Nvidia and X86 license

[cbn]

http://www.tgdaily.com/content/view/41628/135/

I know this is a 2 month old article but looking at the intel roadmap yesterday got me thinking.

At the moment all CPUs past Sandy Bridge are going to be 8 core by default.....but for consumers to use all these extra cores the x86 software needs to be specifically coded "multi-threaded" right?

Is there someway Nvidia could bypass the need for specific "multi-threaded" software with their parallel multi-core technology? So instead of "single threaded" programs only using 1 core of a octo core CPU they could more efficiently be spread out over all the cores?

The reason I am asking is because at the moment I am not seeing myself needing so many CPU cores in the future. I would rather have fewer but faster cores (if anything) the way current CPU software is coded.

What happens to x86 technology once intel starts making 16 and 32 core CPUs.....but most common everyday programs are based on "single thread". Will some "other technology" from the GPU world that is able to spread out "single threaded" programs over many cores start to look at lot more attractive and less wasteful.



Moved from the Video forum.

AmberClad
Video Moderator

 

[Idontcare]

If a program already coded for x86 processors can't be efficiently ran on 8 cores of native x86 CPU goodness then it sure as heck isn't going to scale up and run any faster on a massively parallel GPGPU.

GPGPU's step in where the limited cores of a CPU leave an otherwise embarrassingly parallel program wanting more cores.

Basically you are left with a small class of applications that are embarrassingly parallel and coarse-grained which can run on the hybrid beasts (fusion) that are coming down the pipe:

See far-right hand side of this graphic

But for those "mostly serial" programs they will continue to be best served by the robust processing capability (called an ISA) of a modern x86 CPU, so too the fine-grained parallel programs until fusion is here.

 

[cbn]

Originally posted by: Idontcare
If a program already coded for x86 processors can't be efficiently ran on 8 cores of native x86 CPU goodness then it sure as heck isn't going to scale up and run any faster on a massively parallel GPGPU.

GPGPU's step in where the limited cores of a CPU leave an otherwise embarrassingly parallel program wanting more cores.

Basically you are left with a small class of applications that are embarrassingly parallel and coarse-grained which can run on the hybrid beasts (fusion) that are coming down the pipe:

See far-right hand side of this graphic

But for those "mostly serial" programs they will continue to be best served by the robust processing capability (called an ISA) of a modern x86 CPU, so too the fine-grained parallel programs until fusion is here.

[Thanks for this detailed answer. It is way over my head but you still did a good explaining it to me]

I brought this up because it seems Intel/AMD have seemingly reached a certain type of limit with respect to performance with one CPU core. They can't efficiently go faster (more clock speed) so instead they have decided to go "lateral" instead by implementing more CPU cores.

Could some type of hybrid CPU be constructed that would be able to spread "single threaded" programs out over many cores? Don't GPUs do something like this already?

I'm asking because I am wondering how useful an eight/sixteen/thirty-two core CPU will be to the average consumer?

 

[Nemesis 1]

AMD Bulldoozer promises good power increases. As does Intel SandyBridge . Haswell Looks to be tech change taking on larrabee type qualities.

 

[cbn]

Originally posted by: Nemesis 1
AMD Bulldoozer promises good power increases. As does Intel SandyBridge . Haswell Looks to be tech change taking on larrabee type qualities.

Haswell is the default 8 core CPU I am talking about that is beyond Sandy Bridge.

I wonder what kind of GPU (larabee) type qualities it will have?

 

[Idontcare]

Originally posted by: Just learning

Originally posted by: Nemesis 1
AMD Bulldoozer promises good power increases. As does Intel SandyBridge . Haswell Looks to be tech change taking on larrabee type qualities.

Haswell is the default 8 core CPU I am talking about that is beyond Sandy Bridge.

I wonder what kind of GPU (larabee) type qualities it will have?

Goto-san thinks the future of Haswell could involve Larrabee.

http://pc.watch.impress.co.jp/...html/kaigai02.jpg.html

It stands to reason, at 22nm and 16nm a Larrabee core is going to be really tiny, on the order of a couple mm^2 per core. Integrating 20 or 30 such cores into a heterogeneous (fusion-type) processing ensemble alongside Haswell-type cores would provide a really exciting opportunity for balancing the performance needs of single threaded as well as fine-grained parallel apps while the embarrassingly parallel and coarse-grained applications can then scale like crazy as the thread scheduler throws threads across all the larrabee cores under the IHS.

Basically once the OS support is there for distinguishing between real cores and hyperthreading virtual cores, the support needed to deal with a heterogeneous core environment like this will already be present and require little additional tuning to account for the larrabee cores versus haswell cores capabilities.

 

[Nemesis 1]

I do believe your talking along these lines.

http://www.theinquirer.net/inq...498/ati-ups-gpgpu-ante

This what ya had inmind but haswell with larrabee. I believe This happens on sandy . Haswell I believe is going to be a world shocker with Zero x86 decoders. Since larrabbee carries x86 decoders no need for haswell to carry. As Larrabee has by this time allready done recompiles for X86 programms. Haswell Has it own Vector unit like larrabee Vector unit. Vector unit backends are usually VLIW from C/C++. Just as handheld section is moving to. Amd is setting here with ATI. So for Intel its smart to move towards amd not away.

 

[Nemesis 1]

Thats why I said Intel changing the AVX standard was maybe a warning shot to AMD from Intel saying . Don;t follow us its a dead end for you. Maybe Intel isn't so evil. Intel seems very happy with AMD at 20% . AMD isn't . Intel could lower pricies and reduce AMD to 10% if the wish to . Haswell with Vector unit says one thing to me . Haswell cores will not carry X86 decoders. The Larrabee side kick is there . to protect popa X86 instructions. Since by the time were talking . Larrabee will have alreadu done much of the x86 recompile work . Didn't you think it strange that Intel used native larrabee and than on Sandy they used AVX. None really understand this at this time. But I have a hunch. That If Haswell does get a Vector unit. and FMA. AMDs FMA on BD should tell us a lot . I think it will suck big time. Thats why intel changed spec. A warning shot . Hay don't follow us here . Were going to the land of less complex logic in cpus and Better compilers with software to accomplish our goals its a great endeavor and I wish them Luck on this path they clearly have choozen .

I am willing to bet sizeable amount it won't be on a cisc backend. or ondie x86 decoders.

Problem is thats way beyond my dreamstate.

Heres how if all goes as it should this plays out. At least for me this is idea.

Intel wins cpu war with Haswell.

ATI wins GPU war .

Intel / AMD/ATI combine to form one platform / silcon of = value. For both . But remain competitive in chipset and other segments. This could happen . A deal that makes AMD /Intel both happy./

 

[alyarb]

give some examples of single threaded applications that you think are too slow.

 

[Idontcare]

Originally posted by: alyarb
give some examples of single threaded applications that you think are too slow.

Backtesting in Metatrader 4.

 

[alyarb]

what is the nature of the work? how do you propose to parallelize a naturally sequential series?

 

[Idontcare]

Originally posted by: alyarb
what is the nature of the work? how do you propose to parallelize a naturally sequential series?

Metatrader 4 is a trading platform primarily used in forex markets.

I have no proposal for parallelizing inherently serial code. That is why I want more GHz.

 

[alyarb]

exactly. let's face it, not everything is parallelizable or even worth trying to parallelize. the majority of of applications that people use are single threaded and have been "fast enough" for some years, which is why ultra low power machines for general office productivity are becoming popular even with weak as hell atom CPUs. i foresee high-performance single-threaded applications to become a very special case over the years. many single-threaded workloads will become extinct over time as the good candidates are rewritten for multiprocessing capability. any inherently single-threaded applications that remain are likely to never make it into the never never land of parallelism. to address this, 3 or 4 years from now, we will probably see some very exotic implementations of "turbo mode."

if we are talking about the 22nm timeframe, it wouldn't be outrageous to imagine an 8-core CPU with a 32-core larrabee derivative integrated onto the die. the ability to dynamically power up and power down execution units is going to become critical and it will introduce a mitigation for single-threaded sticks in the mud. If the entire chip runs at ~3 GHz, it's not crazy to estimate that the larrabee region could be powered down and a single haswell core temporarily accelerated to 4+ GHz for the duration of the single-threaded work without violating TDP. Considering 45nm dual cores can reach comparatively high frequencies at very low voltages while operating as a single core, my estimates would have to be unrealistically conservative for a 22nm device.

it is unrealistic and unreasonable to expect intel to give you a ton of cores and a ton of GHz at the same time. it does not economize TDP. more elbow room in the thermodynamics department awaits us as articulate graphene structures become feasible over the next ten years.