AMD what will happen in 2015?

[AtenRa]

With respect to the big core APUs, I can see that being an advantage for gaming laptops.

But for desktops, I think that extra die allocated for the very large iGPU on Kaveri is just too much of a cost adder.

For desktop, I would rather see AMD scale back their integrated graphics dramatically to help them better compete with Intel.

The iGPU is not only for gaming, they are used for GPGPU.

http://forums.anandtech.com/showthread.php?t=2394878&highlight=

 

[AtenRa]

graphics will be next.

We hear this for the last 3-4 years :whiste:

 

[cbn]

The iGPU is not only for gaming, they are used for GPGPU.

http://forums.anandtech.com/showthread.php?t=2394878&highlight=

I still think GPGPU is ahead of its time (and is adding way too much cost to desktop).

For now, they should confine big core APUs to laptops until the technology matures.

P.S. If someone wants to wants use GPGPU on desktop, they can just add a graphics card.

 

[AtenRa]

I still think GPGPU is ahead of its time (and is adding way too much cost to desktop).

For now, they should confine big core APUs to laptops until the technology matures.

P.S. If someone wants to wants use GPGPU on desktop, they can just add a graphics card.

Adding a dGPU adds cost and higher power consumption.

Also large iGPUs helps with lower volume products.

 

[cbn]

Adding a dGPU adds cost and higher power consumption.

It adds cost, but I have never seen an A10 APU cost comparison to CPU + dGPU that was favorable.

In almost all cases I have seen in the past having cpu and gpu separate is higher performance at lower cost.

P.S. For power sensitive form factors like laptop, that is where the APU should be used.

 

[witeken]

We hear this for the last 3-4 years :whiste:

Irrelevant. Skylake will have 20 or 80% more EUs and 2 iterations more of architectural improvements.

 

[cbn]

Also large iGPUs helps with lower volume products.

In cases of niche very small form factor desktops, a Mobile SKU Kaveri could be used (possibly with configurable TDP).

 

[AtenRa]

It adds cost, but I have never seen an A10 APU cost comparison to CPU + dGPU that was favorable.

In almost all cases I have seen in the past having cpu and gpu separate is higher performance at lower cost.

P.S. For power sensitive form factors like laptop, that is where the APU should be used.

Prices from Amazon

A10-7850K = $153.00


Athlon 860K = $90.00

R7 250 = from $87.00

total = From $177.00

A10-7850K has 512 Steam Processors, R7 250 has 384.
So for GPGPU the A10-7850K is cheaper with lower power consumption AND Faster than Athlon + R7 250.

 

[AtenRa]

Irrelevant. Skylake will have 20 or 80% more EUs and 2 iterations more of architectural improvements.

Skylake will have to go against Carrizo.

 

[AtenRa]

In cases of niche very small form factor desktops, a Mobile SKU Kaveri could be used (possibly with configurable TDP).

I was talking about manufacturing.

 

[Phynaz]

Yeap and that is the AMD iGPU superiority.

Which has gotten AMD where?

Features have to have a benefit to the consumer in order to be marketable. Consumers haven't seen seen a benefit in this feature obviously.

 

[Arachnotronic]

Skylake will have to go against Carrizo.

While I don't know what the quality of Intel's Gen. 9 architecture will ultimately be like, fighting the equivalent of a "tock" on the GPU side + 14nm with a 28nm part -- no matter how good that 28nm part's architecture is -- will be very tough, particularly in the market for mobile (i.e. laptop) computing where power efficiency is king.

 

[mrmt]

Which has gotten AMD where?

Features have to have a benefit to the consumer in order to be marketable. Consumers haven't seen seen a benefit in this feature obviously.

This. I think it is already clear after three generation of APU products that the concept is a commercial failure. AMD lost share and volume in every single market APUs were marketed. Since the company can't bleed cash forever, it is already on critical levels in terms of liquidity, let's see what kind of changes Lisa will bring to this area.

 

[AtenRa]

Which has gotten AMD where?

Have 100% of Console market share and cheap 28nm 100mm2 SoC have the same performance at lower power consumption than 22nm FF big core Haswell at 15W TDP.

 

[AtenRa]

While I don't know what the quality of Intel's Gen. 9 architecture will ultimately be like, fighting the equivalent of a "tock" on the GPU side + 14nm with a 28nm part -- no matter how good that 28nm part's architecture is -- will be very tough, particularly in the market for mobile (i.e. laptop) computing where power efficiency is king.

Carrizo will be 28nm HDL, not the same plain 28nm Kaveri used.

 

[monstercameron]

This. I think it is already clear after three generation of APU products that the concept is a commercial failure. AMD lost share and volume in every single market APUs were marketed. Since the company can't bleed cash forever, it is already on critical levels in terms of liquidity, let's see what kind of changes Lisa will bring to this area.

so hypocritical, how can the concept be a failure when intel has effectively been doing the same thing? AMD losing market share with their current product stack seems coincidental, there is only so much amd can do with intel dominating the market.

 

[witeken]

cheap 28nm 100mm2 SoC have the same performance at lower power consumption than 22nm FF big core Haswell at 15W TDP.

Should not confuse imagination with reality.

 

[mrmt]

so hypocritical, how can the concept be a failure when intel has effectively been doing the same thing? AMD losing market share with their current product stack seems coincidental, there is only so much amd can do with intel dominating the market.

Intel was far more parsimonious with iGPU area than AMD. While AMD went straight for 50% area and power hogs, Intel was slowly ramping up GPU and didn't lose focus from power consuption improvements. The two products are differents, Intel sells nice CPU chips with good enough iGPUs, while AMD sells poor CPU with a good but bandwidth choked iGPU.

As a result. Intel made and still makes tons of money with its strategy while AMD doesn't. It's not hypocrisy, really, just judging the numbers.

 

[AtenRa]

Should not confuse imagination with reality.

HP Pavilion 13,3" with A8-6410 28nm SoC

HP Pavilion 13,3" with Core i3 4030U 22nm FF Haswell

 

[cbn]

Prices from Amazon

A10-7850K = $153.00


Athlon 860K = $90.00

R7 250 = from $87.00

total = From $177.00

A10-7850K has 512 Steam Processors, R7 250 has 384.
So for GPGPU the A10-7850K is cheaper with lower power consumption AND Faster than Athlon + R7 250.

Current Newegg prices for R7 250 start at $69.99 After $10 rebate:

http://www.newegg.com/Product/Produc...64&ignorebbr=1

EDIT: I just noticed this was a GDDR5 version. (This changes the analysis I just wrote below to be in favor of the Athlon x4 860K plus R7 250 even more.)

So going by those prices buying the APU saves a person only $7, but then there is the bandwidth issue. (That R7 250 comes with 128 bit DDR3 1800, and it doesn't have to share any of it with the cpu).

So for an OEM Desktop situation where the DDR3 speed doesn't exceed dual channel DDR3 1600, the Athlon x4 860K + R7 250 wins on performance at a cost of $7 more money going by these prices

Now for a DIY gamer build, there is the option for the A10-7850K to use faster DDR3 2400 RAM, but these only exist in 2 x 4 GB configuration (and I am not even sure if it would make the APU faster since the bandwidth also needs to be shared with the cpu).

Meanwhile a budget gamer with the Athlon could get by with a single 4GB stick.

So in a budget gamer build situation, the Athlon x4 860k user only pays $31.99 free shipping for a single 4 GB stick --> http://www.newegg.com/Product/Produc...eId=1&name=4GB

But could use a 2 x 4GB DDR3 1600 kit as well @ $59.99 --> http://www.newegg.com/Product/Produc...35&ignorebbr=1

Whereas the A10-7850K user is forced to use a 2 x 4GB DDR3 2400 kit @ $73.99 plus .99 shipping:

http://www.newegg.com/Product/Produc...069&IsNodeId=1

So going by DIY build prices, the Athlon x4 860K + R7 250 ends up being $36 cheaper if a single 4GB stick is used and $8 cheaper if using 2 x 4GB. (And we still don't know if the A10-7850K would be faster even with the speedier 2 x4GB DDR3 2400 RAM).

 

[monstercameron]

Intel was far more parsimonious with iGPU area than AMD. While AMD went straight for 50% area and power hogs, Intel was slowly ramping up GPU and didn't lose focus from power consuption improvements. The two products are differents, Intel sells nice CPU chips with good enough iGPUs, while AMD sells poor CPU with a good but bandwidth choked iGPU.

As a result. Intel made and still makes tons of money with its strategy while AMD doesn't. It's not hypocrisy, really, just judging the numbers.

so intel make good cpus and good enough gpus but amd makes "AMD sells poor CPU with a good but bandwidth choked iGPU." that is beyond skewed and is pure hypocrisy.

 

[AtenRa]

snip...

We are talking about GPGPU specifically and you are bringing gaming in to the discussion ??

 

[Erenhardt]

nip pointless OT

Yep. I guess amd would love you to pay for two cut down unfunctional chips that would otherwise go to the trash.

If you are debating apu on perf/$ merits, why going for the top of the line compute oriented part? The a8 provides 90% performance of a10 in games for a fraction of price.

And if you tried to do kaveri to carrizo transition then its poor troll attempt.

Those two are designed for different markets. Whole other design concept.

 

[cbn]

We are talking about GPGPU specifically and you are bringing gaming in to the discussion ??

The vast majority of the time the use case is gaming though.

GPGPU is just too niche to spend all that die area for desktop iGPU at this time.

 

[Headfoot]

The problem is there are three categories of GPU use in general desktop/laptop Windows computing, and a faster GPU only helps a small minority of those.

1) No GPU needed loads, 2) Nearly any GPU past 2009 will do minimal-loads (desktop animations, browser GPU offloading for page compositing, etc.), 3) Faster GPU = Faster Task completion loads

A faster GPU in the APU is only relevant for category three. There aren't many tasks that fall into category three, namely gaming, video encoding/decoding and acceleration (and these can be moved into fixed function silicon). This leaves gaming as the main task that is noticeably accelerated by a faster chip (and only by a faster chip). And dGPUs are so much faster than the iGPU that anyone with gaming as a primary purpose won't choose an iGPU. The HSA play seeks to increase the number of things that fall into category 3 but it has so far produced very few results. Further, the portion of things that can't be accelerated by dGPUs and only by iGPUs is so tiny as to be irrelevant. The cap on max performance that the iGPU scheme creates is self-limiting.

iGPUs need to accelerate more things noticeably or be closer to dGPUs in speed in order to be a viable choice to anything more than a small portion of the market. Once APUs get HBM we might have a fight but right now current APUs are just steps on the path to hopefully a product which is good enough to grab real market share. Right now the only market it makes sense in is low-cost laptops that people might want to run a game on every once and a while. And even then, the low end APU SKUs are slower than the top end ones to the point of not being capable enough for gaming anyways, especially paired with low end RAM that you see in low cost laptops

cbn is right -- every choice comes at the opportunity cost of some other choice. Spending generations getting APUs to where they provide tangible benefits to a tangible portion of the market is R&D budget, software budget, development time and actual silicon chip transistor budget that could have been spent on other goals (like entering the Phi/CUDA space more aggressively, or lower power CPUs, or more aggressive ARM adoption, or any other number of things).

2015 will only be exciting for AMD if AMD can step up their software game to leverage their already-paid-for hardware improvements in HSA and GPGPU. Given that Mantle is a pretty nice piece of software, more forward/lateral thinking software than they've ever done before, I think its possible if not unlikely.