not totally repulsive

Keep in mind that VCCAUX is used to power all sorts of stuff, and I don't think it's all documented. So, if you change the FPGA design you may get more/less VCCAUX current. That along with temperature effects may push the voltage beyond its published limits.

If it were mine, I would bite the bullet and put in a proper ldo.

Maybe some wonderful Xilinx guy (e.g. Austin) will drop his $0.02 regarding the constancy of VCCAUX current on S3.

Bob

There will be a temperature variantion on this, and you should verify the drop in operating conditions (it is a diode, and any ringing will be rectified - so a slow one like a Zener is probably a good choice )

Also note that the % Vcc variation is amplified on this.

3.3C +/- 10% or 3.0..3.6 , now becomes 2.2-2.8V, and a 20% window, is now 27.3% - so you'll need tighter starting Vcc levels.

But it will work. I've also looked at using Yellow LEDs as low-cost

1.8V shunt regulators for CPLDs :)

-jg

You realize that there ire tiny switching regulators that are about the same size as the monolithic linear ones?

National semi was any easy tool that will generate a reference design.

I use them and they work great, but you won't find them easy to cobble onto an existing PC board that expected a 3 terminal regulator. Most come in packages with no leads and require a large solder pad for the heat slug in the middle. For very low currents you could use a 5-pin SOT23 package like the LM3674MF-ADJ. It's a nice device but doesn't come in a 2.5V version so you need the external resistor divider to get 2.5V.

For the prototypes you are probably best off looking for a very low drop-out linear device, which are available in 2.5V versions and require very few external components. IIRC NXP makes some that are stable with little or no capacitance.

Regards, Gabor

John,

Dropping the 3.3v to 2.5v with a pn junction is just so easy to do, that I am sure you are not the first one to do this (in fact, we did with a SCR in the lab for a Spartan 2 app note - 2.5v core).

Other than the obvious, that the voltage is likely to vary over temperature beyond the recommended Vccaux range, if you have verified that your design works over the temperature range you need, then you are "done." You need to be sure no one changes the recipe for your zener diode (perhaps just buy all you will need, and then the next revision just use a LDO regulator that operates off of a lower voltage).

Vccaux is used for bandgap reference voltage generators, which will start working at 1.8 volts, and work fine up to beyond 3.0 volts. There are a few more circuits also using Vccaux, but generally, it is not as fussy as say the Vccint. It will affect output timing, as Vccaux is used for the predrivers, and it will also affect LVDS inputs (on some parts where the diff-amp is powered by Vccaux, the latest parts use Vcco for that however).

The reason for the recommended 5% specification, is that we have to specify a lot, and it is far easier to specify all supplies at 5%, rather than have each supply have its own rated range, and then have to characterize everything over all of the ranges.

At less than 100 mA, it is not easy to get ~2.5V from 3.3V any easier than what you have described. An efficient switcher with 250 mW capacity is also not easy to find (switchers are inefficient if used at a power much much less than what they are designed for).

Austin

Certainly. But they need inductors and probably secondary lc filtering before we get to the analog stuff. I believe I mentioned going to switchers next rev, although making Vccaux with an ldo from +3.3 ain't bad. Even if we upgrade to an XC3S1500 at 128 MHz, which we may do, the current will still only be about 100 mA, which is only 80 mW loss in a linear reg.

Dang, the power supplies are more trouble than the FPGAs.

John

[snip]

If they wanted to, these FPGA manufacturers *could* make a FIVE-VOLT-only FPGA.

I think they're just being stubborn.

:-}

Bob

Thanks for the insight on Vccaux internals. Hell, sounds like I can use the diode forever!

As far as tc goes, even a "full diode" tc of 2.5 mV/K won't affect things much. But when a silicon diode has this much forward drop, a good chunk of the drop is ohmic, which has the opposite tc from the pn junction itself. I wouldn't be surprised if this diode were running near its zero-tc point. I'll measure it if I get a chance.

John

Do the right thing after the spin: LDO from 3.3V or switcher from 12V.

Prior to that, how many existing boards do you need to use up, and how many customers can you afford to loose when they do not work reliably?

Even after you test one or two samples over temperature extremes, a sample of one or two is no basis to design a production run on.

Andy

They did make 5V only FPGAs for years and years...few people want them for new production designs anymore ;)

KJ

I think we bought 5 or 6 of the first etch. So far, everything else works (it's a 4-channel DDS-based waveform generator; a future spin will be a 4-ch 32 MHz arb.) We plan to add a couple of features and move one connector, so we won't fab any more of this rev A version.

But based on my measurements and Austin's comments, it sounds reliable as-is, especially since we can check the actual Vccaux on each unit to make sure the diode is behaving. So I'd be happy to sell a few of this version.

We're looking into switchers now. Most of the interesting (small package, synchronous, 1 amp at least) parts won't accept inputs past

5.5 volts, so I guess we'll switch 12 to 5 with a Simple Switcher + schottky, use MSOP synchronous switchers from 5 to 3.3 and 1.2, and an LDO (or even a diode!) for 3.3 to Vccaux.

Why not? We have a reel of 3000 diodes. And we usually check all the supply voltages anyhow; each has a test point. A switcher or an LDO could be far more wrong than any diode is likely to ever be.

John

John;

A TI TPS75003 power supply controller has 2 3A switchers and a 300mA LDO. This part is on the Xilinx/Digilent Spartan 3E board. Its max Vin is 6.5 volts, so as you mentioned you'd have to regulate the voltage down to 5 volts first.

-Dave Pollum

Smiley noted, but there is an element of truth in this.

If you look at the microcontroller sector, OnChip regulatros are very common on newest uC chips, and the Automotive/Industrial sectors put pressure on 5V compliance at least, and operation preferably.

CPLDs are candidates for OnChip regulators, and those are already done (just not at great Icc values).

FPGA's could easily REDUCE the supply rail count, but to move all rails to regulators is a bigger ask, because of the power budgets. FPGAs are in peril of thermal runaway already, do you want MORE heat pumped into that tiny area ?!

Also, unlike uC design teams who are very 'analog aware', FPGA development is rather cocooned in the digital world - Linear stuff !?.

-jg

...usually check...likely to ever be...

At least the LDO/switcher was tested (or at least a statistically significant sample was tested) and it's production optimized to meet published specifications. The diode was neither tested nor optimized for your "specification". The diode could be completely "right" per it's specs, and still "wrong" for your implementation. Such is not the case with the LDO/switcher, so long as your application is within its published specifications.

Not only do you have to check the diode's performance across parts and over temperature, input supply variations, with aging, etc.), you also have to check the FPGA's performance across parts and conditions. What is the part-part (let alone lot-lot) variance in the fpga supply current, over temperature, and with aging? Do both components vary sympathetically or otherwise?

Once you start designing outside datasheet specifications, you lose the advantages of the manufacturer's testing and process optimization to meet those specs over the environment, over a large volume of components. It can be done (and often is), but it takes a lot of expensive testing and analysis over a large sample base (both components and conditions) to ensure it works reliably. It should be the last resort, not the first.

Andy

...usually...likely to ever be... Those are words I like to hear when I purchase a product.

A switcher/LDO would have been designed, tested, and optimized to meet it's specifications (the same ones you would be using) over large production volumes. Such is not the case with the diode in your application; it could be completely "right" per its specs, and not work for your application.

Not only do you have to test a large sample of parts, but over a large sample of conditions (including component aging). And then there's the combination of FPGA and diode. How much does the FPGA supply current change across conditions/environments? Does it vary sympathetically or otherwise with your diode?

When designing outside of datasheet values, you lose the benefit of the manufacturer's design, test, and optimization of that part to meet those specifications over very large production volumes. You can replace that with your own testing of lots of samples over lots of conditions, but it is expensive. It should be a last resort, not the first choice.

Andy

Something that nobody has mentioned directly is that running 50 mA through a Zener diode in the forward-bias mode may (and may not, too) have effects on the long-term reliability of the diode. You say you used a 1 W Zener, so it should be good to 120 mA reverse-biased. Does the data sheet give any spec on forward bias limits? I'd want to know the manufacturer has run some of their typical Zeners forward biased at significant current for an extended period, as this could be something they've never tested for long-term effects.

Jon

Hi Jim,

I think that because FPGAs are on the latest and greatest geometry, building SMPS onto the dice is not a practical proposition. The FPGA manufacturers (and their customers) want faster, smaller, better. Think of all the LUTs a

4A pfet would replace. Also, I'd trust Linear Tech. to do a much better job of a SMPS than an FPGA manufacturer.

Cheers, Syms.

p.s. I think that John's use of a diode drop for Vccaux is just fine. Simple and robust. I wonder if the temperature sensing diode that exists in some FPGAs could be used for this. ;-) (Is that repulsive enough?)

It seems to need a bazillion external parts. And we've been very unhappy lately about TI's products and support.

According to the guy I asked to look into this, The Switcher of Choice seems to be the LTC3411, single msop-packaged synchronous switcher. It behaves well in simulation with a small 4.7 uH inductor and ceramic output caps. We'll try some.

John

Jim,

I'm not sure that I totally agree with the "cocooned in the digital world" statement. If you're working on (modern) FPGAs and you don't know what Ldi/dt and Cdv/dt are then you are in BIG trouble (imho). The magnitude of these quantities, now, impact what needs to be done (not to mention transmission line effects).

It used to be that you could be a "digital" engineer and get by being ignorant of circuit design. Today, I don't think that a "digital-only" designer's designs will be successful.

Regarding power, I don't want X and A to add on-chip regulators to their FPGAs. It's hard enough trying to keep these buggers cool (as you've noted). I've had my fill of custom heat sinks, heat pipes, and 400lfm turbofans. I'll take the extra supplies as the lesser-of-two evils.

Bob