Low cost HBM on the way, will hit mass market soon Plus HBM3 teased up to 64GB

[Greebo]

The real point is no one is using HBM1 0r HBM2 on a gaming card, not even a Pascal Titan X.

Maybe this is to do with price or maybe it is to do with performance. Whichever way you want to look at it both AMD and NVidia have decided to use other solutions for their gaming cards.

Nobody sensible is using HBM because its limited to 4Gb so no good for hgh end cards and for low end cards, they dont need the bandwidth or the power saving so pointless either way.

HBM2 is expensive and as it turned out the bandwidth wasn't needed on Pascal, neither was the power reduction.

Some may argue against that to a degree since Pascal gpus are more limited by hitting their power limits rather than heat as fair as overclocking goes.

With Nvidia's next card you are supposed to have between 16Gb and 32Gb and with the bandwidth needed, gddr5x would just take up too much power to give you the bandwidth so you dont cripple your gpu.

 

[Greebo]

The acid lining to all this is that it implies that HBM2 is too expensive to use in standard consumer GPU's.

That would be a huge shame.

Not really. I suspect a 1080 only costs $100 in components. Would it really matter if with HBM2 it would cost $150?

It only matters if then instead of $600 Nvidia then want $900 for the final card. If they made it $650 people would still buy and wouldnt mind the $50 increase.

 

[drunkenmaster]

Nobody sensible is using HBM because its limited to 4Gb so no good for hgh end cards and for low end cards, they dont need the bandwidth or the power saving so pointless either way.

HBM2 is expensive and as it turned out the bandwidth wasn't needed on Pascal, neither was the power reduction.

Some may argue against that to a degree since Pascal gpus are more limited by hitting their power limits rather than heat as fair as overclocking goes.

With Nvidia's next card you are supposed to have between 16Gb and 32Gb and with the bandwidth needed, gddr5x would just take up too much power to give you the bandwidth so you dont cripple your gpu.

Nvidia had to move away from a GP100 with HBM2 and a large die and move down to a 470mm^2 core and stick it with GDDR5x. What makes you think they made a slower than possible Titan on this gen, for absolutely no reason or they made the biggest die possible within the confines of being fed by gddr5x?

Had they made the bus much wider, it would have used more power and that would have allowed them to feed more shaders but those would need more power also, add those together and you can't fit any more shaders in that what they achieved.

That design is fundamentally power and bandwidth limited... as all cards are. HBM2 would have let them add more shaders, more bandwidth and without increasing power. The problem is for them that HBM2 was going to be available later in this year.

At the cost Nvidia are selling Titan X(seriously that still gets me, lets not do a new name) HBM2 costs wouldn't be an issue except it would reduce the profit on them. Nvidia users just sit around not caring of Nvidia make an utter killing on a card or a fair profit and a fair price though. As such Nvidia have no reason to provide better performance for a fairer price when they can make as much/more money by over charging for something that has a higher profit margin.


Nvidia may have 'launched' a HBM2 card already, but you know, it's unavailable and afaik they still have only showed a single working demo and that was what, a few weeks ago. Even still that demo wasn't shown at an event but was on a bunch of GP100's back in a lab that they connected to. They are so 'available' that Nvidia has yet to show one working publicly yet and every version they've mentioned(Tesla, Quadro) is only stated availability either late Q4 or Q1 next year(and we still don't know they'll hit those dates at all).

 

[Seanspeed]

Not really. I suspect a 1080 only costs $100 in components. Would it really matter if with HBM2 it would cost $150?

It only matters if then instead of $600 Nvidia then want $900 for the final card. If they made it $650 people would still buy and wouldnt mind the $50 increase.

I dont know why you assume this all so cheap. :/

 

[bru]

Not really. I suspect a 1080 only costs $100 in components. Would it really matter if with HBM2 it would cost $150?

It only matters if then instead of $600 Nvidia then want $900 for the final card. If they made it $650 people would still buy and wouldnt mind the $50 increase.

There is no way the components on a 1080 0nly cost $100, the GPU itself probably costs that alone.

 

[Greebo]

There is no way the components on a 1080 0nly cost $100, the GPU itself probably costs that alone.

Sales $5 billion

Cost of goods inc R&D = $2 billion

R&D spend = $1.3 billion

Net cost of goods = $700m or 14% of sales.

Not sure if cost of goods include packaging etc but lets say it doesn't.

So with a 1080 being $599 list price the cost of components is $83.86 and even including R&D costs back its only $239 per card. Of course that for the cards they sell direct. They have to sell to wholesales who take a cut, transport costs and then the final etailer like ocuk making a margin as well so their final sales price wont be anywhere near MSRP.

Of course profit margin will vary across the range but in general you are only looking at $80 to $200 in costs including assembly on even the high end cards so actually component costs will be less.

You have to remember Nvidia like other tech companies make a massive margin. I'm sure I've read somewhere the iphone costs less than $100 in components yet sells for $800.

 

[KillBoY_UK]

Not really. I suspect a 1080 only costs $100 in components. Would it really matter if with HBM2 it would cost $150?

It only matters if then instead of $600 Nvidia then want $900 for the final card. If they made it $650 people would still buy and wouldnt mind the $50 increase.

even if that was true, which its not. Entire lines of graphics chips cost billions of USD and 4/5/6+ years in R&D they need to get back.

Its like the BOM of a disc based game console game is a couple of quid, the cost of development is 10's or 100's of millions and just as much on marketing.

 

[Greebo]

I dont know why you assume this all so cheap. :/

Because of Nvidia's accounts which states their sales is 7 times more than their cost of goods???????

 

[Dave2150]

There is no way the components on a 1080 0nly cost $100, the GPU itself probably costs that alone.

If you think a small GP104 die costs $100 alone then you're on another planet, brother.

 

[Greebo]

even if that was true, which its not. Entire lines of graphics chips cost billions of USD and 4/5/6+ years in R&D they need to get back.

Its like the BOM of a disc based game console game is a couple of quid, the cost of development is 10's or 100's of millions and just as much on marketing.

I have already included their £1.3 billion R&D costs.

 

[Greebo]

If you think a small GP104 die costs $100 alone then you're on another planet, brother.

Well how does Nvidia make 7 times in sales their cost of goods then?

Or are you telling me their $90 cards only cost $5 to make to compensate for the higher end cards which cost lets say, 50% of their retail?

 

[Razor Time]

This chart is for older cards but might give an idea as to what things might cost in terms of raw materials.

 

[Greebo]

Profit margins have improved massively from then. If you look at Nvidia accounts, cost of goods as a percentage of sales has dropped by a third.

Granted revenue from cards will have increased as list price has increased so therefore the costs of the components has not increased from the gtx580 days in order for their margin to be what it is now.

Certainly R&D has increased massively and is now at $345m a quarter and it used to be only $200m a quarter a year or two ago. But thats R&D into the next gpu, not the current pascal series.

 

[Razor Time]

Profit margins have improved massively from then.

Totally agree. If you look at todays 1080 cards they are more in line in terms of bus size and die size to the old 460/560 cards.

I would interested to see a similar chart for the last few generations of cards including this one.

 

[Seanspeed]

Sales $5 billion

Cost of goods inc R&D = $2 billion

R&D spend = $1.3 billion

Net cost of goods = $700m or 14% of sales.

Not sure if cost of goods include packaging etc but lets say it doesn't.

So with a 1080 being $599 list price the cost of components is $83.86 and even including R&D costs back its only $239 per card. Of course that for the cards they sell direct. They have to sell to wholesales who take a cut, transport costs and then the final etailer like ocuk making a margin as well so their final sales price wont be anywhere near MSRP.

Of course profit margin will vary across the range but in general you are only looking at $80 to $200 in costs including assembly on even the high end cards so actually component costs will be less.

You have to remember Nvidia like other tech companies make a massive margin. I'm sure I've read somewhere the iphone costs less than $100 in components yet sells for $800.

R&D costs for a GPU go well beyond singular yearly expenses.

Besides, you're acting like the desktop gaming GPU's are the only thing that's included in these figures, much less their higher end GPU's.

I also see somebody bringing up costs for AMD cards from 2009 as proof of something. Costs have risen considerably since then. Especially for the chip manufacturers. Which would indicate that these figures would be much higher today and that $100 is a very, very terrible guesstimate as to the cost of a brand new process-based, non-cut down, GP104, complete product.

 

[Razor Time]

I also see somebody bringing up costs for AMD cards from 2009 as proof of something.

I just posted a chart showing what gpus cost to make in raw parts at a point in time and nothing more. If there was a more upto date chart I would post it. The chart also shows nvidia gpus from Nov 2010.

If you have a more upto date chart or a break down of what things cost then add it to the conversation.

 

[Greebo]

R&D costs for a GPU go well beyond singular yearly expenses.

Besides, you're acting like the desktop gaming GPU's are the only thing that's included in these figures, much less their higher end GPU's.

I also see somebody bringing up costs for AMD cards from 2009 as proof of something. Costs have risen considerably since then. Especially for the chip manufacturers. Which would indicate that these figures would be much higher today and that $100 is a very, very terrible guesstimate as to the cost of a brand new process-based, non-cut down, GP104, complete product.

Except from that chart you can see cost of components was a maximium of 40% of retail MSRP

If you look at Nvidias accounts since those days, the cost of components to retail sales price has been dropping dramatically as a percentage and is now at least 33% less.

therefore, its not 40% anymore which would be $240 for a 1080gtx, it is $161 based on their current percentage.

If it isnt the right percentage for the high end cards then they have massively reduced the cost of components on lowend cards in order to get their overall cost of components down so low.

Its not rocket science. If your sales are 5 billion and your costs of goods are only 700m, the cards cost peanuts to make.

Yes there may be years and years of R&D to recoup but my point was that if they needed to spend another $20 on memory it wouldnt be a lot in the overall scale of things. Of course products are priced on margins not a fixed amount of profit so $20 on components would mean $140 on the retail price.

 

[Seanspeed]

therefore, its not 40% anymore which would be $240 for a 1080gtx

A far cry from the $100 you originally posited.

Pricing at 2-2.5x what it costs you is entirely common in the sales industry.

Yes there may be years and years of R&D to recoup but my point was that if they needed to spend another $20 on memory it wouldnt be a lot in the overall scale of things.

Yes, there *may* be years and years of R&D you're not considered whatsoever. Kinda like seriously important to mention.

And why do you assume it's only an extra $20 for HBM? You seem to be purposefully low-balling everything and I dont get it.

 

[henson0115]

A far cry from the $100 you originally posited.

Pricing at 2-2.5x what it costs you is entirely common in the sales industry.


Yes, there *may* be years and years of R&D you're not considered whatsoever. Kinda like seriously important to mention.

And why do you assume it's only an extra $20 for HBM? You seem to be purposefully low-balling everything and I dont get it.

You may need to reread greebos post

 

[Greebo]

A far cry from the $100 you originally posited.

Pricing at 2-2.5x what it costs you is entirely common in the sales industry.


Yes, there *may* be years and years of R&D you're not considered whatsoever. Kinda like seriously important to mention.

And why do you assume it's only an extra $20 for HBM? You seem to be purposefully low-balling everything and I dont get it.

$100 was just a wild guess tbh before I looked at their accounts. I admit it's more likely to be $200. No I considered it. Nvidia spend 1.3 billion a year on r and d. I mentioned it twice in earlier posts. They used to spend 800m a year until this last year. Each year they bring new cards put so far enough that most r & d costs is only going to be a year or two max so around 2 billion.

But you are ignoring my points. I'm only looking at component cost. People are saying its hundreds of dollars just for the GPu alone. On their basis a 1080 would cost around $400 for the components plus there is the r&d to recoup on top yet they are currently selling on average for four times their component costs.

So either there is a tonne of cards which they sell for 8 times their costs or the 1080 costs no more than $200 in components costs.

You can't make 2 billion from 5 billion sales and spend 1.3 billion on r&d if your cards cost $300 to $400