Xilinx padding LC numbers, how do you feel about it?

Of course it's ridiculous!

Marketing loves to make engineers hate the company. They want to make the

*managers* look past their feeble engineers for approving brand X versus brand A (forgetting, of course, that most engineers have the lattitude to push for one brand due to the *engineering* tradeoffs).

There are times when you need to shoot the engineer to ship the product (just one more tweak is *not* okay!) but thankfully the engineers get to be pissed off at marketing until they skew the preference toward the "other" brand's mismarketing BS.

Ah, what a wonderful world. What would it be like, after all, if we have to pull our hair out with our CAD tools (PC, synth, P&R, sim) but didn't have to get the least bit upset about choosing the part in the first place?

- John_H

Rickman, why do you ask? Everybody in this newsgroup hates that method, but Xilinx Marketing thinks that it is a more meaningful metric when comparing different manufacturer's devices. And it is, like it or not.

This reminds me of the way horsepower was measured for automobiles: Disconnect the gearbox, the generator, the waterpump, the airconditioner and everything else that might drain power. A red-blooded engineer would say: give me the horsepower where the tire meets the road.... Peter

Rickman, why do you ask? Everybody in this newsgroup hates that method, but Xilinx Marketing thinks that it is a more meaningful metric when comparing different manufacturer's devices. And it is, like it or not.

This reminds me of the way horsepower was measured for automobiles: Disconnect the gearbox, the generator, the waterpump, the airconditioner and everything else that might drain power. A red-blooded engineer would say: give me the horsepower where the tire meets the road.... Peter

How about you try and introduce the meaning of UNITS to them ?

ie let them use LCeq for the marketing DOCS, and also have an absolute, physical count, of LC = X * Y on the die.

As a radical idea, they could even publish the formula LCeq = LCactual * MarketingsIdeaOfLevelPlayingFieldFactor

That way, everyone is happy :)

The problems occur when one dept trys to use a std unit, but hijack it to mean something else. In the end no one knows where they are.... and everyone is annoyed.

-jg

I asked because I wanted to hear the the opinion of others. I read a recent Spartan 3 data sheet and noticed that there is now a footnote explaining they they are padding the number. I just find this pretty absurd. You compared it to car makers measuring horsepower, but that is a measurement that has parameters. Countings LCs is counting and adding a 12.5% fudge factor is marketing, not measurement. It would be more like a car maker saying his engines are 9 equivalent cylinders!

But mainly it is frustrating to always have to calculate the correct number myself. At least marketing has not gotten to my calculator yet. :)

That is silly for you to state this as a fact when it is just your opinion. Have you ever had a customer ask you to adjust your LC counts so they can compare them to the competition? Who says it is "a more meaningful metric"? Your marketing, that's who!

It is one thing to accept that your company lets marketing inflate the numbers, but please don't insult me by trying to justify it.

Am I alone in regard to finding this annoying?

Actually Pete I don't think it is. It's a very arbitrary number where apples, dead tires, horse shit, and tree stumps are all given numerical values which have absolutely no relation to particular customers needs, and declared equal in decsion tree wieghting in ways certain customers would never do the wieghting.

If the company would then derate the part for being unable route a fully used device, or power a fully used device, or cool a fully used device there might, maybe, in a long shot be a justification for the bloat factor.

The reality is that certain applications do not do arrithmentics, do not need much if any carry chain, or multipliers, or even very many F5/F6 muxes, but do need a lot of LUT's.

Being clear up front that you are offering 2 cases of applies, 4 dead tires, no horse shit, and one stump to sit on while eating your apples is a whole lot clearing than tring to figure out just how much shit you get for 37.

Old School math ... John

On a sunny day (Fri, 20 Jan 2006 10:53:31 +1300) it happened Jim Granville wrote in :

How about 'feel gates'? After all my 10W PC speakers are 1000 'music watts'. And this FPGA has a 10 giga gates equivalent feel power?

A. Nonymous.

...snip...

Jan,

I think you are agreeing with me, no? There is a difference between how many "gates" they claim and how many LCs they count. It is somewhat arbitrary how you convert from LCs to gates. You can justify a wide range of conversion numbers. But LCs are counted, not estimated or equivalized.

It is just silly for marketing to think that engineers are going to use their "equivalized" LC numbers in any comparison with another vendor or even within their own products. It's like they think we are too ignorant to realize they are doing it. They are mistaken, aren't they?

Guess I'm not the only one but I keep an excel file updated with all the key xilinx and altera parts that I'll likely use. 2 spreadsheet columns, 1 for fabric columns, 1 for fabric rows, excel calculates the product as well as subtracting the fabric cells taken up by PPC, etc, etc. I also add columns for number of multipliers, IOB delays, etc, etc, (again!). When I occasionally get a crack at a new project start, sort by the key architectural features, and end up with a shortlist of parts, usually narrowed down to 1 from A and 1 from X (I'm convinced they do it deliberately!).

You so>As we have all known for some time, Xilinx has had a habit of padding

There's also a handy part comparison generator (Xilinx only), brought to us by the friendly janitorial staff over at Fliptronics, that can be found over here:

Brian

Rickman, thank you for asking. I am in the process of part selection for a new project and wasn't aware of that particular scam!

Marketing? Heck I would call it outright fraud. The silliness with gate counts is at least somewhat understandable since there is no really meaningful conversion. But if a cookie manufacturer sold boxes of 27 cookies but stamped on the box that there are 30 inside, isn't that considered a crime? They can't defend themselves by saying in the small print that their cookies are 12% better than the competition, and they count as 1.12 each.

-Jeff

The real scam in this, is not solidly disclaiming that you are very likely to be unable to route a fully used part because it lacks routing resources. AND that you are very likely to be unable to use all the pins in many designs because you will exceed the maximum number of toggles per I/O bank. AND that you are very likely to be unable to use all of the LUT's and FF's concurrently as you can not get enough power into the device if it's even close to mostly active as they assume a very high percentage of idle or low speed logic. AND that if you do run it close to max power, it may not be possible to keep it cool. All of this comes out as you read application notes and use the online power estimator, but little if any is visible from the data sheets or other design materials that one would use up front for initial parts selection.

So, if they want to take every advantage number wise to claim usable gates, then they also need to very clearly at the same time explain each of the cases that may force the number of REALLY USEABLE gates to be significantly smaller.

Consider a reconfigurable computing application where you fill the device with small very active computational cores that are very active logic wise ... this drives the dynamic power requirements and heat load thru the roof, forcing the part to be seriously derated. With applications like this it's fairly easy to end up with applications that might have toggle rates that are very very high as a percentage of the device.

or just try and use it as a huge digital delay register with programable taps ... Nearly every LUT and FF is part of a shift register ... at any speed which is a significant percentage of rated clock speed the power and heat will force the design to be seriously derated. When you start to look at the current profile around clock edges to deliver the total power requirements the power estimator gives for these kind of applications, it's pretty clear that much power will hit the power pins only a fraction of the clock duty cycle, even if spreading the current with multi-phase clocks. They do not provide any package level power models to simulate that die isn't seeing much larger ripple than can be observed on the host pcb due to spiking current profiles.

The extreme of this farse would be a vendor claiming six times the logic resources, and 1/4 the price, but only offering routing to allow

1% of the resources to be used ... I'm sure the same vendor would be at the front of the pack to set the record straight about requiring the derating of the part due to resource limitations.

Toys,

Relax.

Is it fraud to not tell you that operating your vehicle at 115 mph might lead to death?

Is it fraud to not tell you that placing a cardboard box on your stove might lead to a fire?

We provide programmable resources. The data sheets provide exact tables for those who care as to the total numbers of everything.

If all you read is the top sheet, then you have just "scratched the surface" are are subject to the 'promotional' side of the presentation: all of the elements that we think are new and exciting, and different.

Sure, go ahead and toggle every single flip flop and IO simultaneously at max frequency, and I will assure you that you will not like the result.

But that is not what our customers do: rather, we provide devices that are programmed by them to solve problems. If the device chosen has too few resources, then they can (and do) go to the next larger device. If the device chosen gets too hot, they can (and do) use better heatsinks, or change their design.

It is called engineering.

Austin

Uhh, Austin, in case you did not read the initial posts, the misleading data sheet numbers is what this thread is about. I searched the Spartan 3 data sheet for LC and Logic Cell. The only mention I see of the counts is on the first page in the table that has the "Equivalent" count which is exactly what this thread is about.

So where is the accurate information you are talking about? "Equivalent" logic cell counts are clearly not "new", so are they "exciting" or just "different"?

If that is the way your company looks at it, why don't they put the

*actual* logic cell counts into the data sheet table rather than the marketing "Equivalent" counts?

But the way Xilinx does it, it's called marketing.

I have no beef with either you or Peter. But please don't insult our intelligence by trying to defend marketing numbers in a data sheet.

Rick,

OK. So I download:

And I see immediately when I search for LUT, the issue. Then, I continue.

And, I see no more references to LUT. OK, so S3 has not provided a count? Yup. I get it.

But, they do have total CLB's, and the comment that has the dreaded

1.125 multiplier in it.

So, I take number of CLBs, and multiply by LUTs per CLB. That wasn't too hard.

But I agree, it seems odd not to have the basic numbers there, without multipliers for "effectiveness."

Before I finish, though, I go off to Virtex 4's data sheet:

And I immediately see two different numbers: 178,176 LUTs in the LX200, and 200,448 "Logic Cells." No footnote. But there are 89,088 slices, so I multiply by 2 LUTs/Slice, and I get 178,176. OK, so far so good.

So, at least S3 had a footnote to explain the "magic," wheras V4 did not even have that. But V4 did have the basic count up front (for the largest part) to help with decoding...

Is this really such an issue?

Perhaps since I know LUT/slice slice/CLB for each family (at least I can always look it up), it isn't as if this is an ordeal to keep track of.

Would I rather that Marketing was not responsible for the data sheet? Heavens No! My paycheck depends on their abilities!

So, the tables do have exact numbers, they just require some investigation to remove the "effectiveness factor" for those who like to count.

By the way, I love to count. In fact, my father was from Transylvania, and as you all know, a role model for two generations now for all Transylvanians has been Sesame Street's 'The Count'.

"Ah Ha Ha! One, Two, Three.... I Love to Count..."

So, for anyone confused, and unable to count, just drop me an email, and I will be happy to provide any details.

And, I will continue to work with Marketing so that the numbers are at least there in some form in the tables.

Austin

Yikes!! - so Xilinx really HAVE abdicated engineering to Marketing ?

That explains the debacle that is ISE 8.1i, and the staggering lack of testing that seems to have gone into that release.

I do start to worry about the culture change at Xilinx.....

-jg

Puzzle: Would it not be LESS effort to just go fix the data sheets, than to have Xilinx engineers repeatedly make fools of themselves here, trying to justify the plain silliness....

Radical Idea: Give Marketing the front page(s) of the data sheets, and let engineering control the rest, and label the pages accordingly. Then _everyone_ knows what they are reading...

Jim,

Marketing has always had responsibility for putting the data sheet together. That is nothing new.

Peter and I have decided to work our how to do this better. We'll see how successful we are...

Austin

Better yet, leave the brochure to marketing and the data sheet to engineering. The data sheet shouldn't be the sales tool any more than the slick marketing brochure is an engineering document.

It is ridiculous having to mine the data sheets for such basic important information as the LUT count. Nobody using these devices gives a flying fig about the number of "marketing gates" or "equivalent LUTs". Those figures simply have no meaning to an engineer, and have no place on a data sheet.

Ray,

OK, OK. We have heard you.

As I said, Peter and I will do our best to influence "truth in counting."

Austin