Future APU's and the end of FX

Michaelius said:
So the moment when AMD manages to secure future games will have good multithreaded support due to putting 8 cores in consoles they stop updating 8 core cpu lines.

No wonder the company is joke.
Click to expand...

I got an FM2 based mITX board (no AM3+ boards of that type worth buying) hoping I would see an eventual Athlon (no GPU) 8 Core type chip.

Yeah. Sigh.
 
atom said:
I guess when it comes down to it AMD's FX series isn't that profitable for them. APU is quite popular and I have caught myself looking at FM2+ stuff which usually ends with me slapping myself back into reality. If the APU stuff that's coming out now was out in June when I built my PC I could have saved some good money. I bought an FX 8350 and an HD 7790. That's $350 bucks right there. I am hoping Mantle is as good as the wizards who praise it's magic say.

All that said, this graphic could be completely false.
Click to expand...

Wizards :) It comes down to the way graphics are handled. The dry version is the bottleneck debate which is boring boring boring boring.

In the video/amd section of the forum has a ask devs about Mantle thread (youtube vid of apu`13)and some good feedback there from the devs.

It goes much further then the outline of having just an API replace another API.
Mantle already allows features that in the past were unable to get driver support because in theory it was possible but the driver was not allowing it.

You program the chip almost directly the methods that developers had to use are no longer bound to the "golden rules". In the past this would lead to copying what was done before because it works now that you can do all of it on the hardware itself means that drivers are no longer a limiting factor.

You can even move an awful amount of compute directly to the chip without going through the cpu.
 
Last edited:
reb00tin said:
I still don't get it. Are we going to get Steamroller for AM3+ ?
Click to expand...

I haven't seen ANYTHING that says we will. Everything that has been released from AMD shows that the only thing we're gonna have for AM3+ (and performance desktop overall) is going to be Piledriver til 2015....which makes absolutely no sense at all which is why I tend to think they do have "something" in the works for release some time in 2014. May not be Steamroller but maybe a Piledriver refresh kinda like C3 stepping for Phenom II back in 2010. But at this point, it's all a guess cause as far as I know, there is jack shit out there saying anything about AM3+.
 
reb00tin said:
Would Steamroller be compatible with 990FX? What would limit it? The old north bridge? Memory bandwidth?
Click to expand...

the biggest limitation from AMD's PoV would be marketing, it would be a tough sell:

"come one, come all. roll up to buy our new premium-brand processors. they use a backward ass platform with none of the latest features a consumer in 2014 will expect, but yes, it is premium really!"

USB 3.1 = no
SATA Express = no
PCIe 3.0 = no
Fast interconnect = no
On chip PCIe = no

AM3+ is dead as a doornail.
 
R3MF said:
the biggest limitation from AMD's PoV would be marketing, it would be a tough sell:

"come one, come all. roll up to buy our new premium-brand processors. they use a backward ass platform with none of the latest features a consumer in 2014 will expect, but yes, it is premium really!"

USB 3.1 = no
SATA Express = no
PCIe 3.0 = no
Fast interconnect = no
On chip PCIe = no

AM3+ is dead as a doornail.
Click to expand...

OTOH who cares, as long as it benches good?

USB3.0 is fine. Sata is blazing fast as is, PCIe 3.0 offers ZERO real world benefit, even to the fastest GPU's. Everything else is meh unless it shows *actual* gains.
 
omega-x said:
OTOH who cares, as long as it benches good?

1. USB3.0 is fine.

2. Sata is blazing fast as is

3. PCIe 3.0 offers ZERO real world benefit, even to the fastest GPU's.

Everything else is meh unless it shows *actual* gains.
Click to expand...

You are referring to a future technology...in 2014-15 the hardware/peripheral scene will be a bit different from today.

1. I agree...for now. USB 3.1, while a nice incremental upgrade, is almost pointless imo since there are hardly any dedicated 3.0 devices on the market other than flash drives.

2. I agree. However, if the interconnect popularity is going to change to SATAe, then SATA devices may become scarce/expensive.

3. PCIe 3.0 gives what, about a 5% gain? (but often only to the highest performing GPUs available right now). I agree that's nothing major or perhaps not even noticeable to most users in most occassions, but there will be new GPU generations released in 2014-2015 that will likely utilize the extra bandwidth of PCIe 3.0 much better than what we have available today.

Points being, if someone where to come on this forum asking for gaming PC build advice sometime between mid-2014 and mid-2015, do you honestly think the most recommended parts will be new Steamroller processors paired with tired motherboards featuring grossly outdated technology because "it benches good"? I don't see that happening, because [H] strives to lay out the facts about real-world gaming/performance results for any and all products they review, not list out results from a bunch of bullshit benchmarks.

I'm rooting for AMD to do the right thing and incorporate all the latest and greatest techs in a very soon to be released chipset. If they don't, then that's on them and they will likely falter some more.
 
On a side note all this means is everyone at the design and engineering group responsible for the next generation LGA 2011 successor and LGA 1150 successor (well at least the desktop end) can relax and take a semi-long vacation. Without the FX series, those listed above can really not do much then relabel their broadwell products to skylake and release it for the exact same price point.

AMD needs to all do something with FX im suggesting to drive Intel to innovate
 
cageymaru said:
You have to think of the GPU on the APU as another smaller processor.

In the future I could see Windows allocating background processes to it. Or maybe you set a program to run on the GPU. A high powered APU with an equally powered GPU on the die unlocks lots of possibilities.
Click to expand...


Possibilities, in the future, sure. Maybe. Eventually...

Until then, anyone who buys a discrete graphics card would arguably be better served by a chip that cuts GPU resources by ~50% or so (leave 'enough' to do any APU/HSA/gobbldygook stuff that may come along) and uses the liberated die space to add another 'Roller module, for a total of 6 cores.
 
Geonerd said:
Possibilities, in the future, sure. Maybe. Eventually...

Until then, anyone who buys a discrete graphics card would arguably be better served by a chip that cuts GPU resources by ~50% or so (leave 'enough' to do any APU/HSA/gobbldygook stuff that may come along) and uses the liberated die space to add another 'Roller module, for a total of 6 cores.
Click to expand...

Do you engineer AMD chips?
 
R3MF said:
the biggest limitation from AMD's PoV would be marketing, it would be a tough sell:

"come one, come all. roll up to buy our new premium-brand processors. they use a backward ass platform with none of the latest features a consumer in 2014 will expect, but yes, it is premium really!"

USB 3.1 = no
SATA Express = no
PCIe 3.0 = no
Fast interconnect = no
On chip PCIe = no

AM3+ is dead as a doornail.
Click to expand...

What marketing :) ? Long long time ago they decided to go APU only (2011). In general that is where AMD can get the most out of the combined strength.
I would not even list the features on the AM3+ chipset.

My problem with AMD is that it has been years when there is only Piledriver , while FM2 already had a refesh in between (Richland) and nothing been mentioned about what AM3+ is going to get.

To point to late 2014/2015 and expect the people with a AM3+ mainboard to keep using it or shops keep stock is rather silly. Why not tell people: no more new cpu for AM3+ . When we saw the 9xxx series that should have meant something to most people.
 
piledriver might be the end of am3+, but you have to wonder if they have AM4 in the works with steamroller FX based chips since DDR4 is coming out any day now.

maybe AM4 isn't on the roadmap because they havn't announced AM4 yet?
 
Wildace said:
piledriver might be the end of am3+, but you have to wonder if they have AM4 in the works with steamroller FX based chips since DDR4 is coming out any day now.

maybe AM4 isn't on the roadmap because they havn't announced AM4 yet?
Click to expand...

AMD explicitly mentioned they intended to merge AM3+ and FM2 into a single socket from then on. In other words, no more separate socket lines.

I wonder if FM2+ is it.
 
Geonerd said:
Possibilities, in the future, sure. Maybe. Eventually...

Until then, anyone who buys a discrete graphics card would arguably be better served by a chip that cuts GPU resources by ~50% or so (leave 'enough' to do any APU/HSA/gobbldygook stuff that may come along) and uses the liberated die space to add another 'Roller module, for a total of 6 cores.
Click to expand...

On an Intel 3770K you can render video in real time on the built in GPU, which frees up the CPU for doing other tasks. People are doing many other things on them also. So the notion of that the GPU on an APU doing tasks even though there are much more powerful video cards installed in a system is old news. What is new news is it becoming mainstream in the future. That's why AMD is developing Mantle and other technology to push their technology to the front. One of the main selling points of Mantle is freeing up resources for the CPU to allow it to do more things.

The really good thing about tech such as QuickSync by Intel and Mantle by AMD is that since they are using the integrated GPU that is just sitting there gathering dust, it costs very little resources to call upon it. You just need to have your program or OS call upon it for tasks. It's like being broke and then finding money under the mattress! If you want to know more about tech like this in use today I would suggest following open source projects. Those programmers like writing code for projects like Handbrake to render video on your video cards instead of your CPU. Cutting edge things that the big tech giants take notice of, but are stuck with tunnel vision.

General QuickSync information straight from Intel.
http://www.intel.com/content/www/us...uick-sync-video/quick-sync-video-general.html

Anandtech looks at QuickSync.
http://www.anandtech.com/show/6864/handbrake-to-get-quicksync-support

Adobe Pro/ Premiere uses QuickSync.
http://software.intel.com/en-us/blo...-plug-in-for-adobe-pro-and-consumer-products/

Stream your games on Twitch.tv using QuickSync to transcode video in real time to free up resources for your CPU.
http://obsproject.com/forum/viewtopic.php?f=20&t=7580

The whole point is that software already exists that uses the GPU on Intel chips for quite a few years. AMD wants to expand upon this. It's new to many because they simply didn't know that you could do these type of operations on the integrated GPU. Since these things run OpenCL you can program them to do other things and that's where Mantle integration comes into play. This is my speculation, but maybe BF4 allows Kaveri users to process destruction calls with the on die GPU? What kind of resources could that free up for the CPU to allow higher frame rates? That's why I see the on die GPU as a really powerful processor waiting to be tapped.

I do agree with you that I'd rather see 8 or 12 Steamroller cores in a package, but as powerful as AMD video cards are in regards to OpenCL performance, I could see them getting very close to an 6 core processor in general performance. My problem is that I already own an 8 core so I'm skeptical as to how much longer will I have to wait to have something to upgrade to.
 
cageymaru said:
In the future I could see Windows allocating background processes to it. Or maybe you set a program to run on the GPU. A high powered APU with an equally powered GPU on the die unlocks lots of possibilities.
Click to expand...

This won't happen.

The GPU portion of the chip can not run native x86 or x64 code, and even if it could it would be absolutely terrible at it. It would require specialized code, and even then only some types of tasks would lend themselves well to the massively parallelized capabilities of GPU processing.

On the other hand, I could see some code using the GPU portion as another extended instruction set like MMX or SSE and run certain portions of the code on it, but this would require a very smart compiler or painstaking manual coding, and specialized binaries just for AMD HSA enabled APU'S. A compiler like this could possibly even dump some code to the ARM core.

There will be no universal running the program or thread of your choice on the GPU or ARM core. Simply can not happen without specialized ported binaries.
 
Zarathustra[H];1040435867 said:
This won't happen.

The GPU portion of the chip can not run native x86 or x64 code, and even if it could it would be absolutely terrible at it. It would require specialized code, and even then only some types of tasks would lend themselves well to the massively parallelized capabilities of GPU processing.

On the other hand, I could see some code using the GPU portion as another extended instruction set like MMX or SSE and run certain portions of the code on it, but this would require a very smart compiler or painstaking manual coding, and specialized binaries just for AMD HSA enabled APU'S. A compiler like this could possibly even dump some code to the ARM core.

There will be no universal running the program or thread of your choice on the GPU or ARM core. Simply can not happen without specialized ported binaries.
Click to expand...


APUs are the future, even Intel sees this, CPU only chips will die off just like 32 bit chips have.

We will see Operating Systems optimised for APU systems within 5 years.
 
Zarathustra[H];1040435867 said:
This won't happen.

The GPU portion of the chip can not run native x86 or x64 code, and even if it could it would be absolutely terrible at it. It would require specialized code, and even then only some types of tasks would lend themselves well to the massively parallelized capabilities of GPU processing.

On the other hand, I could see some code using the GPU portion as another extended instruction set like MMX or SSE and run certain portions of the code on it, but this would require a very smart compiler or painstaking manual coding, and specialized binaries just for AMD HSA enabled APU'S. A compiler like this could possibly even dump some code to the ARM core.

There will be no universal running the program or thread of your choice on the GPU or ARM core. Simply can not happen without specialized ported binaries.
Click to expand...

That's exactly what AMD is working on. They envision automatic direction of code to the CPU or GPU cores with minimal changes to current coding techniques and tools.
 
A CPU is made of of two part CISC and RISC, a GPU just does RISC, with a bit of work you could make the GPU part of the APU do ALL the RISC stuff and have the CPU part only do CISC.
 
Spazturtle said:
A CPU is made of of two part CISC and RISC, a GPU just does RISC, with a bit of work you could make the GPU part of the APU do ALL the RISC stuff and have the CPU part only do CISC.
Click to expand...

Except that's not how CPUs are designed anymore.

The last true CISC processor is arguably the first gen Atom. (In Order Execution Code only like the first Pentiums)

AMD made gains with the original Athlon and Thunderbirds because they were RISC built CPUs that broke down CISC code. They changed the architecture to get around the CISC bottlenecks of then current CPUs.
 
The problem was that you needed a lot of code to make it work , in the future I can see Linux offering this short pipeline to allow code to run directly on the GPU. In a sense something like OpenCL but on the kernel. That people from Microsoft can't or won't implement this and focus on trivial stuff like running OS from cloud (this made me laugh so hard especially after the stupid pawn stars video claiming chrome sucks). Linux would be the gaming operating system in the next 5 to 10 years.

If this does not happen there is always Mantle and that is more or less the short cut to al of the current Operating Systems shortfalls.
 
I really don't understand Amd's road-map going forward.

DDR4 will be available soon, so why wouldn't they incorporate that into their 2015 lineup with Carrizo.

I understand AM3+ is a dead platform, but if Amd has a better core (steamroller) why wouldn't they stick it on a AM3+ socket and sell it? esp since they are going to be sticking with DDR3

You can tell by looking at the current generation of Fm2+ boards that there will no 6-8 core Fm2+ processors, they simply don't have the power circuitry for it.

I don't know I'm baffled by this product road-map. Perhaps Amd has some unmentioned product that is going to come out of the gates. Much like the 290 and the 290x were (never were on any roadmap)
 
ebduncan said:
I understand AM3+ is a dead platform, but if Amd has a better core (steamroller) why wouldn't they stick it on a AM3+ socket and sell it? esp since they are going to be sticking with DDR3
Click to expand...

I think you underestimate what it takes to create a CPU package. Yes, they have the core design's figured out, but they don't have a manufacturing process for 4 module, 8 core Steamrollers. Someone probably made the calculation that based on predicted sales, it would cost more to convert the Piledriver 4 module process into a steamroller 4 module process than it would bring in from sales.
 
ebduncan said:
I really don't understand Amd's road-map going forward.

DDR4 will be available soon, so why wouldn't they incorporate that into their 2015 lineup with Carrizo.

I understand AM3+ is a dead platform, but if Amd has a better core (steamroller) why wouldn't they stick it on a AM3+ socket and sell it? esp since they are going to be sticking with DDR3

You can tell by looking at the current generation of Fm2+ boards that there will no 6-8 core Fm2+ processors, they simply don't have the power circuitry for it.

I don't know I'm baffled by this product road-map. Perhaps Amd has some unmentioned product that is going to come out of the gates. Much like the 290 and the 290x were (never were on any roadmap)
Click to expand...

I always figured that, if like DisplayPort and previous new technologies, that what is preventing the introduction of supported features like DDR4 or PCI-E 3.0 are two things:

  1. Adoption rate.
  2. Market availability.
"Why add support for something (i.e.- DisplayPort, DDR4) if there is neither a large adoption of the feature or the market hasn't shown interest in it?"

DDR4 is going to be expensive initially. There isn't anything conceivably useful for consumers other than higher bandwidth rates and lower power consumption. Adoption rate is definitely going to be slow and costs will be high-- look at DDR3 prices currently. And, how many products will support it outside of enthusiast (Haswell-E) and servers (Haswell-EP/EN) at the start?

And, the first market-wide support from a processor isn't until Intel's Skylake in 2015 (earliest) to 2016 (latest).

With DDR4 coming out in first availability for servers in the middle of 2014(-ish), it makes sense not to add support for it just yet in Kaveri APUs. For 2015 with Carizzo, even makes better sense since there still isn't a large availability of it yet.

I'd expect at least late 2015 to early 2016 for AMD to support it alongside PCI-E 4.0. This will at least have it near feature-parity with Intel's Skylake when that's available as well.

It'll also be hard, if not difficult, for consumers to adopt DDR4 right away since those DIMMs will be expensive initially. Little number of "normal" consumers means little to no market for DDR4 initially until costs come down and OEM computer manufacturers provide support for it. With little to no market availability of DDR4, means no processors will support it right away.

That's my thinking.
 
@octoberasian

Very intuitive. Do you think the market is switching from floor-sitting ATX towers to tiny miniITX boxes on top of the desk? i.e. Is ATX format dying?
 
reb00tin said:
@octoberasian

Very intuitive. Do you think the market is switching from floor-sitting ATX towers to tiny miniITX boxes on top of the desk? i.e. Is ATX format dying?
Click to expand...

In the consumer OEM space, I definitely think it's trending towards smaller form factors. Also, HTPCs practically scream small form factor.

In the enthusiast market, it's going both ways. The people that are building massive systems will continue to build even more massive systems, while those that want smaller systems will build small systems. For example, just look at the advent of enthusiast ITX cases and XL-ATX cases, both of which happened within the past ~3-5 years. The Corsair 900D is bigger than its predecessor, while the Bitfenix Prodigy is a hugely popular enthusiast ITX case.
 
Zarathustra[H];1040437656 said:
I think you underestimate what it takes to create a CPU package. Yes, they have the core design's figured out, but they don't have a manufacturing process for 4 module, 8 core Steamrollers. Someone probably made the calculation that based on predicted sales, it would cost more to convert the Piledriver 4 module process into a steamroller 4 module process than it would bring in from sales.
Click to expand...

These are not intel cpu's. Modules are easy to add and subtract this is the reason AMD went to a modular approach.

Not saying it wouldn't cost them anything to do it, just saying its not as expensive as you let out to be.

On the DDR4 thing, I think we all can agree that the AMD APU's need as much memory bandwidth as possible. They certainly do see gains with additional bandwidth.
 
You know both AMD and Intel only make one design right? They then use a computer program to make all the other designs.
 
Spazturtle said:
You know both AMD and Intel only make one design right? They then use a computer program to make all the other designs.
Click to expand...

I question your ability to make this statement considering you didn't even know that current CPUs are RISC based.
 
techrat said:
The last true CISC processor is arguably the first gen Atom. (In Order Execution Code only like the first Pentiums)
Click to expand...

Do tell how in-order-vs-ooo execution dictates cisc-vs-risc policy. ;)
 
ebduncan said:
These are not intel cpu's. Modules are easy to add and subtract this is the reason AMD went to a modular approach.

Not saying it wouldn't cost them anything to do it, just saying its not as expensive as you let out to be.

On the DDR4 thing, I think we all can agree that the AMD APU's need as much memory bandwidth as possible. They certainly do see gains with additional bandwidth.
Click to expand...

DDR4 is good for AMD that is for sure. But if you can recall last time AMD went from DDR2 to DDR3 it was the high end DDR2 which was still faster. The cpu both had a ddr2 and ddr3 memory controller on board that is prolly the sign you wanted from AMD.

For DDR4 to work in their advantage it might take another generation after the release of DDR4.
 
By the time DDR4 is adopted into mainstream PCs, AMD will have excavator well and truly designed.
 
ebduncan said:
These are not intel cpu's. Modules are easy to add and subtract this is the reason AMD went to a modular approach.

Not saying it wouldn't cost them anything to do it, just saying its not as expensive as you let out to be.

On the DDR4 thing, I think we all can agree that the AMD APU's need as much memory bandwidth as possible. They certainly do see gains with additional bandwidth.
Click to expand...

I don't think the design side of it is bad. As you say, there are automated layout tools that help in copying and pasting it.

I'm talking more about the updating of the production lines. Right now they presumably have two a few different lines, including one for APU's and one for Server/FX chips (which likely share a silicon production line, but have separate lines down the line due to the different sockets/packages they get)

In going to Steamroller they would likely have to process shrink the Piledriver Server/FX line from 32nm down to 28nm like Kaveri, and do some other changes to it, that would likely cost some real money.

It's not just the R&D that costs money in development. Often setting up production is just as, if not more costly.

This is where I think they probably did the cost benefit analysis for a FX Steamroller and looked at disappointing FX sales, and thought it probably wasn't worth it, and that they should focus their engineers and money on APU's instead because that's where the "future" is.
 
Zarathustra[H];1040440270 said:
I don't think the design side of it is bad. As you say, there are automated layout tools that help in copying and pasting it.

I'm talking more about the updating of the production lines. Right now they presumably have two a few different lines, including one for APU's and one for Server/FX chips (which likely share a silicon production line, but have separate lines down the line due to the different sockets/packages they get)

In going to Steamroller they would likely have to process shrink the Piledriver Server/FX line from 32nm down to 28nm like Kaveri, and do some other changes to it, that would likely cost some real money.

It's not just the R&D that costs money in development. Often setting up production is just as, if not more costly.

This is where I think they probably did the cost benefit analysis for a FX Steamroller and looked at disappointing FX sales, and thought it probably wasn't worth it, and that they should focus their engineers and money on APU's instead because that's where the "future" is.
Click to expand...
And, add to that, AMD doesn't have the operating income or research and development funding like Intel.

They turned a profit in the last quarter, but it's nothing compared to Intel.

It's more of a "We have this amount of money to spend, do we spend it on the next APU with a new CPU architecture that actually makes money, or do we spend it on the next FX processor that doesn't sell as much? Remember: We can only choose one."

If that was the case at AMD and I was them, I'd go with the former than the latter.
 
Actually AMD had the option to fund two things, APUs, FX or ARM.

They chose APUs and ARM server SoCs.

Hierofalcon is AMDs main money making play for 2014, being able to fit 8 10Gbps webservers in the same space as 1 1Gbps server took before.
 
xorbe said:
Do tell how in-order-vs-ooo execution dictates cisc-vs-risc policy. ;)
Click to expand...

If you weren't intentionally trying to misrepresent what I said, it would be understood.
 
Spazturtle said:
Actually AMD had the option to fund two things, APUs, FX or ARM.

They chose APUs and ARM server SoCs.

Hierofalcon is AMDs main money making play for 2014, being able to fit 8 10Gbps webservers in the same space as 1 1Gbps server took before.
Click to expand...

how about gpu?
 
You must log in or register to reply here.
Back
Top