ASMBL - hmmm

Been following the rollout of Application Specific Modular BLock (ASMBL) at Xilinx. Lots of gushing words, but what is it REALLY ?

Thus far we can glean : Stripe layout, and areal bondpads Not quite 'Revolutionary': Areal bondpads are nice to have, and the Stripe layout has plus and minus columns :

Plus: Faster local speeds Minus: Only to a certain ceiling Plus: Faster Place/Route Plus: Reduced routing cross points Plus: Faster device Testing, & possible redundancy mapping Plus: More consistant Place/Route migration Minus: Tendancy to wastage : New IP -> New Column

CPLDs have been multiple-larger-block structured for years, and Clock drivers on FPGAs are already have larger fabric elements.

But more fun comes from what the analysts think all this means :

"I believe we'll eventually see FPGAs that are fully application-specific,"

said Bryan Lewis, an analyst with Gartner Dataquest, San Jose. "But this is a nice compromise without losing the economics of having an architecture that can be used by multiple customers."

Er, isn't a 'fully application-specific FPGA' actually an ASIC ? :)

"If Xilinx is doing a chip that is 90% perfect for the app, the price

difference would have to be strong justification to make the leap into an ASIC," Snyder said.

Might be that the more defined stripes make the move to Structured ASICs easier, and this 'MASK FPGA' segment would grow ?

Perhaps ASMBL means 'A Surfeit of Marketing BaLoney' ?

-jg

Reply to
Jim Granville
Loading thread data ...

They make it sound like it is a response to structured ASICs but it sounds a little like the QuickSliver stuff to me. QS is embedding custom blocks along with FPGA like blocks. Also the Structured ASICs are biting the dust but QS has hired 50% of all the top reconfigurable computing people I have a lot of faith in their abilities.

Steve

application-specific,"

is

Reply to
Steve Casselman

Jim,

Now really, Jim, why so negative? Isn't Xilinx justified in rolling out a new architecture with some fanfare? Why is it that others are allowed to make outrageous and sometimes even false claims, and no one cares? Yet when we announce a truely revolutionary FPGA architecture, we get complaints?

If we can architect a device to allow for any collection of "stripes" and have the mask, and the software ready immediately, isn't that worth shouting about?

More and more we see market specific applications that require a more cost effective FPGA (eg software defined radio is a totally different mix of features than automotive entertainment center). Why not be able to target a general purpose device to a specific market segment at reduced cost and better margins? Might get more business that way, right? If Spartan 3 is already cost effective against many ASICs, might this not make us even a better cost/benefit solution?

I do agree that the analysts do not have a clue, however. For example, we are being compared to structured ASICs and other "hard" solutions. Those are misleading and false. They are all dead-ends (just a re-spin of gate arrays with new clothing, or some other ASIC flow with window dressing liberally applied).

As for the details that I am sure you want to hear, you have to wait until the next press release.

Schedule: architectual release, then specific family details, then product announcement. Nothing new here. Been following the same formula for many years. Happy New Year!

Aust> Been following the rollout of

Reply to
Austin Lesea

"Austin Lesea" wrote

I did not think I was negative on the engineering aspects ?

- I listed more pluses than minuses :) ?

fanfare?

Of course, but when the gush exceeds the hard data, expect some analysts to get it wrong....

"truely revolutionary" is a bold claim.

I can see good Engineering & Yield trade-offs in what's released so far, but that's some way short of "truely revolutionary".

Can you clarify 'the mask' ? Does this mean this will become like a block-hard-copy (but still FPGA), where a large enough customer (/market?) can 'select the mix of stripes' and a new die results ? With real care, 'the mask' could even be effectively virtual by using stripe based exposures at the wafer level.

Challenge there will be in the definition, to get devices to reach critical mass, and not go EOL as uptakes do not quite meet forecasts.

So is this not a merchant market device, but an 'ASIC cherry pick' vehicle ?

-jg

Reply to
Jim Granville

Reply to
Peter Alfke

"Peter Alfke" wrote

Sounds like a new marketing manager fresh in from another industry.... ;-)

Even the technical press have taken the slightly misleading "Xilinx Unveils Revolutionary..." and watered it down to the more accurate "Xilinx set to debut novel FPGA architecture" "..set to debut in a 90nm Virtex device in the first half of 2004, segments function blocks into interchangeable columns, rather than squares on a grid."

-jg

Reply to
Jim Granville

Jim,

It is NOT "hard-to-copy"! Retains ALL of the reprogrammability and functionality of a 'true' FPGA -- just does it for a better cost/benefit trade-off.

"Hard-to-copy" means that if you make a mistake, you are toast. It is inevitable that customers do not think of absolutely everything. The new architecture will also allow us to provide EasyPath(tm), which if the customer calls up (which has happened, and boy are they glad they went with Xilinx!), and says "I forgot an inverter" we just modify the test program, and they are back in business IMMEDIATELY, with only a small charge (perhaps very small) for the modified test program work that we have to do. No one else can do this, either! In fact, EasyPath(tm) can be done for a few images (bistreams) and still retain the savings. This fits it well with customers that wish to have single platforms that perform multiple functions (ie cellular basestations which are notorious for needing major changes every three months).

"Hard-to-copy" on a base station? Disaster. End of the world. Remember I worked in telecom for 23 years, and FPGAs saved by rear-end so many times, I can not count them all. Ever had a Class-A recall of 100,000 units from telephone companies? Nightmare. But if they are reprogrammable, we only needed a seed stock of 1,000 units, and then sent the new ones out, got back the old ones, reprogrammed them, and went on thru the whole 100K. Saved the company. Saved the product. Today you could send out a floppy, flash, or download it over the 'net.

That is why we state that "structured ASICs" are just a new description for a "buggy whip" -- not even in the same century as what we are doing.

As for 'revolution', who has ever announced such a chip architecture? Certainly no one in the FPGA industry. Haven't ever heard of it anywhere else.

'Novel' is accurate as well. It is new. Prefer 'revolutionary', as that implies a complete re-thinking, and introduction of radically new benefits and features. The press announcement mentioned 100 innovations. That is no small feat for a new device family.

I suppose Peter and I are guilty of working here, but some of the things that are coming up are awesome.......

Aust> "Peter Alfke" wrote

Reply to
Austin Lesea

Howdy Austin,

Austin Lesea wrote: [...]

Hmmmm. If the customer is just changing the contents of a LUT, does that really require requal'ing EasyPath parts and changing the test program? My understanding of EasyPath was that components and routes are tested, but I guess I don't have an exact definition of "component." Could the contents of a LUT be changed without going through this hassle? What about changing the polarity of a signal (or clk) going to a flop? IOB parameters? Contents of a BRAM?

[...]

And yet there are perfect applications for both buggy whips and structured ASICs. Funny how an improvement on something doesn't necessarily out date everything before it, isn't it?

Where I work, whenever we hear the word "revolution," our ears perk up. Because we know that that history has shown that most real revolutions fail, or at least, the first couple attempts at the revolution do. Both old history and recent history, in all areas, from politics to the telecom world, to the semiconductor world. It is VERY rare. So if I were you, I'd personally shy away from that term unless it truly is revolutionary, in which case, as a shareholder, I hope you're not betting too much on it.

I am not saying I'm against revolutions when they are really called for. I'm just saying that the word has been associated with A LOT of failure and that whenever I hear it coming from the mouth of a marketing droid, I hold onto my wallet.

Indeed they are. It is a very exciting time for FPGAs, semiconductors, and electronics in general. But just like we don't blame you for talking up Xilinx, don't blame us for trying to keep the marketers in check with all the fluff (and few details) they put out.

Have fun,

Marc

Reply to
Marc Randolph

Jim, here are some fictitious press releases that might be to your taste:

Boeing 747: like a 727, just bigger Porsche 911: like a VW, just faster Nokia cell phone: like a Walkie-Talkie, just lighter Transistor: like a vacuum tube, just smaller LCD display: like a CRT, just flatter FPGA: like a CPLD, just more complex

Any spectacular innovation can be de-dramatized. Wouldn't excite any magazine editor, though. Peter Alfke

Jim Granville wrote:

Reply to
Peter Alfke

Marc,

--snip--

If they use a LUTRAM, SRL16, BRAM, then those features are 100% tested.

If they do not use a clock inversion, or do not use a different IO standard, then these are not tested.

Oh yes. I have a buggy whip at home. Use it all of the time. Right.

Most revolutions fail? Hmmm. I suppose you also belong to the "glass is half empty crowd."

Austin

Reply to
Austin Lesea

Fear of revolution? Let's stay away from politics, and look at technology:

Programmable logic was a revolutionary idea (Signetics, MMI, Xilinx) FM radio, LCD display, the PC/Mac, transistors, the printed circuit board, the idea of a stored-program computer, telephony, radio, etc are all revolutionary and successful ideas in electronics.

The 4-stroke gasoline (Otto) and Diesel, and jet engines, powered flight (Wright), movable type (Gutenberg) were all revolutionary inventions that caught on extremely fast. Vaccination and recent progress in medicine must be called revolutionary. (Nice not to worry about polio anymore). DNA testing in law enforcement...

Let's keep at least some of the enthusiasm for scientific progress that permeated my youth, long ago.

Let's hear it for revolution again! Down with pessimism and sarcasm ! Peter Alfke

Reply to
Peter Alfke

"Peter Alfke" wrote

Maybe - I'm still trying to cut through the fluff, to make my own mind up.

worth

One poor aspect of the info so far, is it does not delineate SOFT and HARD features clearly at all. So, I am trying to see what is new, or clever, and may be over-interpreting some info....

Putting the discussion of 'revolution' to one side, can we focus on what really is new, or clever ?

My reading of the fluff so far, is that there is some freedom to stripe-select at FAB time, and so Xilinx can roll new die variants relatively easily to target new market segments - correct, or not ?

Now, if that really is a whole new mask set, then that will have significant NRE costs, esp at 90nm, and so restrict the market reach, but I could imagine a smarter-flow where the stripes are 'printed a stripe at a time', and this would slash NRE to an admin/package/test flow aspect. Q: Which of these applies to 'the mask' ?

-jg

Reply to
Jim Granville

Jim,

---snip---

Correct.

The rest has to wait for the next press release.

Austin

Reply to
Austin Lesea

Jim, it's best to stop reading too much into the press release. We should never even have mentioned stripes and soft and hard, it just sends smart guys like you off in strange directions.

Let's keep it at: Xil>

Reply to
Peter Alfke

Hey, I'd be happy if they get their current hardware and software to operate fully. :)

Peter Alfke wrote:

--

Rick "rickman" Collins

rick.collins@XYarius.com
Ignore the reply address. To email me use the above address with the XY
removed.

Arius - A Signal Processing Solutions Company
Specializing in DSP and FPGA design      URL http://www.arius.com
4 King Ave                               301-682-7772 Voice
Frederick, MD 21701-3110                 301-682-7666 FAX
Reply to
rickman

I don't understand that. I assume we are discussing saving $ by only testing the parts of the chip that will be used rather than the whole thing.

If the base stations are changing every three months, are they just switching between a small collection of choices that are known ahead of time, or are they rolling out new features?

If they are rolling out new features, then they are probably making interesting changes to the FPGA code and probably would want the fully tested version rather than the cheaper one that might not run the new code.

I like the idea of saving money, but I've never worked on gear that didn't get fixed/upgraded in the field.

--
The suespammers.org mail server is located in California.  So are all my
other mailboxes.  Please do not send unsolicited bulk e-mail or unsolicited
commercial e-mail to my suespammers.org address or any of my other addresses.
These are my opinions, not necessarily my employer's.  I hate spam.
Reply to
Hal Murray

Hal,

To use EasyPath(tm) one has to have a stable design, or a set of stable designs (like 3 designs). If you want to upgrade designs in the field, then you might be in trouble, just as you point out.

You are not 100% guaranteed to be in trouble, but the yield of a new untried pattern to a set of EasyPath product will not be 100%. It will not be less than about 90 or 95% either, based on our tests and research.

For some that is acceptable. For others, not. In any event, "Hard-to-copy" offers no ability to change at all, so there is no comparison. There is a 100% guarantee of returning and replacing all parts and boards if there is a required change.

If I am trying to manage my risks, 10% failure vs 100% failure is a big factor.

Nothing is perfect, but we strive to get the best fit for the customer with our product offerings.

Aust>>"Hard-to-copy" means that if you make a mistake, you are toast. It is

Reply to
Austin Lesea

Is this 90-95% based on a complete new pattern, or some small respin of the known good design ?

Interesting numbers. This does point to some usefull tool features :

1) Ability to reuse as much existing place/route in a change as possible ( even to the point of less optimum results ) 2) Ability to create multiple, seeded changes to a design. - ie if your design change can hold 98% of previous proven/tested place/route and give, say 3 version of the 2% 'fix', then in a expensive HW situation you could try these 3, and choose the one that works. 3) Ability to place idle-test-patterns, into spare resource areas, with a view to giving some test coverage.

It's all more work, but does give a satistical method to reduce the non-zero chance of 'must change HW', esp if that HW change is very costly.

Reminds me of a story I heard about the very first Russian embedded uC, (military) where production came serialised with a list of which opcodes did not work, and the designers task was to code around that list. (Intel tried a similar path with their Pentium a while ago :)!

-jg

Reply to
jim granville

Jim,

The coverage (?) of a completely new and different pattern is about 90 to 95%.

That is just because about 90 to 95% of any FPGA is unused (really, just count the number of bits set to a '1').

The overlap of bits that change from one design to the next is probably so random, that there is not much use in trying to keep track.

Aust> > Hal Murray wrote:

Reply to
Austin Lesea

Easypath = EasyScrap...

Hey, buy our excess scrapped inventory!!!!

Reply to
acm

ElectronDepot website is not affiliated with any of the manufacturers or service providers discussed here. All logos and trade names are the property of their respective owners.