we can both agree that Polaris is a more viable product to compete in the 200$ market than either of these in 2016.

It's hardly speculative when guys like samsung and TSMC use these types of services to determine how much they need to charge for wafers.

Samsungs process and TSMC process ending up being very similar in price per gate because of one thing. Yields. The smaller and denser chips are packed the worse the yields get. The more logic that is squished together in a smaller area, the more likely things go wrong. If you look at other papers, you will see this. I.e 65nm has better yields than 40nm, 40nm's has better yields for the same area than 28nm. Add in GF is paying licencing fees to Samsung and because GF is 1/5 to 1/7 the size of TSMC and Samsung in terms of volume, economies of scale work against them.

AMD agreement to GF, has been a pain the bum for them. I wish ATIC with all it's backing, gave AMD a break. But these take or pay agreements have been killing AMD slowly.

Polaris is cheaper to produce than hawaii but I don't think its that much cheaper to produce than Tonga. What needed to happen is AMD cards needed to increase in performance vastly so the price of their products could go up. This is what helps justify a node switch. AMD pricing needs to go up so their gross margins can go up.

 

[USER8000]

Prior to the gtx 1080 launch, if Nvidia didn't launch anything, we could see history repeat itself with AMD having higher pricing prior to Nvidia launching kepler.

The rx480 was performing along the lines up gtx 970 range and slightly better and also similar to the r9 390, another 300-330 dollar card. Both these cards were great value for their generation. So it would be no surprise if AMD was not aggresive with pricing vs them. A 250-260 dollar card would definitely be midrange(4gb model) and it would have provided AMD with the buffer it needs to prevent canibalization.

A rx 480 at 250-259 for the 4gb and 299 for the 8gb, would have been par for the course, if AMD previous behavior is anything to go by. In the last 5 years, the only aggressively priced card from AMD considering it's competition was Hawaii. Hawaii and arguably the fury none x has been the only card on 28nm that has forced Nvidia to lower their prices.

The 7970, 7870, r9 285, fury x, nano came priced high and it was Nvidia cards, that forced them to do price cuts(gk104, gm104 and the gtx 980 ti).

What AMD fans forget is Nvidia puts pressure on AMD too. It's not a one way street and its one o the reasons why Nvidia was able to pick up so much marketshare during 28nm.

Your revising the bold part or telling a lie. What caused the 7870 to drop in price was the gtx 670.
It was 50 dollars more than a 7870 and performed better than a 7950. Hence 7950 prices had to fall and 7870 prices had to fall accordingly. The gtx 660 ti came out well after AMD's price drops including the 7870. Particularly when street prices were belows AMD official msrp.

And it adds more evidence to my case that AMD cut prices to polaris actual launch. Nvidia gx104 series forced price cuts everywhere in AMD lineup including their pitcairns cards.

The thing with the 6870 is that it was made on a very mature node that was dirt cheap. And because it performed worse than a 5870 it had to be priced below it. 40nm and 14/16nm finfet production cost is vastly difference which adds more evidence for my theory. For the die size Polaris is the cheapest on a new node ever. As a node matures, it becomes progressively cheaper to make larger and largers chips on it. the 6870 was made at the peak of maturity on 40nm's. During this generation Nvidia could get away with selling 5xxmm2 chips at 499 and 350. They could't do this anymore. And these wafers being around 3000 dollars and the yields being at the best they are going to get, makes putting larger chips profitable at lower price ranges.

Polaris is the complete opposite. It is made on a brand spanking new 14nm LPP finfet design. Finfet wafers costs between 7000-8500 dollars which made cost per transistor not decrease vs last generation. What his means is if yields were the same, a 400mm2 chip 28nm is the same cost as a 200mm2 finfet chip. The only savings come in the form of yields which is part of every smaller chip. This is well documented. Prior to finfet, wafers might go 15%-20% but because transistor density doubled there was a net savings.

This brings me back to my point. The problem is AMD is charging historically low prices for their 2xxmm2 die(they haven't been this low since the 3870(the 6870 1gb is more equivalent to the 4gb rx480), at a time when production cost is at record high levels. It's a contradiction meaning AMD is not making much money on their cards.

Add in the other wrinkles in their pricing and you can tell AMD is hitting near the floor of where they begin to lose money and its why its price to performance is so screwed up.

The rx 480 4gb is 199 and the rx470 4gb is 179. The rx 480 requires a perfect die which makes it a far rarer die than a rx470. The higher the price of the rx 480, the lower the demand for the card is. What is going to happen when their is only a 20 dollars difference between the two cards?

Basically the rx480 takes demand and sales away from the rx470 and because there is not enough difference in price between the two cards. And because the rx480 is much more supply limited than the rx470, there is a huge loss in potential revenue.

Ideally what you want is there to be enough of a price gap between the two cards that rx480 sales meet the supply and the same with the rx470. What you have in this situation is because the rx 480 is so closely priced to the rx470, most people prefer it and skip the rx470 to get the rx480. What happens as a result is the rx480 supply runs out, and you lose a tremendous amount of rx470 sales, because those people wait for the rx480 to become in stock again

The situation is even worse for the rx460 4gb. The rx 470 4gb is 70-80% faster than an rx 460 4gb, but there is only a 40 dollar difference in price. As a result no one besides the OEMS bite on the rx 460 because it's a no brainer to spend a little more to get vastly better performance. And the cannibalization of sales continue.

This very concept is why there has been a minimum of 50 difference between the x870 and x850 cards and sometimes 100 dollars. Absolutely never has there been just 20 dollars difference between a cut die and a full die. Even memory upgrades cost more than this. The only explanation for this is the rx470 hits floor pricing.

To all the fanboys questioning this, why has AMD set up their pricing so closely space when there will undoubtedly be cannibalizing their own sales?

It's basic marketing, economics and business.

What your wall of text seems to not get is that AMD was selling the R9 390 at the price-points the RX480 8GB is selling at. So to re-cap:
1.)438MM2 GPU
2.)8GB of GDDR5 accross a 512 bit memory bus
3.)Bigger coolers due to much higher TDP
4.)Much more complex PCB due to 512 bit memory controller
5.)More complex power distribution circuitry.

The 232MM2 in Polaris 10 is nearly half the size of the GPU in the R9 390,the RX480 uses half the number of memory chips,and has a less complex PCB and requires less cooling. If anything from what I remember reading here,a lot of that increase was apparently the R9 380 and R9 380X.

Even a 50% increase in cost per transistor would still make Polaris 10 based cards cheaper to make than Hawaii or Tonga based ones . AMD has WSA too,so for them to fulfill it with CPU sales not being so hot,at least helps more in that regard longterm. If anything from what I remember reading here,a lot of the marketshare increase was apparently the R9 380 and R9 380X(could be wrong there). But as you know the R9 380 and R9 380X are using 359MM2 GPUs too and tend to be generally higher in power consumption and cooling requirements,which increases costs.

The RX480 and RX470 have also replaced the R9 380,R9 380X and R9 390. The whole AMD £150 to £300 price-range is now only served by one GPU instead of two.

But OFC as I mentioned earlier,you and your mates spin away that AMD increasing shipping marketshare is a bad thing. They have simplified their whole range whilst making the replacement cards cheaper to make,and at the same time making sure they can try and get more wafers used at Global Foundries longterm.

Also my current card is a GTX960 and my previous one was a GTX660TI,so not sure what the whole point of you bolding that sentence is for.

 

[Piroko]

It's hardly speculative when guys like samsung and TSMC use these types of services to determine how much they need to charge for wafers.

It is speculative in that it talks about the first 16ff process from TSMC that (to my knowledge) didn't see broad adoption and was quickly replaced by the more attractive 16ff+ and several deviations on that. This should be another hint that something major has changed in the cost relation.
By the way, just stumbled over this image:
http://imgur.com/eezbRGE
That would indicate a healthy decrease of cost per million transistors at GF.

Samsungs process and TSMC process ending up being very similar in price per gate because of one thing. Yields. The smaller and denser chips are packed the worse the yields get. The more logic that is squished together in a smaller area, the more likely things go wrong. If you look at other papers, you will see this. I.e 65nm has better yields than 40nm, 40nm's has better yields for the same area than 28nm. Add in GF is paying licencing fees to Samsung and because GF is 1/5 to 1/7 the size of TSMC and Samsung in terms of volume, economies of scale work against them.

Yes, that is a competitive disadvantage for GF, but they would be crazy to pass that down to their customers.

As for yield curves, they seem to be okay for both foundries. We would have heard more from unhappy customers otherwise, like with yield issues of 20nm process. GPUs are also rather well suited for binning and there is a large enough performance gap between the 470 and 460 to fit another China-only bin in (for example).

Polaris is cheaper to produce than hawaii but I don't think its that much cheaper to produce than Tonga.

Tonga is a beefier PCB compared to Polaris:
Tonga
Polaris
More phases, 50% larger PCB, more secondary components. Almost same amount of transistors. And it needs a beefier cooler. I'm quite sure that Polaris beats its parts and assembly cost by a decent margin.

 

[krumme]

The pcb of ref 470 is so tiny it reminds me of ati 4670. Nothing remotely like the 970 7850 or 7970 i have had recently. Pretty sure its fine business unlike all the cpu stuff. Gets used in consoles. Hell its probably one of the only part earning some real money besides the consoles.

 

[Bacon1]

Cheaper and lower cost than gm204 and hawaii. This doesn't mean there is enough savings to cut their costs in half. These cards sold for 330 -399 for their cut chips and 550 for their full chips during their first run.

Amd is selling their full chip at 199 now and 179 for their cut chip. That's a 350 to around a 200 dollar cut. Using a smaller die, brings some savings due to yields and a cheaper PCB/cooler because of less heat but not enought to cut the costs that much.

You have no actual numbers but you somehow think that you are correct and AMD doesn't know how to price their own products to make money?

 

[frozentundra123456]

You have no actual numbers but you somehow think that you are correct and AMD doesn't know how to price their own products to make money?

Actually, that is pretty much true. A good quarter for AMD is when they lose less than they have in the past.

 

[Bacon1]

Actually, that is pretty much true. A good quarter for AMD is when they lose less than they have in the past.

So I guess Polaris does make them more money since they are losing less now, case solved!

 

[Pariah]

So I guess Polaris does make them more money since they are losing less now, case solved!

I take it math isn't your specialty. Throughout this whole thread you have denied that $300 is less than $350. Now you're trying to tell us that AMD losing $81 million in Q2 is better than losing $70 million in Q1. When there is a negative sign in front of a dollar amount on a financial report, the bigger the number after the negative sign, the worse it is.