Hard Hardware Lessons Learned from the Parallella Kickstarter

[eas]

An article in EETimes last month by Andreas Olofsson on his experience launching the Parallella on Kickstarter in 2012.

When someone asks me if I would ever do a Kickstarter project again, my answer is always “No way!” In terms of technology democratization, the Parallella Kickstarter project was a huge success, but it was a financial disaster for Adapteva.

Not having lived through the Parallella experience, I'd have a somewhat different take, which is that, Kickstarter avoids or mitigates some project risks, but every project has risks, and Kickstarter brings some of its own.

[Balaur]

The article heavily insists on Kickstarter as the root cause of all evils (I'm exaggerating of course)
IMHO, the case discussed there is just a poor business plan.

And don't get me started on the many-cores companies!

[janekm]

Well, yeah. If you set your price at $100, and then only after the campaign closes calculate that your BOM cost is $129, and your overall COGS is $179 (!), it's a bit odd to blame Kickstarter...

Not to mention, it's all well and good wanting to keep your manufacturing in the US (and there's obviously cost, efficiency and hassle benefits in many cases), but if you're making $90000 worth of product it's not going to be cheaper...

[EEVblog]

He was on the Amp Hour talking about it:
http://www.theamphour.com/254-an-interview-with-andreas-olofsson-adatevas-ampliative-abacus/

[aolofsson]

Article FAQ:

http://www.adapteva.com/andreas-blog/kickstarter-hardware-economics/

[marshallh]

Kinda seems like Andreas was so knee-deep in taping out the silicon he did not develop a good intuitive feel for the way kickstarters generally work. Most of that stuff falls under "Unsurprisingly".
He took a lot of ballsy moves  (or maybe naive?) and did end up making it work, such as the 99 price target that required subsidization.

While Epiphany seems to be a very poweful chip for a very niche application, it won't succeed as an Arduino or C64/BBC micro in every home, turkey in every pot sort of thing.
The maker crowd that will use that chip is going to be basically immaterial, it needs to find a killer app shipping in something. Not a cell phone, that's pretty dumb. But I could see it in some sort of new base station rollout, or general SDR/analytics application.

[Mechatrommer]

i downloaded and read the epiphany sdk ref but it seems like linux development only, no ref for wingows? they are doomed. and quick read the architecture its like running 16 "INDEPENDENT" apps we have to create separate main0.c - main15.c. thats a pain in the arse sychronizing a single multithreaded application in one single project file. no? imho refinement should be made for multihreaded api in this matter, like create_thread, enter_critical_section etc...

[IanB]

i downloaded and read the epiphany sdk ref but it seems like linux development only, no ref for wingows? they are doomed. and quick read the architecture its like running 16 "INDEPENDENT" apps we have to create separate main0.c - main15.c. thats a pain in the arse sychronizing a single multithreaded application in one single project file. no? imho refinement should be made for multihreaded api in this matter, like create_thread, enter_critical_section etc...

Without knowing any details of the hardware, it seems from your comments that one would use MPI rather than programming from scratch.

[Mechatrommer]

it does have a way to synchronize though using mutex and this "barrier" method. i'm not familiar with the architecture and the MPI standard, i'll let others to make the venture. less example and linux only is a no go for me this far. the spec is tempting but the architecture is so so, so far, or maybe i'm just that ignorant...  its now on sale, they say starting price $99 but in amazon its $149... for anyone interested...
http://www.parallella.org/board/
http://www.amazon.com/Adapteva-Parallella-16-Desktop-Computer/dp/B0091UD6TM

[helius]

No, that's not correct. The Adapteva architecture uses a shared address space, so there is no way to run "independent apps" on it at all, much less a requirement to do so. The 16 cores are independent computers in the sense that they contain their own memory, but the memory of all the other cores is found in the same address space and programs don't really need to know where their data is.

MPI is used for loosely-coupled machines like clusters where different nodes cannot access each others' memory. It will also work on shared-memory machines of course, but it simply makes less sense.

[donotdespisethesnake]

Wow, the author sounds so confused it is not surprising the KS went awry. Or did it? He goes on to say he secured $3.5 million funding as a result - so that is success isn't it?

He says the main message is to factor in your time and other engineering costs (surely goes without saying you would think), but then says the price point was chosen at $99 to get sales, and if a realistic price was chosen, it would have also failed.

He recognises that a KS project can be a loss-leader as a route to conventional funding, or a way to modest funding for organic growth. It can't be both, should really be the main message. If you are doing the loss-leader project, you can cap the rewards to control how much money you will lose.

If you lose on the unit price, you CANNOT make it on the volume!

[Howardlong]

BoM=$125 (just parts, nothing else, no assembly, no testing, nothing, zip nada)

Retail price=$99.

That's all you need to know.

In my experience for that kind of volume (~10ku) and retail price point (~$100), to make a reasonable living in the first world you need a total unit cost absolute maximum 50% of retail cost. Your retail/distribution chain will take 30% of the retail price, leaving you 20% to both live on and invest in new product.

It's a business, not a charity, if the numbers are a non starter, then do something else. Selling 30ku for $3m when it cost you $5m to make them is not a success in anyone's book.

[Mechatrommer]

No, that's not correct. The Adapteva architecture uses a shared address space, so there is no way to run "independent apps" on it at all, much less a requirement to do so. The 16 cores are independent computers in the sense that they contain their own memory, but the memory of all the other cores is found in the same address space and programs don't really need to know where their data is.
MPI is used for loosely-coupled machines like clusters where different nodes cannot access each others' memory. It will also work on shared-memory machines of course, but it simply makes less sense.

yes adapteva uses shared memory architecture, but their multithreading API is somewhat short or at very young age imho, if you read the ref guide you only have 2 functions for multithread shared memory synchronization, mutex and barrier. and there is diagram showing main0.c - main15.c run in each core, no example yet on how to build one single multithread application and how to build elf in windows using windows toolchain.

BoM=$125 (just parts, nothing else, no assembly, no testing, nothing, zip nada)
Retail price=$99.
That's all you need to know.

the author in his link above stated, the loss is due to NRE cost even though they met their less than $100 BOM (COGS). i guess now NRE cost is covered since parallella is on sale now... and they are running production batch now at COGS cost only, well... maybe they need to increase a little bit to cover production cost or whatever non COGS cost. its time to collect what they have lost... they should concentrate on support and marketing department i hope they have enough cost/resources for that too... esp time to refine they documentation, add more practical app example as i mentioned above etc... that alone can help in marketing imho, i hope they success..

[Howardlong]

BoM=$125 (just parts, nothing else, no assembly, no testing, nothing, zip nada)
Retail price=$99.
That's all you need to know.

the author in his link above stated, the loss is due to NRE cost even though they met their less than $100 BOM (COGS). i guess now NRE cost is covered since parallella is on sale now... and they are running production batch now at COGS cost only, well... maybe they need to increase a little bit to cover production cost or whatever non COGS cost. its time to collect what they have lost... they should concentrate on support and marketing department i hope they have enough cost/resources for that too... esp time to refine they documentation, add more practical app example as i mentioned above etc... that alone can help in marketing imho, i hope they success..

Before even putting a retail price forward, you need to know your commercial manufacturing costs. Knowing that that is more than the retail price is commercial suicide. While managing to reduce cost by 50% is laudible, you need to have figured that out before you set the price, not six months afterwards. If you think the retail price dictated by the cost to manufacture means it won't sell, then move on.

If the original purpose of doing Parallella was to use notoriety to attain eventual corporate backing by making $900k of sales at a $1.5m cost, then I'm afraid that too I don't understand. While they may have achieved that goal by hook or by crook, it seems a mighty strange way to do "business": work your nuts off and then give it away.

Even if the COGS was reduced by 50%, by their own figures, they are making only $10 on a $99 retail sale, assuming no returns or other exceptions. That's just not worthy of entertaining at the volumes they're running. You might be able to survive hand to mouth, but there's not enough to be able to invest in new product.

Maybe they really did just want to work for Ericsson. It just seems a strange way of going about it, but I don't think that's really what they had in mind when they started.

Irrespective, hats off to them for describing, warts and all, their tale of woe.

[helius]

you only have 2 functions for multithread shared memory synchronization, mutex and barrier.

It's a RISC so the features are basic. Those correspond to hardware instructions, TESTSET and WAND. Building locks out of them is easy.

and there is diagram showing main0.c - main15.c run in each core

I don't know: for didactic purposes? It makes little sense, especially because there is no process that is being created for each core (as main.c implies). The normal way to start a program for this type of chip is to externally write PC and registers to each core, which may be the same PC and registers. Then release the core from halt. I don't have the board in front of me, but it would be something like

Code: [Select]

eCORE[16][16][PC] = &initfun; eCORE[16][16][RESETCORE] = 0;You could also use interrupts to cause one or more of the cores to begin executing at a chosen address.

no example yet on how to build one single multithread application and how to build elf in windows using windows toolchain.

Since it is a coprocessor, your "application" is running on the host. Linux programs are not generally developed on the windows toolchain.

[bigdawg]

Well, yeah. If you set your price at $100, and then only after the campaign closes calculate that your BOM cost is $129, and your overall COGS is $179 (!), it's a bit odd to blame Kickstarter...

Not to mention, it's all well and good wanting to keep your manufacturing in the US (and there's obviously cost, efficiency and hassle benefits in many cases), but if you're making $90000 worth of product it's not going to be cheaper...

This was just a red herring on his part; I mean his BOM was $125 so even if he would've gone to China, his COGS would still be $130-140 instead of $179.

Nothing wrong with subsidizing goods if that was part of the business plan all along, but it kinda stinks if you couldnt do basic math and than turn around to blame the platform and in effect the backers for the supposed "losses".

[janekm]

Well, yeah. If you set your price at $100, and then only after the campaign closes calculate that your BOM cost is $129, and your overall COGS is $179 (!), it's a bit odd to blame Kickstarter...

Not to mention, it's all well and good wanting to keep your manufacturing in the US (and there's obviously cost, efficiency and hassle benefits in many cases), but if you're making $90000 worth of product it's not going to be cheaper...

This was just a red herring on his part; I mean his BOM was $125 so even if he would've gone to China, his COGS would still be $130-140 instead of $179.

Nothing wrong with subsidizing goods if that was part of the business plan all along, but it kinda stinks if you couldnt do basic math and than turn around to blame the platform and in effect the backers for the supposed "losses".

Well, that kind of depends on how the BOM was calculated... If it was based on for example Digikey quotes then a lot of savings would be possible (but at the expense of a lot of time investment, and that might not have been practical for them). Of course if one just goes naively to a regular CM then the quote coming back will be within 10% of Digikey prices too, and they'll pocket the difference.

It also depends on the makeup of the BOM, connectors & passives can be a lot cheaper in China, as are commonly used chips. But then looking at the schematic the Zynq-Z7010 alone eats up almost half the BOM cost and that one won't come much cheaper in China either so they really didn't have much hope.