AMD reveals potent parallel processing breakthrough (Kaveri IGP will have ....GCN)

The Register ^ | 1st May 2013 | By Rik Myslewski in San Francisco3

[Ernest_at_the_Beach]

Upcoming Kaveri processor will drink from shared-memory Holy Grail

Update....Will require new FM2+ socket.

AMD has released details on its implementation of The Next Big Thing in processor evolution, and in the process has unleashed the TNBT of acronyms: the AMD APU (CPU+GPU) HSA hUMA.

Before your eyes glaze over and you click away from this page, know that if this scheme is widely adopted, it could be of great benefit to both processor performance and developer convenience – and to you.

Simply put, what AMD's heterogeneous Uniform Memory Access (hUMA) does is allow central processing units (CPUs) and graphics processing units (GPUs) – which AMD places on a single die in their accelerated processing units (APUs) – to seamlessly share the same memory in a heterogeneous system architecture (HSA). And that's a very big deal, indeed.

Why? Simple. CPUs are quite clever, speedy, and versatile when performing complex tasks with myriad branches, but are less well-suited for the massively parallel tasks at which GPUs excel. Unfortunately, they can't currently share the same data in memory.

[Ernest_at_the_Beach]

now for the little update on the socket.

From Xbitlabs:

AMD’s Next-Gen “Kaveri” APUs Will Require New Mainboards.

**********************************EXCERPT*************************************

AMD “Kaveri” APUs to Adopt FM2+ Form-Factor

[05/30/2013 11:21 PM] by Anton Shilov

It is not a secret that AMD’s next-generation code-named Kaveri accelerated processing unit for high-performance systems will be significantly different than the currently available chips in many ways. As it appears, the difference is so significant that the new hybrid microprocessor will require power supply circuitry, which means new FM2+ form-factor and appropriate mainboards. It will be impossible to install Kaveri into existing platforms.

AMD Fusion A-series APU code-named Kaveri is projected to feature up to four Steamroller high-performance x86 general-purpose cores as well as GCN architecture-based AMD Radeon HD graphics engine. The Kaveri chip will be AMD’s first high-performance accelerated processing unit with HSA [heterogeneous system architecture] enhancements, such as heterogeneous uniform memory access, CPU and GPU cache coherency and other. Moreover, Kaveri will be AMD’s first high-performance APUs made using 28nm process technology. Finally, the chips will feature new dual-channel memory controller as well as all-new power management.

As a result of massive changes inside Kaveri, the chips will feature new FM2+ form-factor and will only be compatible with new mainboards with FM2+ sockets, according to a new roadmap update issued by AMD. The new chips will therefore be incompatible with existing mainboards, which means that there will be no upgrade path for systems based on A-series “Trinity” and “Richland” APUs.

Given the fact that Kaveri will feature a number of “firsts” for AMD, a new socket was something logical to expect. What is unclear at this point is whether FM2+ mainboards will have any specific requirements and whether they will be more expensive than existing motherboards with FM2 sockets.

It should be noted that among other things Kaveri will reportedly support not only DDR3 memory, but also GDDR5 memory, which will be AMD’s secret weapon to fight for high-performance systems. GDDR5 should enable very high performance of integrated graphics sub-system for all-in-one and mobile systems.

Maximum memory bandwidth provided by dual-channel DDR3 memory sub-system at 2133MHz is 34.1GB/s, which should be sufficient for today’s microprocessors, but may be insufficient for powerful graphics adapters. By contrast, an inexpensive 128-bit GDDR5-based memory sub-system with 3400MHz effective clock-speed enables 54.4GB/s of memory bandwidth, or 60% higher compared to DDR3-based sub-system. While this clearly is not enough for a high-end graphics solution, it should be sufficient for modern mainstream graphics adapters that cost around $100 and which truly provide adequate performance in video games.

There are two problems with GDDR5 memory: it requires point-to-point interconnection and the maximum capacity of today’s GDDR5 chips is 2Gb (256MB). The former fact means that AMD A-series “Kaveri” APUs with GDDR5 memory will come in BGA package and will be soldered to mainboards, which will not be a problem for notebooks and all-in-one desktops, but will eliminate any possible upgrade options. The latter fact means that either AMD Kaveri-based system will be equipped with maximum of 4GB of GDDR5 (as 8GB of GDDR5, like on the PlayStation 4, requires 32 chips), which is not enough for modern general-purpose PCs, or AMD and its partners will have to wait till 4Gb GDDR5 chips arrive.

It is expected that AMD will release its Fusion A-series “Kaveri” APUs will be released in late 2013.

[rmlew]

Sounds like they are bringing thier playstation and xbox chips to general computing. Guess they had to do something. Intel is killing them everywhere.

[Ernest_at_the_Beach]

The Playstation 4 and Xbox One are based on a less powerful chip ...The Bobacat followon called Jaguar.

[Ernest_at_the_Beach]

Reading the comments at Xbitlabs to the article,...apparently both Hynix and Samsung are addressing the GDDR5 limitation mentioned in the article.

[GraceG]

I think eventually we will have desktop computers that will be able to emulate a scanned in human brain, maybe even a non-destructively scanned in human brain.... That would mean you could “scan” in a copy of yourself and have them do things online for you....

That gets scary....

Talk about using your desktop PC for mutitasking....

Then you could have that copy run a robot surrogate... so you could even mutitask in real life... could be handy for getting the housework done in a pinch..

The question is, what happens to your extra instances when you are don using them..... Would you have to merge their extra memories back into yourself, would deleteing it be murder if someone else does it... or suicide if you do it...

[Ernest_at_the_Beach]

Obvious question is will the new motherboard allow current FM2 processors to funcion correctly in the new FN2+ motherboards.

Not seen an answer to that.

[ShadowAce]

[quimby]

Many of the improvements on these newer chips are graphics oriented. The next Intel line, Haswell, has 40% more transistors (1.4 billion). The majority of these are graphics oriented.

[Ernest_at_the_Beach]

Eventually ....that might be a long time away ...

[Ernest_at_the_Beach]

That is the race going on between Intel and AMD...

Good consumers.

Of course,....one DOES NEED Software that takes advantage of the parallel execution units.

Old Windows software,..like Office will show very little improvement.,.....

[Zathras]

I’ve been worried for the last year that the semiconductor industry is being forced to move so fast, that they are not going to recover their development costs.

No one is going to want IVB once Haswell comes out next week.
Same with Bobcat vs Jaguar

Whats the point if no company is making $$$

[OneWingedShark]

Don’t worry about that; the human brain has more neuron-connections than there are stars in the galaxy.
http://theastronomist.fieldofscience.com/2011/07/cubic-millimeter-of-your-brain.html — Note this article is talking about a cubic millimeter.
https://community.ed.ted.com/teded/topics/are_there_more_neurons_nerve_cells_in_the_human_brain_or_more_stars_in_the_milky_way — See the first comment.

[OneWingedShark]

Many of the improvements on these newer chips are graphics oriented. The next Intel line, Haswell, has 40% more transistors (1.4 billion). The majority of these are graphics oriented.

I wonder if mayhaps they shouldn't try using those transistors they have more effectively -- I read about a processor called Objektive which set out to handle OOP in hardware. Then there's things like The GreenArrays GA144 chip (a gross of cores) each optimized stack-machines for Forth.

[ShadowAce]

one DOES NEED Software that takes advantage of the parallel execution units.

I really like my Fedora...

[ken in texas]

Sounds like Sequent’s NUMA on steroids with a healthy dose of Superman. :)

[GraceG]

Don’t worry about that; the human brain has more neuron-connections than there are stars in the galaxy.
http://theastronomist.fieldofscience.com/2011/07/cubic-millimeter-of-your-brain.html — Note this article is talking about a cubic millimeter.
https://community.ed.ted.com/teded/topics/are_there_more_neurons_nerve_cells_in_the_human_brain_or_more_stars_in_the_milky_way — See the first comment.

We could store the location information of every star in the galaxy (down to about 1m of resolution) 10 Kilobyte of info per star in most medium sized data centers today.

The problem is knowing the location of each star as our telescopes cannot scan the needed resolution/distance information. In that way the brain is a lot easier to map.

Also each neuron is made up of a lot of atoms, but their structure is relatively the same, it is a matter of what is connected there, to a mathematical model of a neuron my be simpler and there is a LOT of redundancy in the structure of the each dendrite, axion, etc... when speaking of data.

It would be like modeling a lawn, you only need a complex formula for the breed of grass and then for each addiotional blade you would only need some simple parameters like length of blade, width of blade and direction the blade is pointing and where it is anchored...

Granted that is still a LOT of data, but it may be possible to compress the model down. Time will tell.

If they can get to the point where they can model a cat’s brain, then a chimp brain won’t be long, and then a human brain won’t be far beyond that..

Could be 20 years away, could be 80 years, could be a 1,000 yearsm, could be never.... To do it we need to learn a LOT more about the human brain and make a lot more developments in computer science...

[GraceG]

Sounds like Sequent’s NUMA on steroids with a healthy dose of Superman. :)

NUMA!!!

[catnipman]

Sweet! Should help mine some new bitcoins!

[ken in texas]

It would be like modeling a lawn, you only need a complex formula for the breed of grass and then for each addiotional blade you would only need some simple parameters like length of blade, width of blade and direction the blade is pointing and where it is anchored...

Doing a lawn should be a walk in the park. Al Gore said there are computer systems that model the earth's atmosphere and predict the effects of mankind's harmful activities that result in Global Warming. :)