Why have AMD APUs failed on the market?

[monstercameron]

Except on desktop, I don't see a large iGPU reducing BOM over a small iGPU. (And as I mentioned in my original post here using a larger iGPU to increase performance per watt is most useful for low voltage mobile)



If wanting to spend additional die area on desktop I think it is better to keep the small iGPU and instead increase CPU (up to a reasonable point). In the case of AMD's construction cores I think up to hexcore is a reasonable point.

Bom or build of materials. A larger igp means no need for a dgpu and higher end psu, hence lowering Bom and Bom costs.

 

[cbn]

Bom or build of materials. A larger igp means no need for a dgpu and higher end psu, hence lowering Bom and Bom costs.

Having a single large die (particularly the way AMD is doing things at 245mm2 on 28nm for Kaveri) is more expensive than two smaller dies (one for dGPU and one for CPU). This has to do with 1.) larger dies yielding worse than smaller dies and 2.) larger dies suffering from more edge of wafer loss.

And in practice we have been seeing having the graphics being separate from cpu also being a higher value on desktop.

Really I think the only place where it may make sense to pay the extra cost of integrating a large iGPU is low voltage mobile. This partly because space inside those mobile chassis comes at such a premium and having that extra integration means more space for battery, etc.

 

[monstercameron]

Having a single large die (particularly the way AMD is doing things at 245mm2 on 28nm for Kaveri) is more expensive than two smaller dies (one for dGPU and one for CPU). This has to do with 1.) larger dies yielding worse than smaller dies and 2.) larger dies suffering from more edge of wafer loss.

And in practice we have been seeing having the graphics being separate from cpu also being a higher value on desktop.

Really I think the only place where it may make sense to pay the extra cost of integrating a large iGPU is low voltage mobile. This partly because space inside those mobile chassis comes at such a premium and having that extra integration means more space for battery, etc.

I'm guessing that at this point, it is probably more economical for amd to build big die apus than not use enough wafers and pay glofo.

 

[ElFenix]

why do we get essentially the same AMD thread again and again and again?

 

[Ramses]

why do we get essentially the same AMD thread again and again and again?

Because we don't have the data to make or prove definitively a conclusion. And people like to argue or debate despite the fact. I learn some interesting pov's reading them and throwing my observations out there.

 

[monstercameron]

why do we get essentially the same AMD thread again and again and again?

why dont you guys make a mega thread/stickie thread for this ongoing discussion?

 

[Abwx]

why do we get essentially the same AMD thread again and again and again?

Must be a hit, if not a standard...

https://www.youtube.com/watch?v=7te1yOK0BNM

 

[III-V]

Having a single large die (particularly the way AMD is doing things at 245mm2 on 28nm for Kaveri) is more expensive than two smaller dies (one for dGPU and one for CPU). This has to do with 1.) larger dies yielding worse than smaller dies and 2.) larger dies suffering from more edge of wafer loss.

And in practice we have been seeing having the graphics being separate from cpu also being a higher value on desktop.

Really I think the only place where it may make sense to pay the extra cost of integrating a large iGPU is low voltage mobile. This partly because space inside those mobile chassis comes at such a premium and having that extra integration means more space for battery, etc.

A dGPU is not just a die, though. While it's the most expensive part of a graphics AIB, it's not the only thing, and thus you end up having more costs going the dGPU route.

Even AMD's cheapest dGPU will set you back $70. It's not really arguable -- APUs are more cost effective.

 

[Soulkeeper]

why dont you guys make a mega thread/stickie thread for this ongoing discussion?

I think the one on tomshardware has been a stickie since before bulldozer came out
354 pages

 

[cbn]

A dGPU is not just a die, though. While it's the most expensive part of a graphics AIB, it's not the only thing, and thus you end up having more costs going the dGPU route.

At least three things to consider:

1. The GPU die on the video card is typically clocked much higher than the iGPU of the APU. (So while other components are involved in the video card what silicon is used is better utilized)

2. Video card users are not obligated to use fast dual channel DDR3 (available only in 2 x4 GB) due to the dedicated VRAM. (In some cases the a dGPU user who only needs 4GB can save a good amount of costs because of this.)

3. Video card users should be able to achieve better stock cooler overclocks on their cpu. This because the processor does not have to share the thermal load with a iGPU. We also don't have to worry if the GPU on the video card is going to cause the cpu to downclock while both are under load.

Even AMD's cheapest dGPU will set you back $70. It's not really arguable -- APUs are more cost effective.

$70 will typically get a person a R7 250X (640 stream processors @ 1000 Mhz with 128 bit GDDR5) which is better than any AMD iGPU by a massive amount:

http://www.newegg.com/Product/Produc...bop=And&page=1

(Two R7 250Xs are currently listed at $69.99 AR)

In fact, I have seen R7 250Xs on sale for as low as $54.99 AR in the past, see this thread as an example --> http://forums.anandtech.com/showthread.php?t=2391710&highlight=

 

[Roland00Address]

Because AMD is so far behind in the foundry race and Intel is so far ahead. Numbers from Notebookcheck

Lets compare

28nm AMD
FX7600p 2.7 ghz base clock, 3.6 ghz turbo, 35w tdp, 245mm2 die size on 28nm
0.9 Average Cinebench 11.5 Single Thread, Only 1 Tested Laptop
2.5 Average Cinebench 11.5 Multi_ Thread, Only 1 Tested Laptop
02956 Average 3DMark (2013) - Fire Strike Standard Physics 1920x1080 , Only 1 Tested Laptop
27122 Average 3DMark (2013) - Ice Storm Standard Physics 1280x0720 , Only 1 Tested Laptop

22nm Intel
Core i7-4500u, 1.8 ghz base, can turbo to 2.7ghz dual core and 3.0 ghz single core, 15w tdp, 181mm2 die size on 22nm
1.3 Average Cinebench 11.5 Single Thread, min was 0.79, max 1.34 35 laptops/tablets tested
2.7 Average Cinebench 11.5 Multi_ Thread, min was 1.99, max 2.96, 36 laptop/tablets tested
03640 Average 3DMark (2013) - Fire Strike Standard Physics 1920x1080, min was 2155, max 4056, 29 laptops/tablets tested
30323 Average 3DMark (2013) - Ice Storm Standard Physics 1280x0720, min was 21459 , 31779 29 laptops/tablets tested

14nm Intel
Core M-5Y70, 1.1ghz base, can turbo to 2.6ghz but not for an extended period, 3.5w to 6w tdp depending on OEM decisions due to configurable tdp, 131mm2 die size for core m, 82mm die size for broadwell U @15w
1.1 Average Cinebench 11.5 Single Thread, Hp Envy x2-j001ng 1.89, Yoga 3 Pro 1.13
1.99 Average Cinebench 11.5 Multi_ Thread, Hp Envy x2-j001ng 1.89, Yoga 3 Pro 2.09
02631 Average 3DMark (2013) - Fire Strike Standard Physics 1920x1080 , HP Envy x2-j001ng 2602, Yoga 3 Pro 2660
22657 Average 3DMark (2013) - Ice Storm Standard Physics 1280x0720 , HP Envy x2-j001ng 20946, Yoga 3 Pro 24369

So as you can see Intel has a lower die size and a faster processor on chips with much lower tdp. At 14nm Intel is faster in single thread, and 80 to 90% the performance in multi thread and gpu scores with using 1/6 the TDP. At 22nm Intel wins in everything and if Intel chooses they can keep 22nm up and running with laptops and desktops to compete with AMD and dedicate 14nm only to try to capture market share vs ARM. Thus for the cheaper models intel if they want to can make a "strip down" chip with a much lower base clock they can easily do so just with using 22nm let alone 14nm.

Intels Pentiums and Intels i3s compete with AMD best chips and due to a smaller die size Intel can choose to undercut AMD with prices with OEMs. Thus Intel goal is not marketshare but instead profit maximization. Creating skus with various performances and prices that the OEMs are willing to pay ($50 dollars gets you x sku, $75 dollars gets x sku plus a few hundred mhz, $100 dollars get you a little better) and AMD has little room to compete with price for if intel chooses to they can undercut AMD but they do not for they make more money setting the prices and then AMD figuring out how low they can go while still trying to keep market share and make money.

 

[III-V]

At least three things to consider:

1. The GPU die on the video card is typically clocked much higher than the iGPU of the APU. (So while other components are involved in the video card what silicon is used is better utilized)

Clock speed is rather meaningless by itself.

2. Video card users are not obligated to use fast dual channel DDR3 due to the dedicated VRAM. (In some cases the a dGPU user can save a good amount of costs because of this.)

DDR3-2133 or DDR3-1866 is not that much more expensive than DDR3-1600 or DDR3 1333, and this is especially true for OEMs.

3. Video card users should be able to achieve better stock cooler overclocks on their cpu. This because the processor does not have to share the thermal load with a iGPU. We also don't have to worry if the GPU on the video card is going to cause the cpu to downclock while both are under load.

This doesn't apply to OEMs, where the money is made.

$70 will typically get a person a R7 250X (640 stream processors @ 1000 Mhz with 128 bit GDDR5) which is better than any AMD iGPU by a massive amount:

http://www.newegg.com/Product/Produc...bop=And&page=1

(Two R7 250Xs are currently listed at $69.99 AR)

In fact, I have seen R7 250Xs on sale for as low as $54.99 AR in the past, see this thread as an example --> http://forums.anandtech.com/showthread.php?t=2391710&highlight=

$70 is also what gets you an R7 240 -- the thing to keep in mind is that you pick up a considerable amount of fixed costs when going the AIB route, and you've got compounded margins included (retailer's margin, AIB vendor's margin, and component manufacturer's margin). The die itself is probably $15 or less.

Here's a breakdown of GPU AIB costs:

The idea with iGPUs is to get the cost down as far as possible, and dGPUs simply cannot compete.

 

[cbn]

Clock speed is rather meaningless by itself.

That is true, but these video cards (like the R7 250X I linked) with GDDR5 are not bottlenecked by lack of memory bandwidth (which is something we must always be concerned with iGPUs)

DDR3-2133 or DDR3-1866 is not that much more expensive than DDR3-1600 or DDR3 1333, and this is especially true for OEMs.

All the APU OEM boxes I have seen have DDR3 1600, not the faster DDR3 2133 or DDR3 2400.

Now on the DIY market these 2 x4 GB DDR3 2400 kits are running about $11 to $13 more expensive than a 2 x 4GB DDR3 1600 kit last time I checked. However, even with the fastest dual channel DDR3 2400 kit available, the AMD iGPU is still relatively memory bandwidth bottlenecked. So with dual channel DDR3 1600 (used by OEMs) the memory bandwidth situation only worsens.

P.S. I noticed you didn't see my edit back in post #135 (ie, having a dGPU means the user doesn't need to have a fast 2 x4 GB kit. One 4GB stick can work in many cases and is often specified in lower end gaming builds like these --> http://forums.anandtech.com/showthread.php?t=2389797)

 

[cbn]

$70 is also what gets you an R7 240 -- the thing to keep in mind is that you pick up a considerable amount of fixed costs when going the AIB route, and you've got compounded margins included (retailer's margin, AIB vendor's margin, and component manufacturer's margin). The die itself is probably $15 or less.

The price on R7 240s varies wildly, but in practice there is so much price compression between R7 250X and R7 240 there is really no point in even considering R7 240 unless it happens to be available in a super sale like this one --> http://forums.anandtech.com/showthread.php?t=2398683&highlight=

 

[Kuiva maa]

The APUs were and still are great for Laptops, yet there are only a hand full of AMD Kaveri APU laptops in Germany, USA but none here in Greece.

http://www.e-lenovo.gr/p.LENOVO-Z50-75-POWER-FX-7500-FHD-8GB-W8-1-BLACK.1057199.html

You can find this one. But I agree, one is as good as none. Generally speaking Greek laptop market has much less variety compared to 2-3 years ago. I often get (due to position) requests for affordable gaming capable laptops but every store seems saturated with ULV garbage. Not too long ago, 2 years max, you could get quad i7/GT555 in the 650-750 ballpark and nice quad llano or trinity with 7670M level of discrete graphics for 500 euros or even less. Now you need to pay nearly 1k for an equivalent i7 and AMD units are almost vanished.

Intel APUs are selling gangbusters. It's not the idea of integrated graphics which is the problem, it is AMD's specific offerings. There isn't a strong variety of different parts (just one giant die with parts fused), they have a weak sauce CPU architecture, and they are built on an outdated process node. Until they can fix those things they aren't going to be competitive.

They will never be competitive. These products ,at least on mobile are wooing a crowd that flat out cannot be won. This crowd either

a) Buys whatever can be found at stores.Even If AMD could somehow beat intel on perf/watt and be competitive in prices, they still wouldn't win enough contracts to make any meaningful impact in the market. They would push some extra units, perhaps even double their marketshare there. And that's the best case scenario. Intel owns the market.
b) Knows only intel and will ignore anything that doesn't have the blue intel inside sticker. Even if the competing product is readily available and superior. It just won't happen.

When it comes to x86, AMD can only hope to win the desktop enthusiast, that is savvy enough to know which product is faster and will act as an opinion leader. It is seems nigh on impossible but that's it.

 

[cbn]

3. Video card users should be able to achieve better stock cooler overclocks on their cpu. This because the processor does not have to share the thermal load with a iGPU. We also don't have to worry if the GPU on the video card is going to cause the cpu to downclock while both are under load.

This doesn't apply to OEMs, where the money is made.

I think DIY still matters, but even with overclocking out of the picture the cpu throttling under iGPU load is something to consider.

Another thing to consider for OEM is how much iGPU can be specified for a CPU socket rated at only 95 watts. There is only so much thermal load a stock type cooler (sharing the burdens of cpu and gpu together) can handle.

 

[ThatBuzzkiller]

The fact that they are bad products ?

 

[ElFenix]

The fact that they are bad products ?

142 posts, folks.

 

[Abwx]

At 14nm Intel is faster in single thread, and 80 to 90% the performance in multi thread and gpu scores with using 1/6 the TDP.

TDP specs of mobile Intel products are underestimated, there s some 11.5W HW that consume 18-19W at the SoC.


11.5W Haswell that consume actualy 23W at the SoC level, i guess that this is easier to improve perf/watt with fake specs :

http://www.notebookcheck.net/Toshiba-Satellite-W30t-A-101-Convertible-Review-Update.119960.0.html

But when AMD say that a chip is 15W it is actualy 15W or less, here an A8 6410 wich is 15W TDP :
http://www.notebookcheck.net/HP-Pavilion-13-a093na-x360-Convertible-Review-Update.130928.0.html

So much for the perf/Watt, i can do as well with whatever CPU, including a 100W, and a pen...

 

[VirtualLarry]

TDP specs of mobile Intel products are underestimated, there s some 11.5W HW that consume 18-19W at the SoC.


11.5W Haswell that consume actualy 23W at the SoC level, i guess that this is easier to improve perf/watt with fake specs :

http://www.notebookcheck.net/Toshiba-Satellite-W30t-A-101-Convertible-Review-Update.119960.0.html

But when AMD say that a chip is 15W it is actualy 15W or less, here an A8 6410 wich is 15W TDP :
http://www.notebookcheck.net/HP-Pavilion-13-a093na-x360-Convertible-Review-Update.130928.0.html

So much for the perf/Watt, i can do as well with whatever CPU, including a 100W, and a pen...

Are you honestly telling me that Intel is pulling the wool over their many OEMs' eyes, and that they don't know the specs of the products that they are designing?

Or just perhaps, TDP is not the same as total maximum power consumption.

 

[Abwx]

Are you honestly telling me that Intel is pulling the wool over their many OEMs' eyes, and that they don't know the specs of the products that they are designing?

Or just perhaps, TDP is not the same as total maximum power consumption.

Let s look at the numbers :

Idle : 4.5W load : 36.9W delta 32.4W

Losses on the PSU are about 10% , this reduce the delta to 29.16W at the DC input level, next step is to account for the laptop DC converters that have also about 90% efficency, the resulting delta is 26.244W, we can eventualy substract 2W to be cautious, 24W are left as the SoC comsumption, that is 108% more than the rated TDP.

That is the CPU is benched at 108% over its rated TDP, this allow to display 50% more perfs than if it was benched at 11.5W.

Honnestly.??.

 

[DrMrLordX]

Another thing to consider for OEM is how much iGPU can be specified for a CPU socket rated at only 95 watts. There is only so much thermal load a stock type cooler (sharing the burdens of cpu and gpu together) can handle.

Two things:

1). I haven't seen a 512-shader R7 iGPU in action, but the 384-shader iGPU produces very little heat compared to the 2 Steamroller modules. The sensor is wacky, but it's hard to actually get it to budge, even with stuff like Furmark or GPUPI.

2). No OEM is going to concern itself with "how much iGPU can be specified for a socket rated at only 95W". Never mind that Asus seems to have figured out how to push a lot more than 95W through FM2+, not that any OEM is concerned with that, either. It is not up to the OEM to decide how much iGPU can be specified for the socket. It is up to the OEM to decide how much they want to pay for parts, and how much they think they can charge a customer willing to buy an AMD machine. In the case of Kaveri chips, all the iGPUs fit well within the FM2+ socket limitations @ stock.

If anything, the low typical power draw of a Kaveri chip (which, mind you, almost never draws enough power to reach its rated TDP anyway) makes it stunningly easy to cool unless you're really trying to OC the hell out of the Steamroller modules. Seriously, these Kaveri chips can probably get by at around 1-1.1v vcore @ 3.4ghz or thereabouts. The 95w TDP rating of my 7700k is a cruel joke. The only way it approaches that is when the board sets the default vcore to something stupid like 1.3v. Because hey, turbo!!! Or something.

The fact that they are bad products ?

That's entirely subjective. Mine was cheap, and playing with it is fun. It replaced an older CPU and dGPU and manages to be faster on both counts while definitely using less power than that old power hog of a card I used to use (8800 GTX ACS).

 

[greatnoob]

Let s look at the numbers :

Idle : 4.5W load : 36.9W delta 32.4W

Losses on the PSU are about 10% , this reduce the delta to 29.16W at the DC input level, next step is to account for the laptop DC converters that have also about 90% efficency, the resulting delta is 26.244W, we can eventualy substract 2W to be cautious, 24W are left as the SoC comsumption, that is 108% more than the rated TDP.

That is the CPU is benched at 108% over its rated TDP, this allow to display 50% more perfs than if it was benched at 11.5W.

Honnestly.??.

Maybe that's the entire system power consumption and not the isolated chip itself.

 

[Abwx]

Maybe that's the entire system power consumption and not the isolated chip itself.

There s no mistake from my part, the 11.5W Haswell consume twice its rated TDP, the system doesnt use much more than the idle power comsumption plus the 2.2W i accounted when it is loaded.

 

[cbn]

2). No OEM is going to concern itself with "how much iGPU can be specified for a socket rated at only 95W". Never mind that Asus seems to have figured out how to push a lot more than 95W through FM2+, not that any OEM is concerned with that, either. It is not up to the OEM to decide how much iGPU can be specified for the socket. It is up to the OEM to decide how much they want to pay for parts, and how much they think they can charge a customer willing to buy an AMD machine. In the case of Kaveri chips, all the iGPUs fit well within the FM2+ socket limitations @ stock.

My point is that the socket's power rating is going to limit the maximum amount of iGPU performance we are going to see from these type of chips. So even if interposers and HBM come along to help with other problems I still see a limit to how much desktop APUs (in OEM boxes) will be able to replace dGPUs for gaming.