Why have AMD APUs failed on the market?

[cbn]

I think the market have already spoken and decided. And it wasnt in favour of hexcores.

When I counted OEM boxes at Newegg in this post, the FX boxes actually did pretty good (44 listings vs. 58 listings for FM1/FM2/FM2+).

And this despite all of the AM3+ listings I saw needing a discrete card.

But I am not talking about a hexcore that needs a discrete card.

 

[SPBHM]

it looks like the past 2 years were all about sub 20W, and AMD focused on 35W with Kaveri?

from what I noticed the decent APUs have poor availability (Kaveri and Richland) and are inside to many overpriced and badly designed laptops, and most APUs are just super slow Jaguar/Bobcat...

as for desktops, gaming APUs are hard to sell, since discrete graphics are way faster, and Intel IGPs are fast enough for others

AMD always seems to be at least 2 years late with their releases

 

[MisterLilBig]

If a market leader is already developing a similar technology on open standards, and AMD has access to that market leader patents, why bother with developing their own market standard? What are the chances of AMD making something mainstream? For a company as cash strapped as AMD is, it is better to just follow Intel lead down this rabbit hole and try to come with a more efficient implementation than to try to develop a new standard and then hope software companies will come.

HSA smells of another of AMD's fool's errands.

The HSA rabbit hole was founded with Samsung, Qualcomm, ARM, Mediatek, Imagination and Texas Instruments. Has 12 supporters, 23 contributors and 11 academic institutions behind it.

HSA is not just AMD.
( ht(t)p://www.hsafoundation.com/ )

 

[camelNotation]

I think a big reason they haven't done well is that Intel's graphics performance has been improving so dramatically. Here's what I recall about the Intel/AMD matchups over the last few years:

2011: Sandy Bridge vs. Llano. Even though Llano's GPU was about twice as fast as the HD 3000, the first year that AMD sold APUs, Intel had a version of their own with a competent IGP.

2012: Ivy Bridge vs. Trinity. Trinity's IGP wasn't that much faster than Llano's due to die size constraints. However, Intel shrunk from 32nm to 22nm and could afford a much larger GPU die. Since HD 4000 was 40-50% faster than HD 3000, Intel began to close the gap.

2013: Haswell vs. Richland. AMD couldn't improve much since Richland was a respin of Trinity. Intel improved IGP performance by 20-30% over the HD 4000 with the HD 4400 and released the Iris/Iris Pro graphics, which were just as fast as, or faster than, Richland.

2014: Haswell vs. Kaveri. AMD recouped lost ground, but only a little bit. Kaveri and Iris Pro still trade blows. Mobile Kaveri launch was delayed until mid-year, so very few Kaveri laptops are on the market.

2015: Broadwell/Skylake vs. Kaveri/Carrizo. Broadwell's HD 6000 IGP is scary large.

​

AMD had a great idea in including a competent IGP on their chips, but Intel had the same idea, and has been improving at a faster rate despite their lower performance starting point. Essentially, Intel has been out-AMDing AMD.

That being said, I am looking forward to seeing what Carrizo has in store. AMD will finally have a one chip solution to compete with Intel in the low power notebook market.

 

[Fjodor2001]

I don't really get the moar cores bandwagon. We all know that we have to trade core frequency and power consumption for more cores,

The problem is we're not getting much higher IPC of frequency from one CPU generation to the next these days.

Adding extra cores is a way to provide substantial additional performance in the face of that fact. Sure, you won't be getting 100% more performance when going from 4 to 8 cores in most scenarios. But for the scenarios where you actually need much performance (video encoding, gaming etc), more cores can be made use. So you can get perhaps 50% performance increase in those use cases going from 4 to 8 cores. That is enough to justify it. It corresponds to 8-9 CPU generations, where we're getting ~5% additional single threaded performance per generation these days (1.05^8=1.48).

Especially when we're getting ever more transistors to play with, due to node shrinks.

 

[mrmt]

The HSA rabbit hole was founded with Samsung, Qualcomm, ARM, Mediatek, Imagination and Texas Instruments. Has 12 supporters, 23 contributors and 11 academic institutions behind it.

HSA is not just AMD.
( ht(t)p://www.hsafoundation.com/ )

How many are developing HSA hardware beyond AMD? That's right, none.

 

[ShintaiDK]

When I counted OEM boxes at Newegg in this post, the FX boxes actually did pretty good (44 listings vs. 58 listings for FM1/FM2/FM2+).

And this despite all of the AM3+ listings I saw needing a discrete card.

But I am not talking about a hexcore that needs a discrete card.

If you open the link again its what, 38 FX vs 65 FM.

Looking further on the FX parts:
http://www.newegg.com/Product/Produc...lt=True&page=1

Try remove all the open boxes and refurbs. And you are down to 25 FX. Thats right, 35% of the entire listing was open box/refurb.

But I think you are missing the point, if you look on the thread title again. How many listings does Intel have? And didnt you forget the biggest segment? One that sells 0 non APUs:
http://www.newegg.com/All-Laptops-No...Category/ID-32

So still removing open boxes and refurb. It ends up like what? 235 APUs vs 25 FX?

 

[monstercameron]

I though this was a discussion about why d apus weren't doing so hot and not a troll thread. Intel and amd are doing the same thing more or less yet Intel is profitable. Performance levels are comparable to the layman consumer...so why is amd failing?
Process node disadvantage? No, consumers don't care.
Die size/"cost structure "? Nope, consumers don't care
Product win? Hell yes, how is amd to make money with only a handful of low margin craptops...amd needs more design wins and industry partnerships.

 

[VirtualLarry]

I though this was a discussion about why d apus weren't doing so hot
Process node disadvantage? No, consumers don't care.

They do, indirectly, when they can't get the performance or comparable battery life of their competitor.

 

[MisterLilBig]

How many are developing HSA hardware beyond AMD? That's right, none.

That we know of? None. Would be foolish to think that only AMD is tho. Considering all those hardware founders.

 

[mrmt]

Process node disadvantage? No, consumers don't care.
Die size/"cost structure "? Nope, consumers don't care
Product win? Hell yes, how is amd to make money with only a handful of low margin craptops...amd needs more design wins and industry partnerships.

Why would an OEM want to use an AMD APU if they must either go for a more expensive cooling solution and pay higher prices for AMD APU?

You know that the price structure for OEMs is not what you see in the consumer market, right? And the root cause here is the big die of AMD APUs. This is why we see 0 high volume AMD chromebooks while there are plenty of Intel ones. Intel can bite this bullet and go down in terms of price, and still make a profit, but AMD cannot.

OEMs are shunning AMD offers because there's no money to be made, Core checks Bulldozer APUs, Bay Trail checks cat APUs.

 

[mrmt]

That we know of? None. Would be foolish to think that only AMD is tho. Considering all those hardware founders.

Why it would be foolish? Could these companies, which are far more capable and better run than AMD, reach the conclusion that HSA is a foolish venture and that they will just stay member of the foundation to have access to AMD IP but not commit to develop actual hardware?

 

[NTMBK]

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.

 

[VirtualLarry]

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.

How's Via doing these days, btw? Still using FSB archtecture, and a DX9.0c IGP in the chipset?

 

[Blue_Max]

...maybe AMD would be wise to make an updated version of their chipset to allow a PAIR of 7850k (or newer) APU's on a single board? 8 threads and double the current graphics?

...but that's kinda' pointless as well. The cost would make an i3+graphics more appealing.

They've got such potential there... they'd better make use of it before Intel + Iris Pro close the gap for good.

 

[NTMBK]

...maybe AMD would be wise to make an updated version of their chipset to allow a PAIR of 7850k (or newer) APU's on a single board? 8 threads and double the current graphics?

...but that's kinda' pointless as well. The cost would make an i3+graphics more appealing.

They've got such potential there... they'd better make use of it before Intel + Iris Pro close the gap for good.

Kaveri has no cache coherent off-die interconnect. You can't make it dual socket.

 

[III-V]

For desktop, I think large iGPUs are pretty much a bad idea for both AMD and Intel at this time.

You think reduced BoM and increased perf/watt is a bad thing? Why? That's all APUs really boil down to.

 

[frozentundra123456]

You think reduced BoM and increased perf/watt is a bad thing? Why? That's all APUs really boil down to.

igpus are definitely a good thing. I dont think many would argue with that. Now whether devotion of more than 50% of the die area to it like AMD does, especially with limited bandwidth, is a worthwhile trade off is another question. And the lack of HSA apps makes it even more questionable.

Perhaps it is the right thing, because even intel is devoting ever more die area to gpu with each generation, although I am not sure this is the right path either. Almost seem like they are trying to outdo AMD at their own game, but I really think Intel's emphasis on gpu is directed toward Apple and in competition with ARM.


Edit: I know we are going to get the usual arguments that we dont need it, but I would much, much rather see intel using 14nm to provide a mainstream hex core cpu rather that Broadwell K with a honking GT3 and e-dram. I mean, come on, we have had quad core since 2006 on what 65nm?

 

[DrMrLordX]

...maybe AMD would be wise to make an updated version of their chipset to allow a PAIR of 7850k (or newer) APU's on a single board? 8 threads and double the current graphics?

...but that's kinda' pointless as well. The cost would make an i3+graphics more appealing.

They've got such potential there... they'd better make use of it before Intel + Iris Pro close the gap for good.

AMD would be insane to re-engineer Kaveri (or Godavari) to work in a MP configuration. It has no L3, and L3 has been important for MP AMD configurations since the introduction of HT Assist, at the very latest. Better to imagine a Beowulf cluster of those . . .

Besides, the last enthusiast MP setup they had was a disaster.

 

[Ramses]

Intel APUs are selling gangbusters.

I'm sure that's true, however it's akin to saying a Lada, the most popular car in Russia that is selling quite well (relatively speaking) is a good car because it's selling well. It's selling well because there isn't really an alternative. Other than AMD in this case who apparently can't sell there way out of a wet paper bag. That sort of logic is on the surface sound but it's a bit of a trap.

 

[Blue_Max]

Kaveri has no cache coherent off-die interconnect. You can't make it dual socket.

It was a wild, dumb idea. The cost alone makes it pointless unless they can squeeze two current 7600's onto a single chip.

 

[cbn]

When I counted OEM boxes at Newegg in this post, the FX boxes actually did pretty good (44 listings vs. 58 listings for FM1/FM2/FM2+).

And this despite all of the AM3+ listings I saw needing a discrete card.

But I am not talking about a hexcore that needs a discrete card.

If you open the link again its what, 38 FX vs 65 FM.

Looking further on the FX parts:
http://www.newegg.com/Product/Produc...lt=True&page=1

Try remove all the open boxes and refurbs. And you are down to 25 FX. Thats right, 35% of the entire listing was open box/refurb.

(snip)

So still removing open boxes and refurb. It ends up like what? 235 APUs vs 25 FX?

Removing refurb and open boxes on both the FX line and big core APUs here is what I got:

12 A10
9 A8 (removed cat core based a8-6410 listings)
3 A6 (removed cat core based a6-5200 and a6-6310 listings)
4 A4 (removed cat core based a4-5000 and a4-6250 listings)

21 FX

So a mix of 28 big core APU listings vs. 21 FX listings.

That is not a big discrepancy at all, and remember for whatever reason those FX boxes did not have iGPU.

 

[Shivansps]

APU fails because people either buy a AM3+ or a dGPU or an Intel I3, I5 or Celeron/Pentium.
Low end builds are now using AM1, that get beaten by a 1150 Celeron in CPU/IGP because they dont have DC.
Also people tent to get powerfull Intel dual + gdpus as a cheaper option too, G3220+dgpus like a 250 or a 750TI.

Not a lot of people want to get an APU just to play with its igp, with the addons of having to get expensive memory. ALso the lack and price of ITX FM2 mbs is a problem too.

in out shop (outside US), the best selling combo is a AMD 2650+Asrock AM1, just because is the cheaper option, 2nd best is Intel G3220 with or whiout dGPU. People who want a powerfull pc goes to a FX or an I5, FM2 have literally no demand since AM1 arrived, before that it was common to sell FM2 and A4-4000/5300 as the cheaper option.

How's Via doing these days, btw? Still using FSB archtecture, and a DX9.0c IGP in the chipset?

They are selling these things cheaper than a R-PI... http://www.newegg.com/Product/Produc...82E16813135322

Judging by this http://www.viaembedded.com/en/products/boards/1893/1/EPIA-M920.html they have move forward a bit...

 

[cbn]

For desktop, I think large iGPUs are pretty much a bad idea for both AMD and Intel at this time.

You think reduced BoM and increased perf/watt is a bad thing? Why? That's all APUs really boil down to.

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.

 

[erunion]

First rule of engagement : You cannot sell your products if you dont have any products in the market.

IDC pointed out awhile back that if AMD thinks its products are worth buying but can't get OEMs to adopt them, they need to sell products themselves. Nvidia did this with tegra.

Maybe AMD believes in their products less than Nvidia does?