Starlink Internet Connection equipment - Home/Office user

[woods170]

Do you suppose Greg Wyler is the only one able to solve this problem?

I'd naively figured it was still unsolved. Guess I was wrong about that.

It is one of the reasons why I pointed out that Tim Farrar is a completely unreliable source with regards to insight into Starlink. He estimates a single Starlink terminal to cost between $1,500 and $2,000. That estimate is not supported by any facts.

there is no point in discussing  with Tim as he is just repeating what other profis in the industry have said. The terminal has 1600 chips. There is information that this is As-Ga MMIC (Monolithic microwave integrated circuit)
(https://www.digitimes.com/news/a20200820PD207.html), WIN Semiconductors (https://www.winfoundry.com/en-US) Taiwan.

Emphasis mine.

This goes to show that you don't know what you are talking about. As-Ga MMIC are used in the Starlink satellites, not in the end-user terminals.


Edit: darn, I really should read to the end of the thread before responding. LouScheffer explained the very same thing just 30 minutes after I went to sleep. Darn time-zones...

[Kang54]

Here's a full take-down of Dishy. Go to 30:45 to see the inside of the metal shield, and to 33:10 to see the actual face itself. 35:30 onward is the backside of the face.

[RotoSequence]

Here's a full take-down of Dishy. Go to 30:45 to see the inside of the metal shield, and to 33:10 to see the actual face itself. 35:30 onward is the backside of the face.

Yeeeaaah, this isn't in the ballpark of a $15 part.

[ncb1397]

Here's a full take-down of Dishy. Go to 30:45 to see the inside of the metal shield, and to 33:10 to see the actual face itself. 35:30 onward is the backside of the face.

Yeeeaaah, this isn't in the ballpark of a $15 part.

$2400 for the kit (or a part of the kit) at a volume of 1 million units contracted out to ST Microelectronics (There are ST Microelectronics chips plastered all over the board). Not surprisingly, Hughes Network Systems/Echostar who has potential suppliers for satellite receivers wasn't far off. Aggregate size of the contract also roughly matches the size of the last equity raise conducted by SpaceX a few months before Beta rollout.

https://seekingalpha.com/news/3639484-stmicroelectronics-shares-gain-on-reported-spacex-production-agreement-bi

[RedLineTrain]

I don't know whether to believe Business Insider's sources.  Looking at that board from the teardown, the quantities for these custom chips must in some cases number in the billions.  But even if true, we should expect the next four million dishes to be quite a bit cheaper apiece.

It's a user terminal that will scare off competitors, that's for sure.  They're playing for keeps.

[thirtyone]

Here's a full take-down of Dishy. Go to 30:45 to see the inside of the metal shield, and to 33:10 to see the actual face itself. 35:30 onward is the backside of the face.

Yeeeaaah, this isn't in the ballpark of a $15 part.

$2400 for the kit (or a part of the kit) at a volume of 1 million units contracted out to ST Microelectronics (There are ST Microelectronics chips plastered all over the board). Not surprisingly, Hughes Network Systems/Echostar who has potential suppliers for satellite receivers wasn't far off. Aggregate size of the contract also roughly matches the size of the last equity raise conducted by SpaceX a few months before Beta rollout.

https://seekingalpha.com/news/3639484-stmicroelectronics-shares-gain-on-reported-spacex-production-agreement-bi

I was really curious about this, but didn't have access to the article. Unfortunately it sounds like a load of BS from someone who doesn't understand electronics, unless this is getting warped from the original article. STMicro is NOT a contract manufacturer - they make ICs. They would not be aware of the total assembly cost or manufacturing cost of the whole antenna, just the components that make it. In fact I find it far more likely that they are actually providing foundry services to SpaceX and that total is related to the NREs for their custom masks. These get paid for over the lifetime of the product, and maybe there's some estimate of what that lifetime is (it's way more than one million, I'm sure). There's something that really doesn't make sense if "STMicro" is the "contract manufacturer"

If the YouTube reviewer took dimensions on the 500+ ICs onboard, it's possible to get a max envelope size of the internal die on each chip (I could probably guess by pictures but...time). The big question then is whether it's a CMOS or III-V (GaAs, GaN, etc.). If it's CMOS, I'm a lot more familiar and we're looking at 10 cents per mm^2-ish at very high volume, and you can add up the IC cost. Assembly is nontrivial, usually can make some estimates there too. Related to board area and placement time.

Knowing SpaceX, this is an incredibly early prototype and at this quantity they may want to consider a custom line to package that PCB. I suspect they would want to remove a lot of unnecessary components for their real device (do they really need GPS? I think the Starlink team already said 'no') I'm guessing that's actually the most major component of their next major beta. Anyone want to tweet at Elon and confirm? (that their next rollout is largely related to the next cost-down revision of their dish).

[thirtyone]

I was looking into STMicro and indeed I'm pretty sure they don't really do electronics assembly (may have actually asked, even, for a previous project). It's also unclear if they really do RFICs fabs that aren't on silicon - they make GaN and SiC transistors though (the latter are in Teslas). Makes me strongly suspect SpaceX is actually using STMicro as their fab for their custom silicon on the dish. I also doubt the silicon itself costs that at the one million dish (which is like 500 million parts!) volume. Might be some measure of future business potential or honestly just wrong. There's substrates and assembly that are not cheap either for a product like this.

[LouScheffer]

That teardown was very helpful.  We now know the main architecture:

There is presumably a master processor.  Then, for the antenna, there are 72 controllers (big chips).  Each controls 8 mini chips.  Each mini chip drives/receives from 3 antennas.  Thus there are 72 controllers, 576 small chips, and 1728 antennas.

Each small chip would need 3 low noise amplifiers, 3 transmit amplifiers, and phase shifting networks for each  (2x that if they are going to use both polarizations at the same time, but I'm guessing not).  I'm guessing the small chips would also contain up/down mixers, since shlepping 12/14 GHz signals around normally requires the use of fancy substrate materials (expanded teflon, etc.), and this looks like a regular fiberglass board.  (It needs to be a regular fiberglass board or they would blow their whole budget on just the PCB.)

Getting the cost down will depend mainly on three factors - the cost of the small RF chip (576 of them), the cost of the large controller chip (72 of them), and the cost of the PCB.  Everything else looks like stock consumer electronics items and just a few per antenna.

[Robotbeat]

Do you suppose Greg Wyler is the only one able to solve this problem?

I'd naively figured it was still unsolved. Guess I was wrong about that.

It is one of the reasons why I pointed out that Tim Farrar is a completely unreliable source with regards to insight into Starlink. He estimates a single Starlink terminal to cost between $1,500 and $2,000. That estimate is not supported by any facts.

there is no point in discussing  with Tim as he is just repeating what other profis in the industry have said. The terminal has 1600 chips. There is information that this is As-Ga MMIC (Monolithic microwave integrated circuit)
(https://www.digitimes.com/news/a20200820PD207.html), WIN Semiconductors (https://www.winfoundry.com/en-US) Taiwan.
And the fact that after the words of Greg Wyler in 2019 (who personally invested in this company and possibly wanted to attract other investors) nothing happened about "the $ 15 miracle in 2020" , tells me that $ 2,000 is possibly a bottom estimate ..
I have been hearing from new companies in their first year for 5 years now about the future breakthrough and prices of 1000 USD per ESA FPA terminal. Only after 2 years the authors of these presentations cannot be found anywhere ..

And my congratulations to Space X for a great job - the terminal is really great!!!!!

If they're running Gallium Arsenide ICs on those things, they're going to be bound by the fabrication cost of the wafer and the size of the IC. A fully developed GaAs wafer, I believe, runs around $1000, but I don't really know for sure; that might be too low since silicon CMOS on moderate processes run around that price. We need IC sizes and wafers to know the real cost of the antenna bits.

Cursory research on the size of antenna elements suggest that a 6" wafer should yield something around 8000 devices, for a guesstimated cost to SpaceX of something between 10 and 20 cents per IC.

So between $150 and $300 for the whole array.

[Robotbeat]

That teardown was very helpful.  We now know the main architecture:

There is presumably a master processor.  Then, for the antenna, there are 72 controllers (big chips).  Each controls 8 mini chips.  Each mini chip drives/receives from 3 antennas.  Thus there are 72 controllers, 576 small chips, and 1728 antennas.

Each small chip would need 3 low noise amplifiers, 3 transmit amplifiers, and phase shifting networks for each  (2x that if they are going to use both polarizations at the same time, but I'm guessing not).  I'm guessing the small chips would also contain up/down mixers, since shlepping 12/14 GHz signals around normally requires the use of fancy substrate materials (expanded teflon, etc.), and this looks like a regular fiberglass board.  (It needs to be a regular fiberglass board or they would blow their whole budget on just the PCB.)

Getting the cost down will depend mainly on three factors - the cost of the small RF chip (576 of them), the cost of the large controller chip (72 of them), and the cost of the PCB.  Everything else looks like stock consumer electronics items and just a few per antenna.

What's interesting to me is they put EVERYTHING on that PCB. In principle, should allow them to reduce assembly costs.

The PCB cost might not be too bad. PCBs scale pretty well, and companies work extremely hard to get cost of PCBs extremely low.

Wish the dude had counted the PCB layers and X-rayed it (and maybe uncapped the die...).

[RotoSequence]

What's interesting to me is they put EVERYTHING on that PCB. In principle, should allow them to reduce assembly costs.

The PCB cost might not be too bad. PCBs scale pretty well, and companies work extremely hard to get cost of PCBs extremely low.

Wish the dude had counted the PCB layers and X-rayed it (and maybe uncapped the die...).

They've got their work cut out for them if they want to reduce the costs of a board with 80,000 to 100,000 blind vias. 

[Tuna-Fish]

(2x that if they are going to use both polarizations at the same time, but I'm guessing not).

Yes, I believe the point of the polarization is to drive up aggregate bandwidth per area. For this, there is no requirement for an individual terminal to work with both polarizations at once.

[ncb1397]

Do you suppose Greg Wyler is the only one able to solve this problem?

I'd naively figured it was still unsolved. Guess I was wrong about that.

It is one of the reasons why I pointed out that Tim Farrar is a completely unreliable source with regards to insight into Starlink. He estimates a single Starlink terminal to cost between $1,500 and $2,000. That estimate is not supported by any facts.

there is no point in discussing  with Tim as he is just repeating what other profis in the industry have said. The terminal has 1600 chips. There is information that this is As-Ga MMIC (Monolithic microwave integrated circuit)
(https://www.digitimes.com/news/a20200820PD207.html), WIN Semiconductors (https://www.winfoundry.com/en-US) Taiwan.
And the fact that after the words of Greg Wyler in 2019 (who personally invested in this company and possibly wanted to attract other investors) nothing happened about "the $ 15 miracle in 2020" , tells me that $ 2,000 is possibly a bottom estimate ..
I have been hearing from new companies in their first year for 5 years now about the future breakthrough and prices of 1000 USD per ESA FPA terminal. Only after 2 years the authors of these presentations cannot be found anywhere ..

And my congratulations to Space X for a great job - the terminal is really great!!!!!

If they're running Gallium Arsenide ICs on those things, they're going to be bound by the fabrication cost of the wafer and the size of the IC. A fully developed GaAs wafer, I believe, runs around $1000, but I don't really know for sure; that might be too low since silicon CMOS on moderate processes run around that price. We need IC sizes and wafers to know the real cost of the antenna bits.

Cursory research on the size of antenna elements suggest that a 6" wafer should yield something around 8000 devices, for a guesstimated cost to SpaceX of something between 10 and 20 cents per IC.

So between $150 and $300 for the whole array.

I don't think so. The cheapest item ST Micro has on their website is a 2000 bit EEPROM RFID chip for $.22. Maybe in bulk you can get them for 10-20 cents, but these custom parts aren't likely to cost as much as their cheapest item.

Anyways, the parts are basically black boxes as they apparently don't show up in any catalog. The larger parts that were copied 72 times are roughly 1 cm square or 100 square millimeters. To put that into perspective, the A14 Bionic in the latest iphone die is smaller than that. No way to really know what is inside until somebody opens them, which is probably just a matter of time.

[LouScheffer]

That teardown was very helpful.  We now know the main architecture:

There is presumably a master processor.  Then, for the antenna, there are 72 controllers (big chips).  Each controls 8 mini chips.  Each mini chip drives/receives from 3 antennas.  Thus there are 72 controllers, 576 small chips, and 1728 antennas.

Each small chip would need 3 low noise amplifiers, 3 transmit amplifiers, and phase shifting networks for each  (2x that if they are going to use both polarizations at the same time, but I'm guessing not).

3 antenna per small chip makes lots of sense.  You want the chip to be as close as possible to each antenna.  In the diagram below, hexagons are the antennas and the green circles are the small RF chips.   Using this architecture you can get complete tiling of the array with very short RF connections.

Although representative, this is not exactly what SpaceX is doing.  I rotated the hex grid to make the antennas correspond with the tear-down video (with the flats on the left and right sides).  However this gives a grid for the small chips which does not align with the SpaceX placement.  I fiddled around for a while and could not get the grid SpaceX uses for the small chips.  But I suspect this is just insufficient imagination on my part.

[LouScheffer]

What's interesting to me is they put EVERYTHING on that PCB. In principle, should allow them to reduce assembly costs.

The PCB cost might not be too bad. PCBs scale pretty well, and companies work extremely hard to get cost of PCBs extremely low.

Wish the dude had counted the PCB layers and X-rayed it (and maybe uncapped the die...).

They've got their work cut out for them if they want to reduce the costs of a board with 80,000 to 100,000 blind vias. 

Maybe not.  Unlike (for example) a processor board, not every via needs to work.   They could probably lose a processor, and perhaps a few percent of the RF chips, and the system would still work fine.  You might be able to get the PCB cost down considerably if you don't need perfection.

[joek]

Beyond the view of UT costing is a macro view which IMO is missing here: Is the UT cost of $300 or $2000 material to Starlink's near term success.  IMO no (and those costs will undoubtedly go down significantly in the next couple years).  A few assumptions based on current state (at least as best I can determine).

For the in-space infrastructure segment...
1. Sat unit cost to manufacture is ~$500K (think that is conservative, other chatter indicates could be closer to ~$250K).
2. Internal SpaceX launch cost is ~35M for 60 sats (think that is conservative, other chatter indicates could be close to ~$30M).
2. A sat lifetime of 5 years or 60 months.
3. ...which gives us a cost of ~$9800/mo/sat

For the ground infrastructure segment...
1. Assume 40% of the cost of the sat constellation (admittedly a WAG, but think reasonably conservative).
2. ... which gives us a cost of ~$3900/mo/sat
3. ... and for a constellation of 1440 sats, ~19.6M/mo.

If all $100 of subscriber fees went towards the in-space and ground infrastructure, that would require ~196K subscribers @$100/mo to break even for the entire in-space constellation and supporting ground infrastructure.

So now lets look at buy-in and impact of subscriber terminal costs...
1. Assume actual manufacturing cost is ~$2K, offset by $500 purchase, with net deficit of $1500.
2. Assume net contribution of subscriber $100/mo is ~60% to pay down the UT (over 24 months) and ~40% to the service (in-space and ground infrastructure).[1]
3. ...which increases the break-even to ~525K subscribers @$100/mo.

In short, a bit of sensitivity analysis suggests current Starlink break-even is well within reason, but long term may be highly sensitive to UT cost (and of course subscriber retention).  But in the grand scheme of things I hope and expect that UT cost will decrease rapidly with scale.  In any case, the current debate about UT cost seems a tempest in a teapot.


[1] That was derived by splitting the infrastructure (space-ground) cost portion of a subscriber's $/mo to what could be contributed to paying off the UT deficit.  Obviously subscribers who cancel contracts early will have a negative impact... unless Starlink screws up, I hope and expect that to be in the noise.

[LouScheffer]

That teardown was very helpful.  We now know the main architecture:

There is presumably a master processor.  Then, for the antenna, there are 72 controllers (big chips).  Each controls 8 mini chips.  Each mini chip drives/receives from 3 antennas.  Thus there are 72 controllers, 576 small chips, and 1728 antennas.

Each small chip would need 3 low noise amplifiers, 3 transmit amplifiers, and phase shifting networks for each  (2x that if they are going to use both polarizations at the same time, but I'm guessing not).  I'm guessing the small chips would also contain up/down mixers, since shlepping 12/14 GHz signals around normally requires the use of fancy substrate materials (expanded teflon, etc.), and this looks like a regular fiberglass board.  (It needs to be a regular fiberglass board or they would blow their whole budget on just the PCB.)

Getting the cost down will depend mainly on three factors - the cost of the small RF chip (576 of them), the cost of the large controller chip (72 of them), and the cost of the PCB.  Everything else looks like stock consumer electronics items and just a few per antenna.

Upon closer inspection, I'm changing my estimates.  By counting, there are only 1563 antennas on the board.  Therefore each small chip cannot drive 3 antennas.   The same results can be obtained by superimposing the component side and antenna side, which shows each small chip drives 2 antennas.  It also shows that around the edges, many elements have no corresponding small chip.  This makes sense - you would want the outer edges antennas to be dummies, so that every active element is surrounded by 6 other antennas and sees the same environment.   Also, the digital circuits take some room, and it looks like the RF spots under them are also not driven.

So overall, I now think there are 72 big chips, each driving 8 small chips, each driving 2 antennas, for a total of 1152 active elements in the array.  The others, all round the edge, are dummies to preserve a uniform electromagnetic environment for the active elements, and to make room for the other needed circuits.

[guckyfan]

There was an infrared image that shows all dots equally heated. Probably each small chip drives 3 spots.

[LouScheffer]

There was an infrared image that shows all dots equally heated. Probably each small chip drives 3 spots.

Each small chip definitely drives 2 antennas.   You can see this in the picture below.  I took a snapshot of the copper side from the teardown video.  Then I rotated it as the presenter did as he turned it over.   Then I made the picture partially transparent and overlaid it on the component side, and scaled it to fit.  Below is a picture of two blocks of the array, with 2 controller chips driving 8 small chips, each driving 2 antennas.

I suspect the thermal image shows every dot since the PCB heats up more or less uniformly, but is isolated from the face by a foam spacer with a hexagonal grid, as shown later in the video.

[groknull]

That teardown was very helpful.  We now know the main architecture:

There is presumably a master processor.  Then, for the antenna, there are 72 controllers (big chips).  Each controls 8 mini chips.  Each mini chip drives/receives from 3 antennas.  Thus there are 72 controllers, 576 small chips, and 1728 antennas.

Each small chip would need 3 low noise amplifiers, 3 transmit amplifiers, and phase shifting networks for each  (2x that if they are going to use both polarizations at the same time, but I'm guessing not).  I'm guessing the small chips would also contain up/down mixers, since shlepping 12/14 GHz signals around normally requires the use of fancy substrate materials (expanded teflon, etc.), and this looks like a regular fiberglass board.  (It needs to be a regular fiberglass board or they would blow their whole budget on just the PCB.)

Getting the cost down will depend mainly on three factors - the cost of the small RF chip (576 of them), the cost of the large controller chip (72 of them), and the cost of the PCB.  Everything else looks like stock consumer electronics items and just a few per antenna.

Upon closer inspection, I'm changing my estimates.  By counting, there are only 1563 antennas on the board.  Therefore each small chip cannot drive 3 antennas.   The same results can be obtained by superimposing the component side and antenna side, which shows each small chip drives 2 antennas.  It also shows that around the edges, many elements have no corresponding small chip.  This makes sense - you would want the outer edges antennas to be dummies, so that every active element is surrounded by 6 other antennas and sees the same environment.   Also, the digital circuits take some room, and it looks like the RF spots under them are also not driven.

So overall, I now think there are 72 big chips, each driving 8 small chips, each driving 2 antennas, for a total of 1152 active elements in the array.  The others, all round the edge, are dummies to preserve a uniform electromagnetic environment for the active elements, and to make room for the other needed circuits.

I counted 1336 antennas.  (Double checking of counts and math welcome.)
Bottom group (344):
- rows of 8 to 18 antennas (incremented by 1)
- rows of 17 to 21 antennas (by 1)
- rows of 20 to 22 antennas (by 1)
- rows of 21 and 22 antennas

Center group (648):
- one row of 23 antennas
- 15 pairs of rows with 22 and 23 antennas in each row pair

Top group (344):
- same as the bottom group

The video indicates 79 larger chips, not 72.  I counted 79 as well.

That allows 632 smaller chips.  If two antennas per smaller chip, then 1264, which agrees well with 1336 minus border cells.  As you (I think) pointed out, some larger chips may have fewer than 8 smaller chips attached to them.

I agree that some of the edge cells are likely to be passive/grounded or parasitic to modify the beam edge characteristics.


EDIT: My counts and calculations in this post are wrong.  See the post several down for corrections.
https://forum.nasaspaceflight.com/index.php?topic=50047.msg2159952#msg2159952