AMD ditching VLIW and going SIMD with Southern Island.

[Lonbjerg]

http://www.hardwareluxx.de/index.ph...pu-architektur-wechsel-von-vliw-auf-simd.html

Reason:
VLIW is to inefficient when it comes to other stuff than graphics (like eg. physics) and SIMD is better performing and better suited to develop for.

 

[dakU7]

bye bye bitcoin?

 

[RaistlinZ]

Maybe the street value of my 5870 will continue to climb.

 

[BTRY B 529th FA BN]

dang, just sold my 5870

 

[gorobei]

we now have 3 threads on this if you include the hd7000 big thread.

 

[Fox5]

Can we get an English article?

Also, SIMD and VLIW aren't mutually exclusive, where does it say they're dropping VLIW completely?

 

[notty22]

New AMD Slides Detail Upcoming Radeon HD 79** Series GPU Architecture

AMD has released two slides which details its next generation HD 79** series GPU architecture. The slides were revealed at the AMD’s Fusion developer summit where AMD also officially announced the availability of its Mainstream desktop lineup of Llano APU’s.
Looks like AMD would be ditching the VLIW instruction format and looking forward to move towards the SIMD wide-vector execution. You can check out the brief overview of the Slides along with the screenshots below:

Overview
* Ditch the VLIW instruction format and move to SIMD wide-vector execution
* L1 and L2 read / write caching
* Out of order resource allocation
* ECC data protection on SRAMs and global memory
* Parallel primitive

A few highlights from the talk: Multiple primitive pipelines for setup, etc.
Real caching in L1, L2, separate color / z caches for graphics and atomics
Concurrent tasks
Out of order resource allocation
ECC on srams and drams
Shaders
​

Source

 

[Nemesis 1]

Can we get an English article?

Also, SIMD and VLIW aren't mutually exclusive, where does it say they're dropping VLIW completely?

This is what MS C++ Amp is for Tunes I posted a short discription in another topic about this , I guess ya missed it

 

[Nemesis 1]

Who put AMP is for tunes in my reply. Now I know why when I post without a reread whats happening

 

[Nemesis 1]

http://www.hardwareluxx.de/index.ph...pu-architektur-wechsel-von-vliw-auf-simd.html

Reason:
VLIW is to inefficient when it comes to other stuff than graphics (like eg. physics) and SIMD is better performing and better suited to develop for.

SO I guess this ans. my questin on the VEXprefix

 

[Arachnotronic]

I really must say that by the end of 2011, AMD might just have the performance crowns in both CPU and GPUs.

 

[Nemesis 1]

I am willing to bet my new web browser on that

 

[Blitzvogel]

http://www.hardwareluxx.de/index.ph...pu-architektur-wechsel-von-vliw-auf-simd.html

Reason:
VLIW is to inefficient when it comes to other stuff than graphics (like eg. physics) and SIMD is better performing and better suited to develop for.

I thought I remember hearing VLIW was "the future" of graphics, but I guess that was a while ago, and I guess AMD is betting on non-graphics uses for the new architecture. What worries me is if AMD can keep hold of the performance/watt/transistor crown and do so without having to rely on separate clock domains to do it.

 

[badb0y]

http://www.hardwareluxx.de/index.ph...pu-architektur-wechsel-von-vliw-auf-simd.html

Reason:
VLIW is to inefficient when it comes to other stuff than graphics (like eg. physics) and SIMD is better performing and better suited to develop for.

I like your reasoning.

 

[RussianSensation]

VLIW is to inefficient when it comes to other stuff

I can't wait to use my inefficient 6950 to collect $ from Bitcoin so I can upgrade to the 28nm generation of cards on the cheap. Together with the $250 saved from not getting a GTX580, it might even be a "free" upgrade

I can't wait to see Kepler vs. HD7000 series. 2 brand new architectures (supposedly) on new 28nm tech.

 

[Lonbjerg]

I like your reasoning.

It's not my reasoning, it's AMD's.

 

[Silverforce11]

What is this Bitcoin stuff? I am so behind in the latest fad..

 

[Cerb]

I thought I remember hearing VLIW was "the future" of graphics, but I guess that was a while ago, and I guess AMD is betting on non-graphics uses for the new architecture. What worries me is if AMD can keep hold of the performance/watt/transistor crown and do so without having to rely on separate clock domains to do it.

See "COMPUTE UNIT ARCHITECTURE" on top of the block diagram slide. I imagine they are moving to scalar over VLIW for ease of extracting (including profiling and debugging) good real IPC, which should allow performance to improve in almost all cases, if they designed the hardware in step with its compiler(s) (make the HW simple and fast where the compiler can extract ILP, make the HW more robust where it can't). It still looks much less robust than nVidia, but I would expect that the situation of AMD's GPU performance falling flat compared to nVidia in <complicated program X> will begin coming to an end.

VLIW tends to be a speed demon or slow donkey for general purpose uses, without much middle ground--exactly what you typically find for current AMD GPUs used for parallel computing, where they'll wipe the floor w/ nVidia in one app, and be no faster than a decent CPU in another (exaggeration, maybe, but wild performance variations are typical, compared to nVidia). The more complicated your data is in memory, and the more conditional branches you need to use, the less likely VLIW is your best bet. I personally expected SMT or something like that to be added to VLIW, rather than them moving away from VLIW. TBH.

http://www.hardware.fr/news/11648/afds-architecture-futurs-gpus-amd.html
You may want to throw that through Google Translate, but it's got way more slides. I like what I see.

 

[Fox5]

What is this Bitcoin stuff? I am so behind in the latest fad..

Bitcoin is a fake currency that proponents refuse to believe is a pyramid scheme because it 'follows a mathematical formula' instead of some other arbitrary rules. People have made tons of money on it already, so it's hitting the news big time and now tons are rushing to get into it, and it's heading for disaster very soon.

It's a crypto-currency that gets harder and harder to produce with each coin produced. It is primarily used for trafficking in drugs and computer parts. It has a hard cap on the number of coins that can be produced (21 million iirc), and it is already half-way there. It won't be long now before those buying into it are just throwing their money away.

The creators insist it's not a scam, because those 21 million coins will always have value to the world as untraceable currency. Anyhow, some people have made big (fake) money, and perhaps some still can, but I wouldn't get in on it at this point. It's relevant to AMD's gpus, because they happen to be the best at crunching the algorithm to make new bitcoins.

 

[Ken g6]

As a programmer who's used both CUDA and AMD OpenCL, I'm very pleased with this development. Whenever you program for a GPU, you have to envision your code running on dozens or hundreds of processors in parallel. But to get maximum performance out of VLIW, you had to create vectors of four integers within that already parallel architecture! SIMD should eliminate this additional complication.

For those worried about AMD performance lowering, don't be. The biggest current nVIDIA GPU has 2x512 processors. AMD has some GPUs with thousands, albeit at half the clock speed. This should eventually make AMD more competitive for things like BitCoin...or PrimeGrid.

 

[RussianSensation]

What is this Bitcoin stuff? I am so behind in the latest fad..

Read Post #767 and #772: http://forums.anandtech.com/showthread.php?t=2167303&page=31

Set it up on your system. Both of your cards together should generate near 1 bitcoin per day (that equates to $13-15 USD at current rates).

Essentially, your GPU computes mathematics to check previous transactions for bitcoins in the virtual world (since there is no central authority or a 3rd party which checks them). For doing that, you get a reward - a bitcoin amount. Since bitcoin is not traceable, certain people think it has value as an online currency. As a result, there is some value to this virtual currency - but it changes on a daily basis.

Then you can just convert your GPU "mined" bitcoins into cash payout sent directly to your bank account OR as a gift card (i.e., Newegg, Amazon, etc.)

Let me know if you have any questions!

 

[Dark Shroud]

If you want to post open questions about Bitcoins please do so in the Bitcoin thread. You have a much better chance of getting help there and we don't want to spam other threads about Bitcoins.

 

[frozentundra123456]

I really must say that by the end of 2011, AMD might just have the performance crowns in both CPU and GPUs.

You're kidding right???

At best, they are competitive in the GPU area and totally behind intel on the CPU side. They have had delay after delay and disappointing parts when they finally did come out (cpus I mean). Maybe Bulldozer will finally be out by then, but I am skeptical that they can beat intel. And Llano may be adequate for laptops and have a place there, but for the desktop I see no place for it at all. I will admit that AMD makes some good low end chips, but I doubt that they will take the performance crown. I mean, a dual core I3 2100 competes with and sometimes beats an AMD mid level quad core Phenom II. I know bulldozer is supposedly coming out, but considering the Phenom launch I will be very surprised if it beats Sandy Bridge much less Ivy Bridge.

Maybe AMD is being forced to change direction on their GPU architecture towards GPU computing, but it is troubling to see them making a big change to the thing that they do best.

 

[formulav8]

You sure this re-work is for this years SI? When I read Anandtechs version it sounds like a new gpu design for the next coming years and not this year?

 

[RussianSensation]

You sure this re-work is for this years SI? When I read Anandtechs version it sounds like a new gpu design for the next coming years and not this year?

There has been no official source that confirmed HD7000 to be a GCN GPU. Given that AMD is switching to a brand new untested 28nm tech, do you think they'll also throw in a brand new architecture on it as well? Seems like too much risk to me. For this round, they could just increase the current specs of 6970 and be competitive. But who knows.