[Beowulf] [FoRK] (Pimpin' for) Parallella: A Supercomputer For Everyone by Adapteva

Vincent Diepeveen diep at xs4all.nl
Thu Oct 25 11:23:15 PDT 2012



The factories to produce 28 nm technology have a cost of what is it  
10 billion dollar a factory or so?
Make it afew bilion dollar less.

So what they produce in that factory must pay back for those billions  
it costs to build that factory and
the operational costs of the factory.

The only exception intel for example makes at new factories is for a  
FPGA manufacturer to produce first.
After that intel is using exclusively for mass produced cpu's the new  

How do you want to produce at 28nm a chippie that's 70 gflops a watt,  
though without multiplication unit.

Maybe if you got 75 million dollar you can produce something. Means  
it needs to pay back for 75 million dollar
in production costs first.

Are you getting the point at all?

We're speaking about probably 28 nm TSMC here now. Same technology  
currently getting used to produce GPU's.
Startup costs maybe 5 million dollar?

To produce 1000 cpu's there, you first need a design though.

Writing a program in C code doesn't mean you can print a CPU from it  
in hardware.

There is different hardware languages and for each new technology  
step you need
to produce new design.

So for 28 nm they would need a new design that's DIFFERENT from what  
they have now.

Are you getting point B?


Then the next question is: "how high does this design clock?"

Now an example demonstration design in fpga, that fpga can clock to  
1.5Ghz or so.
Yet if you design your own chip getting above 300Mhz is quite difficult.

You are slowly getting the picture?

This is not like car technology. There is a reason why most clusters  
use massively produced CPU's you know.

It's really expensive to produce CPU's.

If you sell millions of them then, you can produce them cheap.

Now what i understand from the video is that he isn't promising in a  
$100 package a 64 core chip of himself.
He's just promising a SOC that you can plug in to your TV somehow and  
then play video's and browse the web.

That's all.

So i miss the promise of a 64 core or even more core chip that you  
can get at $100 including a SOC in the video at all.

 From how i read their homepage the only thing they got now is a 16  
core CPU design.

This is going to be an expensive 64 core chip if he's gonna produce  
it real low volumes...

If he'd promise a soc with 100k cores or so for small money, then  
things start to get interesting ok,
but only if it's dirt cheap.

The 'dirt cheap' promise is missing everywhere, except for browsing  
the internet.
You don't need a 64 core chip to browse the internet.

On Oct 25, 2012, at 7:59 PM, Tim Mattox wrote:

> Vincent,
> They already have the 28nm design taped out from Global Foundries.
> The short intro video on the http://www.adapteva.com/ front page  
> has that.
> Where are you getting that they would need to do a whole new chip  
> design?
> This is simply a matter of the cost to scale up production beyond the
> samples that they are getting and have already gotten from GF.
> Quoting from the kickstarter page:
> "
> Chip foundry retooling changes will reduce silicon costs to a few
> dollars per chip.
> These changes are very expensive and account for a large part of  
> the funds
> needed to produce a low cost Parallella computer.  Our 16-core  
> Epiphany chips
> have been in the field for over a year and have been tested  
> thoroughly by many
> hardcore developers. Chip product retooling for the sake of cost  
> reduction is
> usually referred to as a "full mask tapeout" and should be  
> considered a low risk
> part of the project. Our current low volume chip manufacturing flow  
> only
> yields 50 dies per wafer.  By creating full production mask sets  
> for our chips
> we will be able to yield 1000's of dies per wafer.
> "
> They do need to design a new PCB for the $99 version, but that's  
> not hard,
> and this is all disclosed on the kickstarter page.
> So, if your problem with their approach is really just about having  
> the CEO's
> daughter in the video... get over it.  I thought that was cute.
> For benchmarks, they reference an industry standard embedded  
> benchmark:
> http://www.adapteva.com/white-papers/more-evidence-that-the- 
> epiphany-multicore-processor-is-a-proper-cpu/
> The 16 core chip isn't going to set any speed records, but as a  
> learning
> tool for parallel processing, it would be great to have in the  
> classroom.
> -- Tim Mattox
> On Thu, Oct 25, 2012 at 1:21 PM, Vincent Diepeveen <diep at xs4all.nl>  
> wrote:
>> Tim,
>> Producing this at 28nm requires a total new design.
>> Every single transistor.
>> And if they wrote it in verilog or something, which is typical
>> language used for FPGA,
>> they have yet another problem... ...of learning a new language.
>> In the video the guy promises basically a SOC with a dual core ARM
>> that you can use to play video's.
>> So that there is a chip of him on that SOC isn't explicitly
>> mentionned nor how many cores it will have
>> nor what it can do nor enhance.
>> I'm not seeing any benchmarks on their website of existing software,
>> say a simple chessprogram that isn't using the instructions the chip
>> doesn't have.
>> So this is a pipe dream for now.
>> The real problem i have with the video. He uses his small daughter -
>> i'm very much against this.
>> Then he basically promises his audience a 'computer' that you can
>> browse onto and play video's.
>> That's all.
>> So that means just decoding a video. We have already hardware for
>> that on such SOC's.
>> So what he promises you simply don't need any chip from this company
>> for.
>> Then if you go to the website, there is an epiphany CPU line, which
>> ofcourse will never make it into 28 nm.
>> Starting up the machines already has a cost of what is it. $5  
>> million?
>> That's just to *start* them.
>> On Oct 25, 2012, at 7:11 PM, Tim Mattox wrote:
>>> Vincent,
>>> They already have working silicon for both the 16 and 64 designs.
>>> It's not a fairy tale.
>>> What they don't have is the full set of photomasks for mass
>>> production.  That
>>> is what the primary expense is that they are trying to fund thru
>>> kickstarter.
>>> It is a little out of date, but this wikipedia article explains  
>>> it...
>>> http://en.wikipedia.org/wiki/Photomask
>>> Also, the epiphany chips are tiny, and thus they would get very high
>>> yield, vs. a big traditional CPU.
>>> -- Tim Mattox
>>> On Thu, Oct 25, 2012 at 1:01 PM, Vincent Diepeveen <diep at xs4all.nl>
>>> wrote:
>>>> Provided that the fairy tale table is true at:
>>>> http://www.adapteva.com/products/epiphany-ip/epiphany-architecture-
>>>> ip/
>>>> under "example configurations'.
>>>> Then secondly the cpu is supposed to be 28 nm.
>>>> How's he gonna produce that with just 750k dollar?
>>>> Having 75 million dollar maybe he can just start produce cpu's.
>>>> Compare with AMD. They're still producing at 32 nm. That's their
>>>> latest bulldozer CPU ok.
>>>> Intel moved on to 22 nm now and hardly has production at it under
>>>> control.
>>>> This is billion dollar companies.
>>>> You know there is many things in life that with a more clever  
>>>> design
>>>> and better software you can
>>>> do better. Yet producing CPU's is a very specialistic type of area
>>>> where just to start a x-billion dollar
>>>> factory produce for you some CPU's.
>>>> CPU's tend to get cheap if you produce big masses of them.
>>>> Provided the fairy tale table is true - to get this entire board +
>>>> connectors + CPU sold for under $100,
>>>> he needs to produce massive amounts of those CPU's.
>>>> Having 64 cores @ 800Mhz it's effectively from my viewpoint seen,
>>>> assuming a 1 instruction a cycle.
>>>> That's like compared to the IPC of a core2 Xeon i have here:
>>>> Say he gets IPC 0.25 on the cpu's.
>>>> At the core2 i get nearly 1.7. Actually more like 1.5x at core2,  
>>>> yet
>>>> i intend to improve that.
>>>> Even at 1.5 that's factor 6 difference in IPC.
>>>> Then it's 700Mhz, see the table, elsewhere he claims 800Mhz.
>>>> I don't see how he will get it to 700Mhz to be honest. Maybe he'll
>>>> get it to 300Mhz in 28 nm.
>>>> If i speak to some hardware architects - they don't believe  
>>>> it'll run
>>>> @ 700Mhz that easily. It requires a LOT of people
>>>> to optimize it to clock it to 700Mhz in those complicated nanometer
>>>> technology. This is GPU architects ok.
>>>> Also simple manycores.
>>>> The only 'advantage' he has with his epiphany chip is that it isn't
>>>> having complicated instructions like multiplication nor division.
>>>> That might make it a tad easier.
>>>> Let's say he manages. So the 2.5ghz core2 Xeons here.
>>>> That's factor 4 faster nearly.
>>>> So we're at already factor 24 difference in speed.
>>>> 64 cores / 24 = 3
>>>> So 3 Xeon cores @ 2.5ghz is roughly equivalent for non-multiplying
>>>> operations to a single 64 core chip of this @ 28 nm.
>>>> How is he gonna produce it with 750k dollar at 28 nm at an  
>>>> affordable
>>>> price. Of  course the wattage claimed i don't believe.
>>>> No one does, but that's not so relevant for now.
>>>> Claiming x-gflops is not so interesting if you cannot multiply -  
>>>> and
>>>> his chip cannot.
>>>> It can run chess software however in theory. Simulations that do  
>>>> not
>>>> require multiplications. In theory.
>>>> Now suppose he gets there after all. A $100 board that's  
>>>> relative low
>>>> power, say eating a watt or 25. With 64 cores
>>>> that effectively is same speed like 3 core Xeons i got here  
>>>> (L5420).
>>>> Of course that's useful hardware for in the living room.
>>>> Those Xeons are on ebay $15 now or so?
>>>> Entire node is under $150. If you buy in dirt cheap you might  
>>>> manage
>>>> $100 maybe.
>>>> The problem of CPU's like this is to produce them dirt cheap
>>>> actually.
>>>> Yet if we look to more useful ARM cpu's and MIPS cpu's that have  
>>>> many
>>>> cores - no mobile phone is using them.
>>>> So mass producing this, which would get the price down - that's for
>>>> now a pipe dream.
>>>> They wouldn't be able to use this chip of course as it can't
>>>> multiply.
>>>> Now for me that's not an issue, yet price versus performance is.
>>>> The low power story no one believes of course.
>>>> You know if he would want to built a new stealth boat for $750k
>>>> that's more real than produce a 64 core CPU cheap,
>>>> as just the production line of a cheap 64 core cpu means you gotta
>>>> sell millions of them to get it to $100 a SOC.
>>>> On Oct 25, 2012, at 5:23 PM, Eugen Leitl wrote:
>>>>> ----- Forwarded message from Jeff Bone <jbone at place.org> -----
>>>>> From: Jeff Bone <jbone at place.org>
>>>>> Date: Thu, 25 Oct 2012 10:09:55 -0500
>>>>> To: fork at xent.com
>>>>> Subject: [FoRK] (Pimpin' for) Parallella: A Supercomputer For
>>>>> Everyone by
>>>>>       Adapteva
>>>>> X-Mailer: iPhone Mail (10A403)
>>>>> Reply-To: Friends of Rohit Khare <fork at xent.com>
>>>>> Ok folks.  This looks to be the Arduino or RaspPi for parallel
>>>>> computing.  I'm psyched about it and I'm one of the unclosed
>>>>> "potential big backers" mentioned.  This isn't going to replace
>>>>> G*PU in the data center but this gets cheap exposure to the
>>>>> masses.  I think this thing has some real potential in the
>>>>> hobbyist / maker / hardware hacker / pedagogical space.
>>>>> Be nice to have a much more open (like this) solution here, in  
>>>>> this
>>>>> space.
>>>>> Final push here.  If they get within striking distance I'll
>>>>> probably step in to close it out.
>>>>> $0.02 (well, really, a whole lot more.)
>>>>> jb
>>>>> Begin forwarded message:
>>>>>> From: Kickstarter <no-reply at kickstarter.com>
>>>>>> Date: October 25, 2012, 10:00:08 CDT
>>>>>> To: jbone at mgrep.com
>>>>>> Subject: Project Update #18: Parallella: A Supercomputer For
>>>>>> Everyone by Adapteva
>>>>>> Reply-To: No Reply <no-reply at kickstarter.com>
>>>>>> Project Update #18: We REALLY need your help in spreading the  
>>>>>> word!
>>>>>> Posted by Adapteva
>>>>>> We really wish we could have been this ready from day one of the
>>>>>> Kickstarter, but we couldn’t wait with the launch. Time has  
>>>>>> been
>>>>>> our enemy from day one at Adapteva. After the project is  
>>>>>> funded we
>>>>>> will tell you about all the details. It’s kind of an  
>>>>>> interesting
>>>>>> story.
>>>>>> We are now working the press hard from our end and have some
>>>>>> potential big backers that we are will try very hard to close  
>>>>>> with
>>>>>> today and tomorrow.
>>>>>> There is still hope, but we REALLY could everyone’s help in
>>>>>> blanketing the earth with Parallella news over the next two days.
>>>>>> We will try to make it as easy as possible, so here are some  
>>>>>> ideas/
>>>>>> tweets/posts suggestions to start with. (most of them inspired by
>>>>>> backers!)
>>>>>> ----------------------------------------------------------------- 
>>>>>> --
>>>>>> --
>>>>>> -----------------------
>>>>>> Email/Facebook/Ideas:
>>>>>> The project is STILL fundamentally about parallel programming but
>>>>>> we have now shown that we have a ready prototype of a REAL  
>>>>>> working
>>>>>> computer. At $100, maybe some of your non-techie friends would
>>>>>> line one, so anyone who feels comfortable spreading the word
>>>>>> through facebook/email, please do.
>>>>>> Maybe it would be a great computer for your kids to hack around
>>>>>> with? Who knows, maybe it would even inspire a life-long passion
>>>>>> for programming and electronics? Maybe get one for a non-techie
>>>>>> brother, sister, parent and help them set it up in their living
>>>>>> room? I am sure they would love you even more.:-)
>>>>>> http://www.kickstarter.com/projects/adapteva/parallella-a-
>>>>>> supercomputer-for-everyone
>>>>>> Thank you!!
>>>>>> ----------------------------------------------------------------- 
>>>>>> --
>>>>>> --
>>>>>> ----------------------------------------------
>>>>>> Tweet Suggestions:
>>>>>> A $100 Zynq based computer, are you kidding me? http://kck.st/
>>>>>> UGQjG3
>>>>>> An ARM based 5 Watt computer that runs Ubuntu Desktop for under
>>>>>> $100. http://kck.st/UGQjG3
>>>>>> Parallella a smart(er)TV for $100? http://kck.st/UGQjG3
>>>>>> Parallella: The first truly open OpenCL hardware platform.  
>>>>>> http://
>>>>>> kck.st/UGQjG3
>>>>>> Parallella: 18 CPUs at 800MHz, 25 GFLOPS, 1GB DRAM, $100. Sweet!
>>>>>> http://kck.st/UGQjG3
>>>>>> Parallella a hackable platform. http://kck.st/UGQjG3
>>>>>> The $100 credit card sized Parallella combines FPGA, ARM, and
>>>>>> DSPs. http://kck.st/UGQjG3
>>>>>> Parallella runs OpenCV: http://kck.st/UGQjG3
>>>>>> Parallella will be great for SDR: http://kck.st/UGQjG3
>>>>>> At $100 wouldn’t the open Parallella be a great education
>>>>>> platform? http://kck.st/UGQjG3
>>>>>> What could Parallella do for DIY drones? http://kck.st/UGQjG3
>>>>>> What could Parallella do for robotics? http://kck.st/UGQjG3
>>>>>> Parallella, the ultimate “thin client”. http://kck.st/UGQjG3
>>>>>> Parallella, a computer your kids can carry around in an Altoid  
>>>>>> tin
>>>>>> can. http://kck.st/UGQjG3
>>>>>> ----------------------------------------------------------------- 
>>>>>> --
>>>>>> --
>>>>>> --------------------------------------
>>>>>> News Site submission suggestions:
>>>>>> (Pick any of the tag lines from the Tweet section or make up your
>>>>>> own)
>>>>>> http://hackaday.com/contact-hack-a-day/
>>>>>> http://news.ycombinator.com/submit
>>>>>> http://www.reddit.com/
>>>>>> tipbox at gizmodo.com
>>>>>> tips at techcrunch.com
>>>>>> www.engadget.com (tip us link at the bottom right corner)
>>>>>> www.drdobbs.com
>>>>>> www.stackoverflow.com
>>>>>> ----------------------------------------------------------------- 
>>>>>> --
>>>>>> --
>>>>>> ------------------------
>>>>>> Forum/Community Posts:
>>>>>> If you are an active member of a forum, please consider posting a
>>>>>> question that would start interesting dialogue about Parallella.
>>>>>> Please do NOT spam, but if you think your community would
>>>>>> genuinely benefit from the Parallella platform, the please post.
>>>>>> We are REALLY interested in finding out what these communities
>>>>>> think about the project. If you are a developer and you can speak
>>>>>> with confidence that you will be able to complete a project on  
>>>>>> the
>>>>>> Parallella, even better!
>>>>>> http://xbmc.org/
>>>>>> http://slashdot.org/
>>>>>> http://opencv.org/
>>>>>> http://www.khronos.org/opencl/
>>>>>> https://bitcointalk.org/
>>>>>> http://www.minecraftforum.net/
>>>>>> http://www.emulator-zone.com/
>>>>>> https://meego.com/
>>>>>> http://www.raspberrypi.org/ (note: suggestion for collaboration
>>>>>> only, not competition!)
>>>>>> http://beagleboard.org/discuss (note: suggestion for  
>>>>>> collaboration
>>>>>> only, not competition!)
>>>>>> http://www.adafruit.com/contact/
>>>>>> http://123kinect.com/kinect-forums/
>>>>>> http://forum.xda-developers.com/
>>>>>> http://setiathome.berkeley.edu/forum_index.php
>>>>>> http://foldingforum.org/
>>>>>> Care to comment? View this update on Kickstarter →
>>>>>> Unsubscribe from this project's updates with one click  
>>>>> _______________________________________________
>>>>> FoRK mailing list
>>>>> http://xent.com/mailman/listinfo/fork
>>>>> ----- End forwarded message -----
>>>>> --
>>>>> Eugen* Leitl <a href="http://leitl.org">leitl</a> http://leitl.org
>>>>> ______________________________________________________________
>>>>> ICBM: 48.07100, 11.36820 http://www.ativel.com http:// 
>>>>> postbiota.org
>>>>> 8B29F6BE: 099D 78BA 2FD3 B014 B08A  7779 75B0 2443 8B29 F6BE
>>>>> _______________________________________________
>>>>> Beowulf mailing list, Beowulf at beowulf.org sponsored by Penguin
>>>>> Computing
>>>>> To change your subscription (digest mode or unsubscribe) visit
>>>>> http://www.beowulf.org/mailman/listinfo/beowulf
>>>> _______________________________________________
>>>> Beowulf mailing list, Beowulf at beowulf.org sponsored by Penguin
>>>> Computing
>>>> To change your subscription (digest mode or unsubscribe) visit
>>>> http://www.beowulf.org/mailman/listinfo/beowulf
>>> --
>>> Tim Mattox, Ph.D. - tmattox at gmail.com
>> _______________________________________________
>> Beowulf mailing list, Beowulf at beowulf.org sponsored by Penguin  
>> Computing
>> To change your subscription (digest mode or unsubscribe) visit  
>> http://www.beowulf.org/mailman/listinfo/beowulf
> -- 
> Tim Mattox, Ph.D. - tmattox at gmail.com

More information about the Beowulf mailing list