Re: Requests for feedback on future AT91SAM9 development kits

[bits elided]

Ulf Samuelsson wrote:
D Yuniskis skrev 2009-12-16 22:20:
Ulf Samuelsson wrote:
I am involved in some internal discussions on how the future development kits for high end ARM products should look like.

Have some ideas, but I would be interested in feedback from any users on the current crop of kits. AT91SAM926xEK etc. General comments on this are also welcome.
Remember that the SAM9 products are mostly microprocessors (non-flash) intended to run things like Linux/WinCE etc.

* What don't you like?
* Single board computer vs tool based on CPU module?

CPU module. Chances are I will *not* be using any of the I/Os
you are going to waste time, money and real estate implementing.
Make an OPTIONAL cheap passive backplane into which the module
can plug so folks can use PCI cards for the I/Os they might
want (purchased off the shelf from commodity vendors instead
of a low volume "specialty" manufacturer like atmel).

The current crop of Evaluation Kits are just that.
They are not Development Kits, allowing you to run your own application
before you have your own hardware. A weakness, in my opinion.

PCI connectors have crossed my mind.
Not with a real PCI bus, since the AT91 processors don't have that,
but rather a mix of serial peripherals like UART, SPI , I2C, SSC etc
made available on multiple connectors using the same connectors as the PCI bus.

I think that I have seen 1 customer out of maybe 400 Atmel ARM9 customers
which has put a PCI bus in their own design, so this is not something
people will find useful.
I do think personally it would make sense to have AT91's with PCI or PCI Express,
and then of course a development board should allow connections to standard cards.
I think especially WLAN would be nice to add this way,.

My intent might not have been clear... :<

Here's one scenario:
- user plugs your "module" into a "motherboard" (lets not
get pedantic on terminology here) that *you* manufacture
and sell at a *reasonable* price (i.e., not what you think it
is *worth* to the user in terms of the work fabricating
something like this that it allegedly SAVES him -- since NOT
USING YOUR PRODUCT AT ALL will save him that work/money just
as easily -- but, rather, "for *your* DM+DL on that item")
- user grabs some number of COTS PCI cards and plugs them
into that motherboard (perhaps a gigabit NIC, a SATA controller,
a BT transceiver, etc.)
- user links/loads the sample drivers provided for those cards
to his application
User now has a viable platform for investigating the suitability
of your processor to his intended application *without* having
to spend much time or money hacking together "one off" prototypes
of his hardware.

The user obviously knows that any real hardware that he develops
will have a different cost basis than this "prototype". Just like
he knows it will have a different physical form factor, etc.

And, he knows that any software deployed on that final hardware
will have different performance characteristics.

But, he can *quickly* decide if pursuing a solution based on
your product makes sense to him. And, he doesn't have to make
a huge investment (time *or* money) to come to that decision.

If he decides to pursue this approach, this crude prototype
enables some software development to begin immediately.
Concurrently, *real* hardware can be in the works that will
allow the development effort to migrate to the "final"
target more quickly.

OTOH, users who don't fit this collection of COTS I/O's can
design a daughtercard (a mother-of-a-different name?) with
their I/Os and just piggyback your "module" onto/into it.

In the end, I think there is a better solution, which will provide
the pins to standardized connectors, instead of committing them to
different kind of physical interfaces like RS232, EEPROM etc.

It is a twist of the old 10 pin header available on the STK500.
Instead of putting just the I/O ports out on each header,
I would like to standardize the use of the header.
I:E: UART.TXD is ALWAYS on pin 0 of a header.
UART.RXD is ALWAYS on pin 1 of a header.
I2C.SDA is ALWAYS on pin 7 of a header etc.

When multiplexing allows it, you may have several functions per header,
otherwise you could have a backplace with 5 UART headers, 4 SPI headers
and 2 I2C headers etc.

I don't see that buying you -- or your user -- much of anything.
Use an *existing* standardized connector: PCI. Connect at a
much higher level, architecturally.

Connecting some level translators and a DB9/25 to a "standardized
header" saves you an hour or two of a technician's time (?).
You're saving "crumbs" and missing the Main Event.

When you have to design a strain gauge amplifier with A/DC
interface to that card of yours, those few man-hours are going
to be insignificant. OTOH, if the user can just *buy* a COTS
DAS card and plug it into some "Industry Standard" interconnect
mechanism (i.e., a PCI bus), then you have saved man-*months*.

Port some open source drivers (preferably NOT GPL'ed) for
a few industry standard cards to this "motherboard+module"
so folks can see how they can merge those hardware features
into their designs.

Until there is a real PCI bus, there won't be any industry standard cards.

And who better than to undertake the cost of designing and
implementing that sort of *bridge* adapter (that will
probably NEVER have any chance of commercial success as a
"product" of its own) than the company who most stands
to benefit from it??

I.e., that "motherboard" design mentioned earlier will never
be the basis for a real product (well, I should learn never
to say "never" :> ). It is a cost of doing business for
you. A means of making it easier for users to investigate,
evaluate and *develop* products using your components.

Why do you think so many products end up x86-based? Because
it is just *so* easy to slap together something using a PC
as a development target. (You yourself said this uP is
targeted for "folks running Linux" and not "folks designing
microwave oven controllers"). I've been developing under
a PC target with *no* intention of deploying on that
architecture -- simply because I can piece together all
of the various I/O that I need (along with drivers) so
much easier using that "Standardized Interconnect" bus.
If I decide the network interface is just not going to be
fast enough using a given technology (e.g., 10 vs 100 vs 1000)
then I can just swap out a card without having to spend
months designing and debugging another "one off" prototype
with a *different* network interface :-/

* Memory size

DIMMs. Let the customer provide his own. *OR*, sell industry
standard DIMMs to the customer AT VOLUME PRICES despite the one-off
quantities.

DIMMs are driven by the PC market and they are usually 64 bit or wider.
A 64 bit DIMM would waste half the memory.
With DDR-2, the AT91 interface is 16 bit, so you then waste 75%
(I don't know, but I assume DDR2 DIMMs are still 64 bit).

So what? If the user wastes 98% of a part that he has lying in his
desk drawer, it's 98% of $0 that's wasted. And, he can chose to
waste 98% of a DIMM that is twice as large 10 minutes from now
(instead of wondering how he is going to get another few megabytes
onto this "custom" development board)

*I* can swap out a DIMM as easily as I can swap out a 100TX NIC
for a gigabit NIC. I don't need to get my boss/client involved.
I don't need to get purchasing involved. I don't need to wait
weeks (?) for atmel to process my order and ship the "special"
parts that I need. So, I can see what happens in a new
configuration without spending anything (money or prestige)
to do so.

An old style DIMM connector optimized for 32 bit and custom memory
module, would allow the user to build a prototype using their own
preferred memory vendor.

I don't think it makes sense to use DIMM if you have a CPU module.

* Too hard to use because of <...>
* I like the following feature (please keep it): <...>
* What connectors should be available. Location of connector
* What physical format?

CPU module. Something that makes it easy for folks to
prototype around (i.e., no BGAs).

There are some chips which are available in TQFP208.
These packages are very large compared to BGA.

Yes. But prototyping with BGAs is a PITA. I can layout
a board and have it assembled by a technician here if I
avoid BGAs. And, have it in my hands very quickly.
And, run through another iteration just as quickly.
With BGAs, it has to be fabbed outside. It gets more
involved (read: management, purchasing) to do that.
So, you don't want to do it often -- lest folks start
raising eyebrows wondering why you are "wasting" all
this money, etc.

I think it may make sense to deliver a CPU module with
large amounts of SDRAM, and a small dataflash,
with empty pads for adding large parallel NOR
as well as the NAND flash of your choice
in TSOP so that they can be handsoldered to the module.

Why stop there? I.e., decide what your philosphy is
going to be and then just stick to *that*. If you think
the user *needs* flash on the module, then *put* flash
on the module. If you think the user will provide
his own flash on a duaghtercard (or, my personal preference,
leave the flash out entirely and just treat ram as flash),
then why waste real estate on an option?

* Additional peripherals wanted?

None. Move them onto daughtercard(s) or motherboard.
I see a CPU module containing
* CPU,
* SDRAM,
* Flash,
* Power Management
* Ethernet PHY.
* Possible a tiny 8 bit controller doing systems control
* Battery
* Optional Micro-SD card connector.

Again, you're engaging in feeping creaturism. Put those
things that *must* be there on the card. And, those things
that are really hard to do "off card" (e.g., one of the Cirrus
processors I used had a VCO for some of the video timing;
moving this onto a daughter card would have been a nuisance)
Those things that a user might chose *not* to use don't
belong on the card (unless you can do it for *zero* cost
and zero *space*/power).

I.e., why not include a power supply on the card, too? :<

* Developed by Atmel vs outsourcing to board vendor

I doubt folks would care too much -- it depends on how
atmel treats it in terms of pricing. I.e., if it becomes
a cost center that has to pay for itself, then it will be
far too pricey. OTOH, if atmel considers it a loss leader
and sells it for "the price of the components" (using
big volume pricing) then you'll find people more willing
to "give it a try".

From comments in other thread here, it seems like this
might be anathema to atmel. If *smaples* are so hard to
come by, I suspect an *inexpensive* development kit
would be likewise.

Low cost development kits has generally been the AVR realm
and usually there are no problems getting hold of these kits
at low cost or as giveaways.

Theoretically there should not be a problem getting hold of these kits.
Atmel like most electronics companies outsource to CEMs,
sometimes you are at the mercy of their whims.

Perhaps you should consider a different marketing/sales model?

Think about the folks that are likely to be buying this:

For large corporate users (possibly with deep pockets),
there is often a lot of bureaucracy in the way to making
a purchase. The last set of development tools I purchased
in such an environment took three months to run the paperwork:
"And *why* do we need this...?"
Yes, people are aware of this.
Ideally, you should only have to ask your personal manager for approval.

No. Ideally I shouldn't have to ask *anyone* for approval!
I.e., if it was cheap enough I could buy one out of my own
pocket, play with it and *then* decide to make a pitch for
a design-in based on it. That's where the PICs have made
a big impact. I don't think most folks like them. But,
they play with one at home. Or, at school. And its no
surprise that when they get into a position where they
have to make an engineering recommendation, they go with
something they already *know*.

No, I don't think many folks are going to reach into
their own pockets just to avoid dealing with their boss
or purchasing department. But, if your price point is set
that low that this is *possible*, then it is highly likely
that your boss won't feel like he is sticking his neck
out putting through a request for something like this.

If it has to go 2-3 levels up, there is a lot of resistance in the system,
which does not gain the vendor, nor the engineer.

All things sold on TV-Shop is normally so cheap that you can buy
without asking your spouse. Peg Bundy not counted.

Atmel has to decide what business they are in: selling
components or selling development tools and components.
If you consider tools to be a product and expect to make
money off them, then you approach them with an entirely
different mindset than if you consider the tools just
a means for selling components.

I did some work on a device that cost a few hundred dollars
to manufacture (DM+DL). It was priced at several *thousand*
dollars -- and *none* were ever sold! They were intentionally
given away to sell *other* products that had even more
promising profit margins and volumes. I.e., like giving away
gold plated toilet paper dispensers -- in order to sell
the toilet paper that is used in them!

Things should be as cheap as possible, but not cheaper.

This all depends on what you consider as part of your
"cost". I.e., *I* see tools as a cost of doing business
so their "cost", IMO, is DM+DL. There is no "profit"
inherent in them. They have a role in selling other
items that drive the profits. If, OTOH, some bean counter
there wants to amortize the cost of developing the
hardware for your "module" over the total quantity of
modules that will be sold, then there is a huge accounting
cost reflected in your "cost" -- one that might make sense
when compared to the cost for a company that *wants* to
use your product and undertakes development of a platform
on which to evaluate it. The problem here is that if
some *other* company has a lower "cost" (because they
sell more of these development kits, or because they
underwrite their costs for developing them), then you
are at a terrible disadvantage.

I'm not fond of Microchip parts. It seems like we're
still living in the days of GI! :< But, they *do* seem to
have the right idea when it comes to development tools
and kits. Consider the cost of designing, building, selling
and supporting those tools just "part of the cost of
doing business" (i.e. SWALLOW IT!) and you'll probably
win more designs in the long run.

That's the AVR strategy.
For ARM, there is a lot of third parties so there is not a lot of S/W tools development for OS/Compilers etc. Only programming S/W and drivers.
..
The typical simpler boards are like $500-800, which is one level
above your typical first level manager signoff,
but that seems to be low enough, for this not to be a problem.

Faced with exactly these sorts of numbers, I opted for the
"lets explore the application on a cheap platform -- a PC -- and
see if it makes sense before throwing money and time after
a *prorpietary* platform". For $800, most develppers
would probably rather have their boss buy them a new PC!

It is fairly easy to borrow boards, and sales guys tend to
give away to major customer anyway.

And if you aren't a major customer, tough luck. So, you
effectively limit how many *new* customers you gain.
As each time the customer has to make a choice for a
new product, they have to go through this analysis
again. And the politics, etc. (and, if you discontinue
products, you just exacerbate the problem and your
reputation within the company -- "Didn't we just buy
one of those things LAST year?")

I think that if the boards could be brought down to below $300
it would be good, but you also would like to have a lower price point.

When you go to your local department store, do they charge
you for parking? Do they charge you if you take a drink
from their water fountain? Or, use their bathroom facilities?
They consider these costs of doing business. Indeed, those
stores at which the patron must pay to park (i.e., because the
store itself has no parking to offer) tend to pay a price
in terms of sales volume: if you can find the products
that they sell offered elsewhere, you *buy* elsewhere. It
is particularly annoying to pay to park at such a store
and then discover that the products they offer don't fit
your needs (i.e., I have wasted money *and* time -- like
buying an overpriced development tool, spending time
evaluating the device in my application, and *then*
discovering that it is unsuitable for my needs: anyone
want to buy a slightly used boat anchor?)

The ATNGW100 with an AP7000 (AVR32) processor sells for less
than $100 and has 32 MB of SDRAM and 16 MB of Flash and runs Linux.

The big difference here, I think, is that the guys working on the 32 bit
AVRs deciced for low price in the beginning, and ordered
in volumes high enough, that allowed them to get price breaks.
If you order 100 boards here, 100 boards there, the price of
components (non-Atmel) becomes signifiant.

Again, it depends on how you look at this activity in the
Grand Scheme. No doubt, your sales force "treats" clients
to free lunches. They think nothing of spending hundreds
of dollars weekly doing this (for each sales person).
So, clearly, they have decided that this is a necessary
cost of doing business. It is reflected (indirectly) in
the prices charged for the parts that are eventually sold.

I'm saying, treat the development tools in the exact same
manner. No, you don't have to *give* them away (though you
don't mind giving away meals!). But, pricing them at
DM+DL -- or, even better, DM! :> -- makes it clear to the
purchaser that you aren't trying to make money off of
this as a "product". Instead, it's "sales support"
(like free lunches, tech support, etc. -- do you charge
$3.95 per minute for your tech support hotline?? :> )

I've seen Freescale sell their iMX25 development kit for $1500,
and that is of course much worse.

And there are some folks who will buy. And a lot who won't.
What this says is they either think they have a monopoly on
this particular sort of device in the market *currently*;
or, they are happy with the number of customers that they
have and their sales volumes.

After all, isn't *that* what you REALLY want??

There is always someone at the top that rather take the money.

Yup. And that works. Sometimes. And in some markets.
But, when it doesn't, its awful hard to regain a good
reputation as being "in partnership with" your customer.

I look at the businesses that I, as a consumer, "enjoy"
doing business with. And those that I don't. It is
easy to see the things that the "good" have in common and
how they differ with the "bad". The latter seem to always
be trying to figure out how to nickel and dime me -- "sorry,
you can't return that without a receipt", "sorry, the sale
ended yesterday", "Sorry, you will have to talk to the
manufacturer to handle that defect"... OTOH, the former
seem like they are working *with* me: "well, there is a
two item limit clearly printed on this coupon, but I'll let
you have three at that price", "yes, the sale was over
yesterday but I'm sure if we could afford to sell it for
that price yesterday, we can probably sell it for that
price today -- let me get a manager to approve this for me"

My goal as a *vendor* is to keep my customers thinking of
me in that former case and not the latter...
.

Relevant Pages

  • Re: TV Tuners - Why poor choices in US Markets ?
    ... There is not much of a price difference between a multi-format ... card and a region specific card suggesting that the additional ... cost of adding this technology is minimal. ... one standard world wide - that's not going to happen. ...
    (comp.os.linux.hardware)
  • Re: Whats really been killing PC gaming? - A Long Comeback
    ... To some degree I agree w/Mag though, top tier video card prices ... >$300-400 USD is the norm now, ... 7800GTX comes in about the same price. ... >>> really increased in user cost at all. ...
    (comp.sys.ibm.pc.games.action)
  • Re: Whats really been killing PC gaming? - A Long Comeback
    ... To some degree I agree w/Mag though, top tier video card prices ... $300-400 USD is the norm now, ... >> really increased in user cost at all. ... >> Price got inflated along the way, ...
    (comp.sys.ibm.pc.games.action)
  • Re: OT: US Presidents and Foreign Policy
    ... That's free money. ... You buy something for $100.00 on a credit card and the retailer ... In effect the retailer puts the price up to cover ... > the cost of getting transferrable funds in other ways is greater. ...
    (comp.lang.cobol)
  • Re: HTTB Spoilers
    ... In combat, the controller of the opposing minion plays ... Reaction, aus/val, no cost ... Cancel a combat card played by the opposing minion as it is played (no ... A vampire can play only one Death Seeker each round. ...
    (rec.games.trading-cards.jyhad)