New roadmaps out

[samm]

Unless somehow, all future software is magically optimized to get the most out of Hetergenous computing and Mantle, I dont see APU accomplishing anything in the enthusiast class. However in the mobile and HTPC markets these can rock out. Again I strongly believe in AMD as the only real competitor to Intel, and such a driving force for Intel to improve their future products. And i hope they do get better software support since I saw performance gains on the FX line after W8.

Questions for those who were there at the conference or during one of the live demos, did AMD demonstrate anything other than games and their amazing performance on one chip vs the 4770K? I wonder if these chips offer something new when it comes to single threaded performance or compute performance when compared to their Intel competition?

 

[Tsumi]

Everyone has the same train of thoughts that Geonerd has, show us then we believe but outright blind trust in software solutions by 3rd party is hard to swallow.

/facepalm facepalm facepalm

Did I say anywhere it will be successful? The ONLY thing I said was that THERE IS ABSOLUTELY NO REASON AN APU CAN'T BE AN ENTHUSIAST CLASS CPU. MAYBE YOU NEED TO KEEP REREADING THAT SENTENCE UNTIL YOU UNDERSTAND WHAT IT MEANS.

Unless somehow, all future software is magically optimized to get the most out of Hetergenous computing and Mantle, I dont see APU accomplishing anything in the enthusiast class. However in the mobile and HTPC markets these can rock out. Again I strongly believe in AMD as the only real competitor to Intel, and such a driving force for Intel to improve their future products. And i hope they do get better software support since I saw performance gains on the FX line after W8.

Questions for those who were there at the conference or during one of the live demos, did AMD demonstrate anything other than games and their amazing performance on one chip vs the 4770K? I wonder if these chips offer something new when it comes to single threaded performance or compute performance when compared to their Intel competition?

AMD has been pushing devs and the focusing a lot on the gaming industry, just look at Mantle and its adoption in future games.

 

[NaroonGTX]

Everyone has the same train of thoughts that Geonerd has,

Lol nope.

 

[Geonerd]

/facepalm facepalm facepalm

Did I say anywhere it will be successful? The ONLY thing I said was that THERE IS ABSOLUTELY NO REASON AN APU CAN'T BE AN ENTHUSIAST CLASS CPU. MAYBE YOU NEED TO KEEP REREADING THAT SENTENCE UNTIL YOU UNDERSTAND WHAT IT MEANS.

AMD has been pushing devs and the focusing a lot on the gaming industry, just look at Mantle and its adoption in future games.

Gee, that's nice. That and a buck will get me a cup of coffee.

In year or two, maybe, this whole GPU/APU/HSA scheme might begin to bear fruit.
Until then, it's just so much marketing monkey hooting and hollering. Anyone care for a round of Buzzword Bingo?
Until then, x86 is where it's at.

I'll say it again, and since you like to shout....
"SHOW ME THE APPS!"

Until then, please feel free to chug the sweet fruit flavored liquid, but don't get all high and mighty when people ask legitimate questions.

 

[Tsumi]

Gee, that's nice. That and a buck will get me a cup of coffee.

In year or two, maybe, this whole GPU/APU/HSA scheme might begin to bear fruit.
Until then, it's just so much marketing monkey hooting and hollering. Anyone care for a round of Buzzword Bingo?
Until then, x86 is where it's at.

I'll say it again, and since you like to shout....
"SHOW ME THE APPS!"

Until then, please feel free to chug the sweet fruit flavored liquid, but don't get all high and mighty when people ask legitimate questions.

Maybe you should go read the thread again.

Some person was claiming the APU only route was abandoning the enthusiast market for good. I'm saying that's false. Of course, people can't read the thread, that's why they say stupid things like this.

 

[luminousone]

HSA is a game changer, its not just some more instructions or piece of software, its more fundamental then that.

HSA interacts with parts of the x86 arch that haven't been touched sense i386, it changes how memory works, it changes how cache works, it changes what the gpu can access at runtime.

On current systems you have separate gpu and cpu memory, and I am not just plainly talking the physical separation of where the memory is located, I am talking about how each system can or can't access the others memory, the need to copy things back and forth between those memories even on systems with unified memory(current apu's, intel igpu, etc).

HSA doesn't need to copy between those 2 types of memory, because the hMMu allows the different memories address spaces to overlap, effectively making the entirety of system and gpu memory available to both the cpu and gpu. The cpu only need give the gpu the location of the memory to be worked on, without the need to copy anything. The GPU can access cpu memory, such that temporary variables in the stack or heap can easily be accessed without any special activity to pre-upload that into the gpu.

And it is safe for the GPU to access memory in a willy nilly fashion as the GPU has access to the CPU's cache protocol, meaning no need to flush the cache before sending the gpu a workload.

And further, because AMD really though this all through, they have created HSAIL, which system languages can be compiled to, so C/C++ can compile directly into GPU code, and LLVM and anyother compiler that supports this can do so WITHOUT SOURCE CHANGES!, that to say that will be generally optimal, but any vectorizing compiler will be able todo this.

HSAIL is neat as well, as it can sit along side your x86-32 or x86-64 code inside your (MS).exe or ELF binary's. meaning that supporting HSA breaks compatibility with NOTHING.

Many software applications need only be recompiled to support HSA, if these applications already use openCL they will still see a benefit from the vastly improved memory situation provided by HSA(elimination of needless memory copies).

HSA is a game changer, Intel will be forced to join sooner or later, their idiotic approach of throwing ever wider and wider registers at the problem of parallel datasets suffers from a rapid rate of diminishing returns, require extensive optimization beyond what the compiler can provide(its a bitch to optimize around 256bit/512bit registers and follow numeras rules for instruction pairing), HSA via GCN or similar arch(hell nvidia fits this description), are easy to optimize for, being only slightly more complicated that simple single issue pipelines.

Again HSA is AWESOME.

 

[NaroonGTX]

Gee, that's nice. That and a buck will get me a cup of coffee.

In year or two, maybe, this whole GPU/APU/HSA scheme might begin to bear fruit.
Until then, it's just so much marketing monkey hooting and hollering. Anyone care for a round of Buzzword Bingo?
Until then, x86 is where it's at.

I'll say it again, and since you like to shout....
"SHOW ME THE APPS!"

Until then, please feel free to chug the sweet fruit flavored liquid, but don't get all high and mighty when people ask legitimate questions.

Seriously, you can't possibly be this dumb. Go read the thread again and perhaps you'll stop with this nonsense.

 

[Mackintire]

Unless somehow, all future software is magically optimized to get the most out of Hetergenous computing and Mantle, I dont see APU accomplishing anything in the enthusiast class. However in the mobile and HTPC markets these can rock out. Again I strongly believe in AMD as the only real competitor to Intel, and such a driving force for Intel to improve their future products. And i hope they do get better software support since I saw performance gains on the FX line after W8.

Questions for those who were there at the conference or during one of the live demos, did AMD demonstrate anything other than games and their amazing performance on one chip vs the 4770K? I wonder if these chips offer something new when it comes to single threaded performance or compute performance when compared to their Intel competition?

I answered your question earlier. AMD did not provide the information. User's that leaked the results of the engineering samples did. In case you missed it....Single threaded performance for the steamroller cores is approximately the same has intel's nehalem chips. So this is a big improvement. To put this into perspective that puts AMDs single threaded instruction per clock IPC within 15% of intels current chips.

But steamroller based units will be going up against Haswell and or Broadwell chips. So I expect AMD will still be running a 18-20% IPC disadvantage.

And....AMD will be running 28nm while intel is using 22-14nm die size.

AMD needs to get to 20nm as soon as possible regardless of whether they use steamroller or excavator cores.

 

[ebduncan]

Amd is going to 20nm.

Anyways there is tons of debate about Kaveri. First steam roller cores have made a large improvement over the previous versions.

HSA isn't the strong point of Kaveri. It will take time and application support will be critical. However HSA is the newest iteration on how to get the most out of today's processors. Amd introduced the first Dual core cpu. Amd introduced the first cpu with built in memory controller.

The trend continues. Amd will release the first cpu with HSA. While it certainly won't make a big difference now, it is the future of Cpu's.

 

[NaroonGTX]

HSA pretty much is the strong point of Kaveri. It has to start somewhere, and this is the first chip to have HSA at an architectural level. Excavator will be the point where HSA is baked in at system-integration level. That will happen about a year after Kaveri launches in Q1 2015 with the Carrizo APU, where Carrizo's goal is to completely dwarf the performance of Kaveri.

 

[Pieter3dnow]

Some person was claiming the APU only route was abandoning the enthusiast market for good. I'm saying that's false. Of course, people can't read the thread, that's why they say stupid things like this.

Not some person their own roadmap of 2010/2011 showed that Excavator would be APU only. And now AM3+ is not going to be practical for future use. So when NaroonGTX states no steamroller FX in 2014 there is no need for it in 2015 either.

And yes enthusiast market is cpu based .

 

[luminousone]

HSA support in applications will happen rather quickly on linux platforms, windows will take a bit longer, but either way HSA is really really easy to implement into software, basic support only requiring a vectorizing compiler that supports the HSAIL target, and a recompile of the software with this compiler.

Microsofts C++AMP supports HSA, and support for C/C++/etc can be obtained via LLVM.

And HSAIL bytecode sits along side x86/64 bytecode in your EXE/ELF binary files, so supporting it costs almost nothing. And the benefits of providing basic support are very high. I would be shocked if you didn't see a good number of developers support HSA within the first year, and Developers that already use tools such as OpenCL will see a boost as well from a supporting compiler as the ability to avoid needless memory copies would still provide a sizable performance benefit.

 

[Tsumi]

Not some person their own roadmap of 2010/2011 showed that Excavator would be APU only. And now AM3+ is not going to be practical for future use. So when NaroonGTX states no steamroller FX in 2014 there is no need for it in 2015 either.

And yes enthusiast market is cpu based .

Have you read anything AT ALL!?

HSA's design goal is to ACCELERATE CPU TASKS USING THE IGPU! That's why AMD is pushing HSA aggressively, and it makes sense to get the tools as user friendly as possible so that more developers would be inclined to use them. And then there's the entire HSA foundation to consider.

 

[Pieter3dnow]

Have you read anything AT ALL!?

HSA's design goal is to ACCELERATE CPU TASKS USING THE IGPU! That's why AMD is pushing HSA aggressively, and it makes sense to get the tools as user friendly as possible so that more developers would be inclined to use them. And then there's the entire HSA foundation to consider.

No i just hit the random reply button from time to time because I like people to ask me if I read anything and just for fun make them post things in BOLD and underlined.
A company always stands behind their products when did AMD not pursuit to push their product or tries to get more allies to see the good work that they are doing.

 

[omega-x]

HSA is brilliant.

It will set the direction of computing for decades.

That said: Kaveri might not be the dogs bollocks of performance (or might be, if the accelerations possible are more drastic than 2x), but the next few generations and die shrinks will whip it into shape while intel rests on their laurels and when that parity happens, we'll be right back into progress.

 

[luminousone]

HSA will also appear on ARM based processors, both in the server and mobile spaces, and not just on AMD manufactured ARM chips, as its an open standard.

HSAIL can as easily sit along side ARMv8 bytecode as well as x86 bytecode in ELF binaries. Their is work being done to allow java to use HSAIL, so maybe we will see it make its way into android phones after Google figures out how the hell to make dvm work with 64bit arm.

 

[tybert7]

maybe a dumb question but... why are there no 8 core kaveri apus for the desktop incoming?

I thought some newer games could take advantage of more than 4 cores, and with mantle incoming, they are favoring more cores to tackle performance rather than ipc, so restricting kaveri on the desktop to 4 cores would make the 8350 the superior desktop cpu... is that intended?

 

[Wildace]

pc cant handle it, if you want an 8 core keveri buy a ps4....

jk

 

[ccityinstaller]

maybe a dumb question but... why are there no 8 core kaveri apus for the desktop incoming?

I thought some newer games could take advantage of more than 4 cores, and with mantle incoming, they are favoring more cores to tackle performance rather than ipc, so restricting kaveri on the desktop to 4 cores would make the 8350 the superior desktop cpu... is that intended?

Honestly I think it is a matter of both die and power constraints..Even @ 28nm, AMD is packing a quadcore cpu in with a fairly large GPU all while staying in under a ~86W (IIRC??) TDP..They could certainly do a "high performance" model with 8 cores, but that would most likely drive the TDP up 1.5~1.8X, and require a much larger die (50~65% depending if they kept the L2 cache the same, which I doubt)...

Do you remember the backlash against AMD for having the nerve to launch a 220W TDP 5Ghz 8 core CPU? They were laughed at and spit on in the forums, all because they chose to TELL THE TRUTH about how much power a highly clocked 8 core CPU consumes..

IF Intel were to launch a 5Ghz (say 4.5Ghz "base clock Haswell part (let's call it a 4790K) they would advertise it using the base clock's TDP, instead of the MUCH HIGHER TDP of the max turbo bin @ 5Ghz..

AMD choose not to do that, and I applaud them for it..Suppose they launched it and used the non turbo TDP of say 160W..Many people with old AMD3+ boards (with cheap 4+1 Power Phases) would rush out to buy them, and they come crying when they forced all 8 cores to run @ max Turbo Speed and their MB VRMs burst into flames..

 

[Pieter3dnow]

Honestly I think it is a matter of both die and power constraints..Even @ 28nm, AMD is packing a quadcore cpu in with a fairly large GPU all while staying in under a ~86W (IIRC??) TDP..They could certainly do a "high performance" model with 8 cores, but that would most likely drive the TDP up 1.5~1.8X, and require a much larger die (50~65% depending if they kept the L2 cache the same, which I doubt)...

Do you remember the backlash against AMD for having the nerve to launch a 220W TDP 5Ghz 8 core CPU? They were laughed at and spit on in the forums, all because they chose to TELL THE TRUTH about how much power a highly clocked 8 core CPU consumes..

IF Intel were to launch a 5Ghz (say 4.5Ghz "base clock Haswell part (let's call it a 4790K) they would advertise it using the base clock's TDP, instead of the MUCH HIGHER TDP of the max turbo bin @ 5Ghz..

AMD choose not to do that, and I applaud them for it..Suppose they launched it and used the non turbo TDP of say 160W..Many people with old AMD3+ boards (with cheap 4+1 Power Phases) would rush out to buy them, and they come crying when they forced all 8 cores to run @ max Turbo Speed and their MB VRMs burst into flames..

That what this whole thread is almost about AMD "gambles" on software developers using HQ huma theoretically a 4 core APU setup should provide more processing power then a 8 core cpu. http://www.youtube.com/watch?v=4YV6z6Fgw48

This is what AMD sees as the future. Nothing wrong with that by the way. But to get there some steps need to be taken. Between now and then for performance you still would need a videocard next to your APU.

And if someone was to make an optimized APU benchmark it would ROTLMAO any Intel cpu. Like what you see already with OpenCL. So the performance is there but the foundation for every developer to use this is not.

 

[Tsumi]

That what this whole thread is almost about AMD "gambles" on software developers using HQ huma theoretically a 4 core APU setup should provide more processing power then a 8 core cpu. http://www.youtube.com/watch?v=4YV6z6Fgw48

This is what AMD sees as the future. Nothing wrong with that by the way. But to get there some steps need to be taken. Between now and then for performance you still would need a videocard next to your APU.

And if someone was to make an optimized APU benchmark it would ROTLMAO any Intel cpu. Like what you see already with OpenCL. So the performance is there but the foundation for every developer to use this is not.

You would always need a video card next to the APU for high end gaming performance... AMD isn't aiming for the APU to replace discrete graphics in the high-end market, they're aiming for the APU to replace the traditional CPU from the high-end market.

 

[NaroonGTX]

maybe a dumb question but... why are there no 8 core kaveri apus for the desktop incoming?

I thought some newer games could take advantage of more than 4 cores, and with mantle incoming, they are favoring more cores to tackle performance rather than ipc, so restricting kaveri on the desktop to 4 cores would make the 8350 the superior desktop cpu... is that intended?

There's a few reasons.

One is that the only reason there were octocores in the first place with the FX label is because they were derived directly from the Orochi Opteron chips. They were literally the same die, just with a few of the HyperTransport links shut off. They never had plans for an octocore Kaveri die, however the original version of Steamroller had plans for GDDR5 and up to 3 CU's (modules, in other words, up to six cores), but those plans got scrapped along with the first version of Steamroller, so after they went back to the drawing board, they settled on DDR3 only and up to 2 CPU CU's, for 4 cores max.

They technically very well could make a Kaveri with 4 modules, but the die size would be bigger as a result, and for little reason. A lot of enthusiasts would like moar coarz, but the main reason Kaveri is existing is to push for HSA. Look at the GLOPs each part produces: the CPU portion produces around ~130 GFLOPs or so, while the GPU pushes like 750+ GLOPs. If the GPU gets overclocked, it easily pushes the 1 Teraflop barrier.

The main point of integrating a power GPU on the die (besides efficiency) is for HSA compute. This will allow applications that support it to be sped up tremendously, especially with the shared pool of memory. The thing is that Kaveri is the first APU that has "true" HSA, and for now it is only an architectural implementation. Excavator will be there at the system/hardware integration level, and will provide even more performance that will eclipse that of Kaveri.

As for Mantle, it makes sense that AMD chose to keep Vishera for the "Performance" segment. At APU13, there was a developer who showed a slide where they mentioned how with Mantle, they could take an FX-8350, downclock ALL cores to 2GHz, and STILL be completely GPU-bound. This makes things such as blazing-fast IPC totally irrelevant, and this benefits everyone, regardless of whether their CPU is Intel or AMD. One of the largest parts of Mantle, for gaming anyway, is the reduction of CPU overhead, which allows for many more draw calls to be made to the GPU, increasing performance.

In regards to Kaveri vs Vishera, you will see Kaveri spanking Vishera and Trinity/Richland in many benches, even some multi-threaded ones. The 8350 only scales up to around 650% instead of 800% like it should, because of the CMT penalty. Kaveri has Steamroller x86 cores, which in the uArch, fixed that issue, so the scaling is 100% for each individual core. This means you will see Kaveri even possibly beating some of the hexacores.

 

[MacLeod]

There's a few reasons.

One is that the only reason there were octocores in the first place with the FX label is because they were derived directly from the Orochi Opteron chips. They were literally the same die, just with a few of the HyperTransport links shut off. They never had plans for an octocore Kaveri die, however the original version of Steamroller had plans for GDDR5 and up to 3 CU's (modules, in other words, up to six cores), but those plans got scrapped along with the first version of Steamroller, so after they went back to the drawing board, they settled on DDR3 only and up to 2 CPU CU's, for 4 cores max.

They technically very well could make a Kaveri with 4 modules, but the die size would be bigger as a result, and for little reason. A lot of enthusiasts would like moar coarz, but the main reason Kaveri is existing is to push for HSA. Look at the GLOPs each part produces: the CPU portion produces around ~130 GFLOPs or so, while the GPU pushes like 750+ GLOPs. If the GPU gets overclocked, it easily pushes the 1 Teraflop barrier.

The main point of integrating a power GPU on the die (besides efficiency) is for HSA compute. This will allow applications that support it to be sped up tremendously, especially with the shared pool of memory. The thing is that Kaveri is the first APU that has "true" HSA, and for now it is only an architectural implementation. Excavator will be there at the system/hardware integration level, and will provide even more performance that will eclipse that of Kaveri.

As for Mantle, it makes sense that AMD chose to keep Vishera for the "Performance" segment. At APU13, there was a developer who showed a slide where they mentioned how with Mantle, they could take an FX-8350, downclock ALL cores to 2GHz, and STILL be completely GPU-bound. This makes things such as blazing-fast IPC totally irrelevant, and this benefits everyone, regardless of whether their CPU is Intel or AMD. One of the largest parts of Mantle, for gaming anyway, is the reduction of CPU overhead, which allows for many more draw calls to be made to the GPU, increasing performance.

In regards to Kaveri vs Vishera, you will see Kaveri spanking Vishera and Trinity/Richland in many benches, even some multi-threaded ones. The 8350 only scales up to around 650% instead of 800% like it should, because of the CMT penalty. Kaveri has Steamroller x86 cores, which in the uArch, fixed that issue, so the scaling is 100% for each individual core. This means you will see Kaveri even possibly beating some of the hexacores.

Good read. Hadn't really heard it put like that. How much of that depends on 3rd parties signing on and programming for that? Or is it the way it's gonna be in the future regardless?

 

[NaroonGTX]

Do you mean Mantle? In regards to Mantle, engine developers can choose to bake it directly into the engine since it's an API, in which case when it comes to the rest of the developers working on a game, it would be as simple as flipping an on/off switch when coding their game. DICE is pretty much the first dev to talk about Mantle in their engines, and EA has announced that pretty much all of their upcoming games will have Mantle support as a result. So the adoption will be rather wide-spread.

I think the next version of CryEngine will have it, since Star Citizen runs on that and will have Mantle support, and there was another CryEngine game that will have it, though I forgot what it was called. Once more and more people see the sheer performance boosts and optimization they can get with it, it will spread like wild-fire. Since developers themselves have been asking for it, I can see it becoming popular and eventually being in virtually every modern engine used on AAA games on the PC.

It will also make console-to-PC ports much better since the efficiency will be there; the X1 and PS4 both are x86-64-based and both have GPU's based on the GCN architecture, so that means quicker, well-optimized ports/versions in the PC's future with little hassle.

 

[samm]

From the last couple of posts AMD is guaranteeing its future by releasing protocols that are easy to use by developers and get the most out of their hardware. Pretty neat.

However it seems as though theyre gonna ignore software designed before the advent of HSA and mantle.
Gaming engines are covered since they use similar hardware in the consoles. $$$ to be made by game studios.
Other software is key here. Have they gotten any strong backers from say MS or Adobe or the such?

 

[NaroonGTX]

In terms of HSA, there are a lot of companies in the HSA Foundation. You can see which companies have officially become members on the official website. HSA is fairly-strongly backed right now, and I can only assume it'll get better as time goes on. This is as big as the Athlon 64, perhaps bigger.

We'll get more details at CES 2014, which is about a week or so before the 'initial launch' of Kaveri on Jan. 14th. Hopefully we see some benchmarks before then, possibly next month if AMD starts sending out chips to various partners and associates.

 

[MacLeod]

In terms of HSA, there are a lot of companies in the HSA Foundation. You can see which companies have officially become members on the official website. HSA is fairly-strongly backed right now, and I can only assume it'll get better as time goes on. This is as big as the Athlon 64, perhaps bigger.

We'll get more details at CES 2014, which is about a week or so before the 'initial launch' of Kaveri on Jan. 14th. Hopefully we see some benchmarks before then, possibly next month if AMD starts sending out chips to various partners and associates.

Yeah that's what I was mainly asking about. I knew Mantle had a promising future with AMD sewing up most of the gaming world with consoles and all that but wasn't sure how bright the future was for HSA and how big a gamble it was.

 

[Tsumi]

It is a fairly big gamble. In fact, AMD pretty much gambled it's entire future on it when they began designing Bulldozer back in what, 2006? Actually, it probably started with the acquisition of ATI.

There's not a lot they can do about older software. Maybe they can have some sort of layer in-between that can detect and convert the instructions for HSA computing, but it won't be nearly as efficient as a program designed to take advantage of HSA. The best they can do is to convince developers to use their tools by making it as easy to use as possible. At best, all it would be is a simple copy and paste, or just a recompilation of the software.

 

[Digital Viper-X-]

It is a fairly big gamble. In fact, AMD pretty much gambled it's entire future on it when they began designing Bulldozer back in what, 2006? Actually, it probably started with the acquisition of ATI.

There's not a lot they can do about older software. Maybe they can have some sort of layer in-between that can detect and convert the instructions for HSA computing, but it won't be nearly as efficient as a program designed to take advantage of HSA. The best they can do is to convince developers to use their tools by making it as easy to use as possible. At best, all it would be is a simple copy and paste, or just a recompilation of the software.

I'm assuming a lot here since I haven't read up on HSA

is HSA putting the GPU in the place of the dedicated FPU? or will it be working as along side it? I
Does the GPU have access to the l2 cache on the CPU?

Sorry newb questions all at once.

 

[Tsumi]

I'm assuming the iGPU has its own L2 cache, as L2 cache tends to be core-specific (or module specific in AMD's case).

As far as I know, it won't completely replace the FPU.

 

[os2wiz]

From the last couple of posts AMD is guaranteeing its future by releasing protocols that are easy to use by developers and get the most out of their hardware. Pretty neat.

However it seems as though theyre gonna ignore software designed before the advent of HSA and mantle.
Gaming engines are covered since they use similar hardware in the consoles. $$$ to be made by game studios.
Other software is key here. Have they gotten any strong backers from say MS or Adobe or the such?

Adobe is a key partner in the HSA Foundation. All their future software will be HSA enabled. HP is part of the foundation, Oracle and JAVA will have all their software HSA enabled. Other big names will be announced in the next few months.

 

[xorbe]

As far as I know, it won't completely replace the FPU.

That has to be impossible, unless there's a "tightly coupled mode" like the original StrongArm design featured for the 3-wide vector co-processor.

 

[KazeoHin]

If oracle hops on the HSA train, we will see a huge improvement in FP calculations in Java, that means one thing and one thing only...


MINECRAFT WILL RUN SO FAST.

 

[NaroonGTX]

AMD hasn't announced any plans to get rid of the FPU, so if it doesn't get replaced, it would be some processor in 2016 or beyond.

AMD had plans for the Bulldozer-style uarch for quite some time. I remember there was a mention of a uarch very similar to the modular concept being talked about around 2003~2004. HSA was a major reason for them acquiring ATI back in 2006.

 

[octoberasian]

I'm assuming a lot here since I haven't read up on HSA

is HSA putting the GPU in the place of the dedicated FPU? or will it be working as along side it? I
Does the GPU have access to the l2 cache on the CPU?

Sorry newb questions all at once.

I'm assuming the iGPU has its own L2 cache, as L2 cache tends to be core-specific (or module specific in AMD's case).

As far as I know, it won't completely replace the FPU.

As far as I can tell, if the iGPU on Kaveri is similar to the GPU cores in the Radeon cards, they all have an L2 cache already.

GCN 1.0:
http://www.anandtech.com/show/4455/amds-graphics-core-next-preview-amd-architects-for-compute/5

GCN 1.1:
http://images.anandtech.com/doci/7457/HawaiiArch.png

I believe Kaveri is using the GCN 1.1 core first introduced in the 7790, so it will have an L2 cache.

I assume that when we get to Excavator, this will no longer be the case-- the CPU and GPU units will be pretty much "fused" together and will most likely share one unified L2 cache.

As for the FPU, I believe the intention for AMD's Fusion (hardware) and HSA (software) is to have the GPU units assist or complement the FPU. I remember someone mentioned in an old post that the GPU and FPU can only perform certain FPU functions unique to their functions, so the FPU isn't going away anytime soon. For example, the GPU is primarily vector and rasterization calculations (outside of AVX instruction set), but does great when it comes to SP and DP number crunching. The FPU does stuff like FMAC, MMX and AVX, which I don't believe any GPU units can do those currently unless they're redesigned. And, I assume that for AMD that will not happen until Excavator.

This is all assumption based on what I've read so far up until now. All we can do is assume what and how on these APUs and HSA work for the moment until more literature is revealed.

 

[SystemCrash]

I have some serious doubt about this roadmap accuracy.

It show that Kaveri should be available at the end of 2013, when they clearly mentioned that it will be only be available in early 2014.

Also, they didn't release much info about the upcoming kaveri... a lot of ppl expected more juicy informations

 

[NaroonGTX]

The roadmap is accurate. Kaveri starts shipping to customers in Dec. 2013. Customers = OEM's and ISV's -- AMD doesn't sell their products directly to us, the consumers. They send them to OEM's and ISV's who then put them in rigs for others to buy, and they send them to retailers to sell in PIB (processor in a box) packaging.

Kaveri is available from OEM's and ISV's on Jan. 14th, 2014, and hits retail in Feb. 2014.

 

[Tsumi]

As far as I can tell, if the iGPU on Kaveri is similar to the GPU cores in the Radeon cards, they all have an L2 cache already.

GCN 1.0:
http://www.anandtech.com/show/4455/amds-graphics-core-next-preview-amd-architects-for-compute/5

GCN 1.1:
http://images.anandtech.com/doci/7457/HawaiiArch.png

I believe Kaveri is using the GCN 1.1 core first introduced in the 7790, so it will have an L2 cache.

I assume that when we get to Excavator, this will no longer be the case-- the CPU and GPU units will be pretty much "fused" together and will most likely share one unified L2 cache.

As for the FPU, I believe the intention for AMD's Fusion (hardware) and HSA (software) is to have the GPU units assist or complement the FPU. I remember someone mentioned in an old post that the GPU and FPU can only perform certain FPU functions unique to their functions, so the FPU isn't going away anytime soon. For example, the GPU is primarily vector and rasterization calculations (outside of AVX instruction set), but does great when it comes to SP and DP number crunching. The FPU does stuff like FMAC, MMX and AVX, which I don't believe any GPU units can do those currently unless they're redesigned. And, I assume that for AMD that will not happen until Excavator.

This is all assumption based on what I've read so far up until now. All we can do is assume what and how on these APUs and HSA work for the moment until more literature is revealed.

I don't think they'll do a fused L2, as I've said, the L2 cache tends to be specific to the core/module. What I think we will see is a shared L3 cache.

 

[NaroonGTX]

I remember a quote from a blog post from an AMD engineer a while back in response to someone who asked about a universal unified L2 cache. Basically, it would be extremely problematic and not worth it, which is why there are only individual clusters of shared L2 cache for each module.

When it comes to HSA, the CPU & GPU will have their unified virtual memory, no need for shared L2 caches however.

 

[ClariorHincHonos]

If oracle hops on the HSA train, we will see a huge improvement in FP calculations in Java, that means one thing and one thing only...


MINECRAFT WILL RUN SO FAST.

Technically they've started to already. Flash offloads to GPU now, or can use CPU if it's disabled.

*Edit - Oops, nevermind. I confused the two always-updating-annoying-ass-apps.