Adapteva Announces 700 MHz, 4096-Core Processor

[Tab54o]

If 1 core can fit inside a space of .1mm2 (.004") then they could easily fit 4096 inside 524.3mm2. Of course only with 28nm process. That is a ridiculously small core.

 

[eddieroolz]

4096 cores, now I thought that was not coming for another decade or so.

 

[hannibal]

Consumer market can utilise 1-4 cores at most... Normally one or two... So you would have an CPU that 2 cores are making the work and rest 4094 would be resting... If they really can make this (The Area51 is right in there!) it is useable in server and other special aplications.

 

[goodguy713]

the issue is people with 1990 pc's out there that are so dated pentium 2 and pentium 3 .. pentium 4's and the lame pretend games on facebook like angry birds have dumbed down the need to upgrade or purchase new computer systems that have advanced feature sets .. we should all be thankful that at least they started doing APU's otherwise we'd still have a tonne of lame ass computers with intergrated graphics like intel hd3000 lol or ati 1600 series .. that having been said i think a lot of these developers are more concerned about micro transactions as in games that cost 1- 10 dollars and have very minimal system requirements and are backwards compatible with tablets and smart phones .. when you have beasts like this in the back of the room its hard to argue with some one only looking for money vs speed and funtionality / quality... Personally i really think that someone should come up with a universal program that can automatically bypass software write protection and divide the processes evenly between multipule cores instead of being programmed for just one or two let the software provide efficent resourse utilization by sorting out the file work path loads and and assining different functions based on priority... where as each function is sorted out via irq request and prioritized on an in active core .. or really what im saying is make it really easy to multi thread applications in order to reduce time and cost to market If there was an actual consumer version of software like this ... i would buy it in a heart beat.. because having a 6 core or 8 core or even 64 core system would have a minimal effect unless you could proactively rewrite or reorganize software for your specific needs ..

 

[alxianthelast]

[citation][nom]wiyosaya[/nom]5.6 times ASCI RED performance and only 80W. I wonder how much power ASCI RED consumed?[/citation]

1.2 megawatts according to a quick google search. Which is surprising when I was guessing 2 megawatts.

 

[wolfram23]

Consumer market can utilise 1-4 cores at most... Normally one or two... So you would have an CPU that 2 cores are making the work and rest 4096 would be resting... If they really can make this (The Area51 is right in there!) it is useable in server and other special aplications.

What part of this did you read and think that it was for consumer market? Was it the part where they compare it to super computers, or where they talk about how amazing it will be for crunching numbers?

OT: That's pretty impressive if they get it working

 

[wolfram23]

the issue is people with 1990 pc's out there that are so dated pentium 2 and pentium 3 .. pentium 4's and the lame pretend games on facebook like angry birds have dumbed down the need to upgrade or purchase new computer systems that have advanced feature sets .. we should all be thankful that at least they started doing APU's otherwise we'd still have a tonne of lame ass computers with intergrated graphics like intel hd3000 lol or ati 1600 series .. that having been said i think a lot of these developers are more concerned about micro transactions as in games that cost 1- 10 dollars and have very minimal system requirements and are backwards compatible with tablets and smart phones .. when you have beasts like this in the back of the room its hard to argue with some one only looking for money vs speed and funtionality / quality... Personally i really think that someone should come up with a universal program that can automatically bypass software write protection and divide the processes evenly between multipule cores instead of being programmed for just one or two let the software provide efficent resourse utilization by sorting out the file work path loads and and assining different functions based on priority... where as each function is sorted out via irq request and prioritized on an in active core .. or really what im saying is make it really easy to multi thread applications in order to reduce time and cost to market If there was an actual consumer version of software like this ... i would buy it in a heart beat.. because having a 6 core or 8 core or even 64 core system would have a minimal effect unless you could proactively rewrite or reorganize software for your specific needs ..

What? What the heck does this article have to do with people running old PCs? Or consumer programs and multithreading? Or cheap games???

People will write programs that take advantage of all the cores of this PC, they're called scientists and they work at universities in the Computing Sciences departments and physics labs (to name a couple).

Also, more cores does not necessarily mean faster even if everything is multithreaded. Depends on what needs to be computed, for example PhysX (by Nvidia) would actually run significantly better on a quad core CPU than a 512 core GPU.

 

[f-14]

[citation][nom]Bones2525[/nom]Yes, but can it run CRYSIS.[/citation]
can you read minimum system requirements for said software.
[citation][nom]jacekring[/nom]Stop with that already...the iPad 3 can probably run Crysis. Crysis is old and dated now...can it run BF3 on max settings is a better question?[/citation]
no, again, you can not read minimum system requirements, maybe if steve jobs coverts it to super mario brothers gameboy style you can play it on icrapple products, but until then that is the biggest negatory on the planet until apple throws in the towel and converts to windows OS like it was forced to do with hardware just to stay relevant.

 

[jaybus]

Supercomputers have absolutely unbelievable energy costs. Reducing energy consumption is one of the top priorities for people designing them.

Much of that energy cost, though, is in cooling the computer room. The cooling system is, at best, 50% efficient, meaning it takes at least twice as much electrical power to cool the computer room as it does to power the computer nodes. To cool a computer that dissipates 200 kW requires a cooling system with at least a 400 kW (1.3 million BTU) rating, which is about 30 typical home central A/C units.

Let's be honest. At least for supercomputers, it's all about scaling up performance to maximize the available cooling capacity. It's not about being green. It's that we are already near max allowable energy consumption due to waste heat and need lower TDP processors in order to continue to make performance gains.

 

[vittau]

[citation][nom]Marco925[/nom]Imagine that task manager!!!! (that is if it was x86 based)[/citation]
It would use all it's processing power just to render it.

 

[Guest]

I could see these generating a pretty significant amount of heat once you reach the 256+ cores.

 

[IndignantSkeptic]

... Personally i really think that someone should come up with a universal program that can automatically bypass software write protection and divide the processes evenly between multipule cores instead of being programmed for just one or two let the software provide efficent resourse utilization by sorting out the file work path loads and and assining different functions based on priority... where as each function is sorted out via irq request and prioritized on an in active core .. or really what im saying is make it really easy to multi thread applications in order to reduce time and cost to market If there was an actual consumer version of software like this ... i would buy it in a heart beat.. because having a 6 core or 8 core or even 64 core system would have a minimal effect unless you could proactively rewrite or reorganize software for your specific needs ..

I'm not very knowledgeable about things like this but I think maybe you'll be excited by OpenACC.

 

[bit_user]

There won't be many workloads that won't run better on a GPU or a MIC chip with fewer, faster cores.

The blurb doesn't say how they're going to get data on & off chip, which is already a huge bottleneck for GPUs. AMD's HD 7970 might get only 16.5 GFLOPS/W, but I'll bet it has a lot more memory bandwidth than this thing. Moreover, it's nearly as fast (3.8 TFLOPS) and it's here today (and it will run Crysis!).

 

[bit_user]

[citation][nom]bit_user[/nom]AMD's HD 7970 might get only 16.5 GFLOPS/W[/citation]
I forgot that this figure is for the entire board: GDDR5, PCIe 3 x16, fan, and display PHYs. By the time Adapteva adds memory and some kind of I/O, they're not going to do much better.

Yes, there are algorithms which aren't SIMD-friendly, but this thing won't exactly be a piece of cake to program, either. I don't see why anyone would take the risk of developing custom software (and with only 32k of memory per core, it will have to be very customized) for a chip from a company that could just disappear over night.

Between GPUs and Intel's Knight's Corner, there just won't be a big enough market for this thing. I'm predicting they'll die long before they even tape out the 4096 core chip. I'd even put money on the 1024 core chip never seeing the light of day.

 

[goodguy713]

[citation][nom]f-14[/nom]can you read minimum system requirements for said software.
no, again, you can not read minimum system requirements, maybe if steve jobs coverts it to super mario brothers gameboy style you can play it on icrapple products, but until then that is the biggest negatory on the planet until apple throws in the towel and converts to windows OS like it was forced to do with hardware just to stay relevant.[/citation]

i was trying to capture the fact that there are still a lot of dinosaur pcs out there that dumb down system requirements in order to play / run modern software .. and on the high end spectrum why dumb down what is classified as consumer use... personally i dont really care for the term because it seems like they know whats best for us and were supposed to flock like sheep to what ever they classify as consumer / enthusiast friendly why not have a 64 core system that runs crisis and bf3 .. my point being is while it might be over kill.. but i mean wouldnt it be nice to be able to take a game like that and have it utilize all 64 cores vs 4 cores .. makes for an interesting concept .. in my perfect world every one has 4 ssd drives in raid and intel extream editions.. 32 gbs of ram and an arm and a leg twin corsair h100 water blocks w/ custom tubbing .. lol sick over kill waist yes.. but it sounds like a 7970 quad fire dream.. lol

 

[Pherule]

Title is misleading. How can they announce a 4096-Core processor that they haven't built or tested yet? It should be "Adapteva Announces designs for / the planning of a 700 MHz, 4096-Core Processor."

Still want to see Intel doing something like this for the desktop end. Reduce core count to 512 but increase frequency to 2-3Ghz, then you have a computer than can do some nice floating point calculations whilst still being capable of current desktop applications.

 

[san_lex]

~80 watt TDP - nice!

 

[sonofliberty08]

intel said it will be 10GHz Pentium CPU too, but......

 

[robisinho]

The amd 7990 should get somewhere in the neighborhood of ~7000 max GFLOPs. These chips in parallel should match that in only 100w. A single (4096 core!) chip does 5.6k GFLOPs.

 

[bit_user]

[citation][nom]Pherule[/nom]Still want to see Intel doing something like this for the desktop end. Reduce core count to 512 but increase frequency to 2-3Ghz, then you have a computer than can do some nice floating point calculations whilst still being capable of current desktop applications.[/citation]Intel is (slowly) getting there.

Remember Larabee? It lives on and is now in its 3rd generation, called Knight's Corner. They scaled back their plans, focused on parallel computing research and are now about to offer a 64-core processor for server & HPC applications. If that works out well and they make progress with technologies, like transactional memory, that make the parallelism easier for software to exploit, we may see these so-called MIC architectures brought to the desktop.

Until then, we'll have to be content with GPUs from AMD and NVidia. Their cores are looking increasingly like those of general purpose CPUs, with each new architecture. While they're not the easiest things to program, they're vastly more mature and balanced than what any small company can produce in a first-gen parallel computing platform.

 

[bit_user]

[citation][nom]robisinho[/nom]The amd 7990 should get somewhere in the neighborhood of ~7000 max GFLOPs. These chips in parallel should match that in only 100w. A single (4096 core!) chip does 5.6k GFLOPs.[/citation]Assuming Adapteva can get working silicon at acceptable pricing and before their performance is completely obsolete, and assuming they can get customers to take the chance on integrating them into designs and writing the custom software that will be needed to use them, then you still have to consider the issues I cited about board vs. chip power and memory bandwidth.

I've read that the GDDR5 memory controller was one of the larger, hotter, and more challenging blocks on both AMD's and NVidia's GPUs. If Adapteva is going to support it, it will probably blow their power scaling. If not, it will certainly blow their performance scaling.

I've seen this story time and again: small company promises revolutionary architecture, then you hear nothing and eventually find out they went bust because they underestimated the technical and business challenges needed to go from an interesting concept to a successful product.

When comparing vs. GPUs, you need to distinguish between vaporware and real products that one can buy in a local electronics store. Even if this thing does become a reality, GPUs will be on at least their next architecture by then - faster, more power-efficient, and maybe even a bit easier to program.

BTW, the 3rd competitor it has are FPGAs. For tasks involving low memory bandwidth requirements, needed by customers willing to invest a lot in development, they offer even more speed and efficiency.

 

[freggo]

[citation][nom]hannibal[/nom]Consumer market can utilise 1-4 cores at most... Normally one or two... So you would have an CPU that 2 cores are making the work and rest 4094 would be resting... If they really can make this (The Area51 is right in there!) it is useable in server and other special aplications.[/citation]

Hardly. As a former game programmer I can assure you that if you make a n-core CPU at least a dozen programmers will come up with an application that will run smoothly using a minimum of 2n-cores