Power supply for FPGA

[joniengr081]

Hi, I have searched and found that there are several options for the power supply for FPGA Kintex Ultra Scale. To find the optimum is a challenge. One option is to use step down buck converters and other option is to have several single channel buck converters. I appreciate if anyone come with a suggestion regarding selection of buck converters.

[AndyC_772]

I would have said Enpirion, who did a fantastic range of fully integrated converters that were specifically designed for exactly this purpose.

They were bought by Altera, who bought them for this very reason.

They in turn were bought by Intel, who didn't care about that product line at all, tried and failed to sell it off, then discontinued it completely with no warning and no opportunity for a last time buy. Not that I'm bitter about that or anything 

If you find a suitable alternative - I know TI do something similar, but they're not available to buy last time I checked either - then do please let me know.

[jayk]

There are lots of choices.  Have a look at the Xilinx eval board designs.  Googling 'ultrascale power supply' turned up a bunch of options.  Also there's some helpful free tools out there (LTpowerCAD, LTpowerPlanner, TI webench). Xilinx UG583/UG1085 are helpful.

Whichever way you go, be sure to pay attention to sequencing.

[EverydayMuffin]

Monolithic Power Systems have a good selection and have done some Kintex US reference designs.

https://www.monolithicpower.com/kintex-ultrascale-reference-design

I believe you can upload your XPE file to their website and they will suggest parts.
https://www.monolithicpower.com/contact/xilinx-xpe-file-upload.html

[joniengr081]

How about Analog, Maxim and NXP. They also have dc-dc buck converters might be suitable for FPGA power supply. Any comments on these manufacturers ? 

[Someone]

https://www.eevblog.com/forum/fpga/what-is-the-standard-procedure-to-design-the-power-delivery-network/

How about Analog, Maxim and NXP. They also have dc-dc buck converters might be suitable for FPGA power supply. Any comments on these manufacturers ?

There are another dozen or so manufacturers you still have not mentioned, so possibly thousands of parts....  but still no mention of what you are actually looking for. Low cost? small size? high efficiency? multisource? input and output specifications?

The manufacturer of the power chips is the least of your worries.

[strahd_von_zarovich]

Well, of course using modules which provide all the voltages you need is very tempting. However, when you can't buy them you are in very big trouble. I had to redesign FPGA boards because of chip shortage even if I used discrete components. 
However, If you use discrete components, at least you can produce them with small changes. So, use a regulator for 1V, and use another regulator for 3.3V for example, there are lots of options when you do this. You can use any voltage regulator which has enable, power good pins and enough performance(voltage tolerance, current capability, noise specifications).
Try to find footprint compatible alternatives and try to use standard packages. Companies like Infenion produce very good regulators but their packages...  , you can't find any alternative to them.

Best regards.

[luudee]

I don't think there is a way around DC/DC converters for internal power rails of modern chips. I have
just finished a design where we needed both 0.8V and 1.0V, each roughly supplying 20A.
The decision really boils down to:

A. Use of the shelf modules
   + easy to use
   + pre-made, with all proper components
   -  Expensive
   -  Can be bulky

B. Design your own
   + much cheaper
   + more flexibility
   + smaller, more compact
   -  Must be very careful designing the circuit, read and follow the data sheet and app notes
   -  PIA to select the proper components
   -  PCB layout can be tricky

 Most use cases do not need PMBUS (and similar) programming interfaces. I usually stay
away from them, as the voltages in our application do not vary.

If your project is small qtty, go with option A, if you will be doing mass production, I would recommend
option B. I also usually make a small test PCBs and evaluate my power supplies before making the
main PCB with all the expensive components (e.g. FPGA).

Good Luck !
luudee

[asmi]

That FPGA can consume tens of amps of current on it's Vccint rail, and if my memory serves me has ±3% regulation requirement over PVT. These are fairly strict requirements, so if you've never designed such high-current applications, you will be best served by first designing a small PCB just for this converter circuit, and you will be able to characterize it to confirm that it meets specifications.

Also don't forget about heat management - 90% efficient circuit at 20A@1V will dissipate 2 watts of heat, there are usually several DC-DC converters on a typical FPGA board which also generate heat, and FPGA itself can get very toasty (even low-end Artix devices can get very hot and require a heatsink in certain scenarios - I've seen A35T heating up to over 60°C with a small heatsink!), so you have to make sure the board will not fry itself once you load it up and let it run for some time. Since high layer count PCB are still rather expensive, you will be hard-pressed to make your design as compact as possible, making heat management even more critical.

[asmi]

Most use cases do not need PMBUS (and similar) programming interfaces. I usually stay away from them, as the voltages in our application do not vary.

Programmable devices can be useful for margining to increase precision of your regulation which will naturally vary board-to-board due to tolerances of components, or if you SoC supports multiple power modes (many high-end MCUs and a quite a bit of CPU SoCs support that).

If your project is small qtty, go with option A, if you will be doing mass production, I would recommend
option B. I also usually make a small test PCBs and evaluate my power supplies before making the
main PCB with all the expensive components (e.g. FPGA).

Using modules also tend to allow for more compact solutions because they typically take less PCB area then equivalent "discrete" converters, so if solution size matters to your design, modules are a way to go. They also tend to be better in terms of EMI.

[luudee]

Most use cases do not need PMBUS (and similar) programming interfaces. I usually stay away from them, as the voltages in our application do not vary.

Programmable devices can be useful for margining to increase precision of your regulation which will naturally vary board-to-board due to tolerances of components, or if you SoC supports multiple power modes (many high-end MCUs and a quite a bit of CPU SoCs support that).

If you know what you are doing, you will be able to design proper supplies that exceed
the requirements, without having to "tune" your DC/DC converters. Imagine having
to tune 1000 PCBs, lol

If your project is small qtty, go with option A, if you will be doing mass production, I would recommend
option B. I also usually make a small test PCBs and evaluate my power supplies before making the
main PCB with all the expensive components (e.g. FPGA).

Using modules also tend to allow for more compact solutions because they typically take less PCB area then equivalent "discrete" converters, so if solution size matters to your design, modules are a way to go. They also tend to be better in terms of EMI.

Nonsense. Using discrete components you can usually make smaller DC/DC converters,
unless they do vertical stacking, which is dangerous due to cooling issues.


luudee

[asmi]

If you know what you are doing, you will be able to design proper supplies that exceed
the requirements, without having to "tune" your DC/DC converters. Imagine having
to tune 1000 PCBs, lol

No amount of "knowing what you're doing" (nor arrogance displayed here) is going to defeat automatic voltage margining using system controller MCU. Not to mention having support for advanced power management modes with reduced voltage for lower-power modes.

Nonsense. Using discrete components you can usually make smaller DC/DC converters,
unless they do vertical stacking, which is dangerous due to cooling issues.

Nonsense if what you've just said, what I said is called "experience".

[Someone]

If you know what you are doing, you will be able to design proper supplies that exceed
the requirements, without having to "tune" your DC/DC converters. Imagine having
to tune 1000 PCBs, lol

No amount of "knowing what you're doing" (nor arrogance displayed here) is going to defeat automatic voltage margining using system controller MCU. Not to mention having support for advanced power management modes with reduced voltage for lower-power modes.

Foreign language collision?
Voltage margin vs stability (phase) margin.

Tight voltage requirements on new parts make it harder and harder to achieve without some active trim/calibration at least for static errors, programmable regulators are becoming the cheaper solution.

Nonsense. Using discrete components you can usually make smaller DC/DC converters,
unless they do vertical stacking, which is dangerous due to cooling issues.

Nonsense if what you've just said, what I said is called "experience".

The (surface mount) integrated power modules use the dense packing to increase operating frequencies, 3,4,5MHz and climbing which is impractical with discrete parts. Their power densities are hard to beat.

[asmi]

Foreign language collision?
Voltage margin vs stability (phase) margin.

I was mostly using TI parts when I was beginning, so I tend to use their terminology. "Margining" is an ability to slightly adjust reference voltage (the range varies from part to part, but usually covers at least ±5%) and thus indirectly adjust output voltage, this is in contrast to programmable voltage, in which case you can set output voltage directly in a more-or-less wide range.

The (surface mount) integrated power modules use the dense packing to increase operating frequencies, 3,4,5MHz and climbing which is impractical with discrete parts. Their power densities are hard to beat.

Yeah, which incidentally often introduces a problem of having to deal with a lot of heat localized on a relatively small area.

[joniengr081]

I was looking at the schematic of a development board which is Kintex UltraScale KCU105. I have attached a page for VCCINT which look like 40 A power supply. There are some questions. Is that a constant or transitional current in the beginning ?
Can this DC-DC converter MAX15301 which they use for VCCINT can deliver that much current ? If we need to design a power supply for Kintex UltraScale then what are the options for VCCINT ? I mean which chips are recommended for that purpose.

[luudee]

I was looking at the schematic of a development board which is Kintex UltraScale KCU105. I have attached a page for VCCINT which look like 40 A power supply. There are some questions. Is that a constant or transitional current in the beginning ?
Can this DC-DC converter MAX15301 which they use for VCCINT can deliver that much current ? If we need to design a power supply for Kintex UltraScale then what are the options for VCCINT ? I mean which chips are recommended for that purpose.

The Max15301 is just a controller. It by itself, can not deliver 40A current. The external switching FETs, however, help it to provide high current.

I constantly design high current DC/DC supplies like this one without the need for "tuning".

luudee

[joniengr081]

Thanks for your reply. Then I think this chip supplies the gate current to the power FETs. I will try to find those in the schematic.

[joniengr081]

Hi,

It's not easy to understand this power supply for VCCINT. The four FETs are rated around 30 A each but how are they connected is still a mystery to understand. These are N channel which means that when there is a positive voltage at the gate they will conduct. Two two FETs FDMS8018 are parallel and the other two FETs  FDMS8558S are also parallel but then they are connected in series, two parallel on the top and then two parallel on the bottom. The Gate voltage are DL and DH. What is the function of them in MAX15301. How the 1V is produced and what is the function of DCRN and DCRP ?

[langwadt]

Hi,

It's not easy to understand this power supply for VCCINT. The four FETs are rated around 30 A each but how are they connected is still a mystery to understand. These are N channel which means that when there is a positive voltage at the gate they will conduct. Two two FETs FDMS8018 are parallel and the other two FETs  FDMS8558S are also parallel but then they are connected in series, two parallel on the top and then two parallel on the bottom. The Gate voltage are DL and DH. What is the function of them in MAX15301. How the 1V is produced and what is the function of DCRN and DCRP ?

it's a buck converter https://en.wikipedia.org/wiki/Buck_converter . The top fets are the switch, the bottom fets the "diode"
DCRN and DCRP measures current via the voltage across the inductor or a current sense resistor

[joniengr081]

I have found that "VCCINT_FPGA" is actually is the power supply for VCCINT and VCCINT_IO Banks in the schematic of KCU105.

[joniengr081]

DH and DL are high and low side FET Gate Drivers but the four FETs: 2 x FDMS8018 and 2 x FDMS8558S are N-Channels, right ? I still refer to the same schematic in attachment.

The output pins are OUTP (27) and OUTP (26). I don't understand how the output current is sensed on DCRN (31) and DCRP (30).

VCC12_VCCINT is the power from main supply 12 V but how we get VCCINT ? Is that input supply ? if yes then where it is generated ?

[joniengr081]

I also don't understand about VCCINT_FET_SWITCH, is this switching ? if yes then where it is coming.

[langwadt]

I also don't understand about VCCINT_FET_SWITCH, is this switching ? if yes then where it is coming.

it is the mid point between the two sets of fets, connected to the inductor L57

[joniengr081]

Yes, it is connected to the inductor L57 but other than that it is not connected any other place, right ? I mean this is not driven at any other place.

I also don't get the input supply to that chip. In addition to the 12 V, do we need to provide PWM signal as an input because it look like a switch 

[langwadt]

Yes, it is connected to the inductor L57 but other than that it is not connected any other place, right ? I mean this is not driven at any other place.

I also don't get the input supply to that chip. In addition to the 12 V, do we need to provide PWM signal as an input because it look like a switch

you need to take about 20 steps back and start by learning how to read a schematic, what a buck converter is and then read the datasheet, your questions make no sense