In order to make an FPGA, you need a huge number of flip flops to hold the state of all the bits in all of the look up tables that do the logic. This alone means that while they're really flexible at implementing any design, they will take much more power both in holding all those bits, and in the routing. Because FPGA implementations are necessarily bigger than equivalent hard wired logic, any connections will be longer, having more capacitance and resistance, which either requires more drive power to maintain a given clock rate, or slower speed to keep within the power/heat limits of a given chip.
So, while I applaud the creative thinking, it's not an original question, and there are good reasons it won't work out.
Semiconductors themselves are really difficult to scale down to smaller feature sizes, but fabs do often provide standard cells that simplify the process, at a cost of worse flexibility and performance. Even with standard cells, scaling up the number of LUTs linearly increases the routing possibilities quadratically, so scaling is not simple. On top of that, routing gets more difficult, as speed-of-light latency adds up.
Also, wires do already run separate from logic, with most process nodes having at least half dozen wire layers, if not double that, all to support a single layer of transistors.
There are open-source FPGA designs, but they are all for very simple FPGAs, due to the difficulty of scaling.
Correction. I used the wrong units. I stated about 7 wafers. But I fat-fingered
the calculation. I'm now showing about 1224 (1cm) chips giving 816 wafers. I
still claim that Terafab can handle that volume. TSMC numbers show that this
would be less than 2% of a single day run. Larger chip area would change the
numbers but I'm pretty sure that TSMC can currently handle the demand. Terafab
will likely be even better.
People have been hoping for "Open Source" hardware for as long as they've been hoping for Open Source gaming. 30 years plus, at this point. Sorry to disappoint, but, save for a few niche areas, it won't happen.
It's a real peachy sentiment, but if FPGAs were better than CPUs/GPGPUs then we'd already be using them. In practice, creating large FPGAs is a wasteful and expensive process that cannot be scaled to a non-research volume market and compete against ASICs or even general accelerators.
Nobody has done it ... therefore there is no reason to do it?
A highly parallel, high speed, massive I/O device to run your robot
that can be upgraded over the wire with new features? Nothing to see here.
Nobody would want to improve a product like a car or a robot "over the wire".
A billion robots whose hardware gets better overnight? Nah. Pointless.
The onus is on you to describe how the economics, fabrication and design is supposed to make sense here.
Because, again, it's a real peachy idea to want everyone to have a superpowered FPGA computer. The reason your iPhone isn't an FPGA comes down to the fact that you'd need ~20x the die space to implement a basic AArch64 core, which would drive the costs to ~$10,000-$20,000 per-handset and reduce your battery life down to the half-hour territory. If you combined the yearly capacity of TSMC, Samsung, Micron, Intel and CXMT, there would not be enough silicon to make a million competitive FPGAs, let alone a billion.
So, explain it. Why would robotics or smart car manufacturers prefer a 150 watt FPGA that is slower and 10x more expensive than a 15 watt CUDA SOC? AI is not a panacea, it cannot make these FPGAs consume less power than an ASIC or take up less die space than a RISC processor.
So you'd like a YC presentation deck? I'll call Paul Graham for a lunch meeting :-)
Point 1: I am creating an FPGA in the robot hand that learns repetitive motion
which will reduce the load on the main AI. The "Learned Hand" idea is based on
the human ability to "get learning into the hand", for example, doing typing
or playing piano. Notice that you don't have to think "adjust the fingers to
play a C chord". A similar system would allow stable walking with no AI input,
including fast recovery from tripping. It self-modifies (learns) over time.
Point 2: The phone ASIC makes sense because you have a very well specified,
very engineered environment that will be replicated a billion times and will
never require "over the air updates". Self-driving cars and robots will need
to adapt to external environments, new sensors, etc. "over the air". If you
replace a robot hand you'll only needs a few seconds of training to make it
repeat precise motions of the previous hand. Each hand could be factory
calibrated in the FPGA to a known-good standard.
Point 3: So you believe that Musk's Terafab won't be able to make a
million FPGAs? By my calculation you could fit about 128k FPGAs on a
16 inch fab wafer. So we're talking a run of about 7 wafers per million FPGAs.
Point 4a: Economics: There are very large companies that create FPGAs so
there is a known market. However the TAM (total addressible market) is much
larger and being held back due to high cost of initial learning (a high end
development board is several thousand dollars). Reducing this cost will
greatly expand the market (look at Raspberry Pi for an example).
Point 4b: Economics: FPGA programming in Verilog is not hard to learn
so anyone with a clever idea could build the circuits. Making it "nearly
free" (under $100) would open the TAM. Rather than the artificial paywall
the company makes money licensing the designs which is how ARM got rich.
Another revenue path is consulting on board design.
Point 5: Fabrication: A low-cost large LUT chip costs nearly the same as a
low-cost trivial LUT chip when factoring all of the costs. Since LUTs are
replicated once designed adding more LUTs to the design is a minimal cost.
Point 6a: Design: The key issue in design comes down to timing and routing.
The timing tree gets larger so clock skew becomes an issue. But this can
be overcome in a large FPGA by arranging clock domains so processes can
run on parallel clocks, in my case one per robot finger.
Point 6b: Design: Current FPGA designs use a lot of chip real estate with
wires and switching. I suggested a novel design that would first layer the
routing wires and then put the LUTs in a second layer. While this will
make the design and fabrication more expensive it will shrink the chip
(thus more per wafer) and improve the crossbar issues. Indeed, Intel has
been creating a "back of wafer" power distribution. Something similar for
wire routing would be a serious breakthrough. Or even better, use the new
"gate all around" technology to create "a vertical LUT" with minimal space.
Point 7: Creativitiy: Musk is putting hardware in space. You could easily
create replicated FPGAs and leave them unpowered. If an FPGA fails due to
radiation you switch it off and use a dark hardware replacement. If the
failure is "we found a faster algorithm" you can do "over the 'air'" updates.
Point 8: As Musk would say: this is only a matter of moving atoms and electrons.
There is no physical reason this can't be done.
etc. Now that I've done my part perhaps you could refute it point-by-point?
1. This is a software-domain issue that can be accelerated by off-the-shelf SIMD solutions, CUDA, CPUs and ASICs faster than an expensive FPGA.
2. There is not enough silicon in the world to make FPGA-only SOCs for the majority of smartphone customers.
3. Yes. Your calculation is not accounting for the size that would be necessary to make the FPGA competitive with CUDA/RISC and binning.
4A. Addressable market is not equivalent to demand.
4B. ARM gets screwed over by their biggest customers, and their smaller customers get away with using unlicensed clone chips. Their IP is not CUDA-level lucrative.
5. LUTs are not a bottleneck for mass-producing FPGAs, logic is.
6A. At extreme cost, which your "addressable market" is not interested in recouping. Again, look at the current demand for giant FPGAs.
6B. Even with Gates-All-Around design, you are still creating FPGAs that are gigantic and redundant compared to cheaper ASIC, GPGPU or CPU/SIMD chips.
7. Why do you think Elon Musk is choosing GPUs instead of FPGAs, then? Do you think he forgot that FPGAs existed when evaluating space-bound hardware?
8. Moving atoms and electrons is the easy part, doing it cheaply/efficiently/wastelessly is an asymptote challenge. Itanium and Xeon Phi died for a reason.
ok. I made a suggestion not a pitch deck. All ideas start somewhere.
There are always reasons ideas won't work but nothing seems to be
a fundamental show-stopper. Silicon, for example, is the second most
abundant element in the earth's crust so your point 2 seems questionable.
(Yes, I know about Spruce Pine and the wafer chain, etc.). I did not
suggest using FPGAs in a cell phone.
Running an FPGA seems less power hungry than running VLA chains to
generate fine motor actions based on a dozen finger sensor tactile
feedback signals arriving every few nanoseconds and responding in
real time, not to mention the lag going from finger to brain.
Offloading that task to intelligent local loops seems less power
hungry as the VLA only needs to know about exceptions, some of
which can even be "learned local handling" without the VLA (or
so goes my research goal).
Anyway either some person will see an opportunity or they won't.
I'm quite happy pointlessly hacking "Learned Hands" for robots.
So this is the end of the debate on my part. Thanks for the feedback.
In order to make an FPGA, you need a huge number of flip flops to hold the state of all the bits in all of the look up tables that do the logic. This alone means that while they're really flexible at implementing any design, they will take much more power both in holding all those bits, and in the routing. Because FPGA implementations are necessarily bigger than equivalent hard wired logic, any connections will be longer, having more capacitance and resistance, which either requires more drive power to maintain a given clock rate, or slower speed to keep within the power/heat limits of a given chip.
So, while I applaud the creative thinking, it's not an original question, and there are good reasons it won't work out.
Semiconductors themselves are really difficult to scale down to smaller feature sizes, but fabs do often provide standard cells that simplify the process, at a cost of worse flexibility and performance. Even with standard cells, scaling up the number of LUTs linearly increases the routing possibilities quadratically, so scaling is not simple. On top of that, routing gets more difficult, as speed-of-light latency adds up.
Also, wires do already run separate from logic, with most process nodes having at least half dozen wire layers, if not double that, all to support a single layer of transistors.
There are open-source FPGA designs, but they are all for very simple FPGAs, due to the difficulty of scaling.
Correction. I used the wrong units. I stated about 7 wafers. But I fat-fingered the calculation. I'm now showing about 1224 (1cm) chips giving 816 wafers. I still claim that Terafab can handle that volume. TSMC numbers show that this would be less than 2% of a single day run. Larger chip area would change the numbers but I'm pretty sure that TSMC can currently handle the demand. Terafab will likely be even better.
People have been hoping for "Open Source" hardware for as long as they've been hoping for Open Source gaming. 30 years plus, at this point. Sorry to disappoint, but, save for a few niche areas, it won't happen.
It's a real peachy sentiment, but if FPGAs were better than CPUs/GPGPUs then we'd already be using them. In practice, creating large FPGAs is a wasteful and expensive process that cannot be scaled to a non-research volume market and compete against ASICs or even general accelerators.
Nobody has done it ... therefore there is no reason to do it?
A highly parallel, high speed, massive I/O device to run your robot that can be upgraded over the wire with new features? Nothing to see here. Nobody would want to improve a product like a car or a robot "over the wire".
A billion robots whose hardware gets better overnight? Nah. Pointless.
The onus is on you to describe how the economics, fabrication and design is supposed to make sense here.
Because, again, it's a real peachy idea to want everyone to have a superpowered FPGA computer. The reason your iPhone isn't an FPGA comes down to the fact that you'd need ~20x the die space to implement a basic AArch64 core, which would drive the costs to ~$10,000-$20,000 per-handset and reduce your battery life down to the half-hour territory. If you combined the yearly capacity of TSMC, Samsung, Micron, Intel and CXMT, there would not be enough silicon to make a million competitive FPGAs, let alone a billion.
So, explain it. Why would robotics or smart car manufacturers prefer a 150 watt FPGA that is slower and 10x more expensive than a 15 watt CUDA SOC? AI is not a panacea, it cannot make these FPGAs consume less power than an ASIC or take up less die space than a RISC processor.
So you'd like a YC presentation deck? I'll call Paul Graham for a lunch meeting :-)
Point 1: I am creating an FPGA in the robot hand that learns repetitive motion which will reduce the load on the main AI. The "Learned Hand" idea is based on the human ability to "get learning into the hand", for example, doing typing or playing piano. Notice that you don't have to think "adjust the fingers to play a C chord". A similar system would allow stable walking with no AI input, including fast recovery from tripping. It self-modifies (learns) over time.
Point 2: The phone ASIC makes sense because you have a very well specified, very engineered environment that will be replicated a billion times and will never require "over the air updates". Self-driving cars and robots will need to adapt to external environments, new sensors, etc. "over the air". If you replace a robot hand you'll only needs a few seconds of training to make it repeat precise motions of the previous hand. Each hand could be factory calibrated in the FPGA to a known-good standard.
Point 3: So you believe that Musk's Terafab won't be able to make a million FPGAs? By my calculation you could fit about 128k FPGAs on a 16 inch fab wafer. So we're talking a run of about 7 wafers per million FPGAs.
Point 4a: Economics: There are very large companies that create FPGAs so there is a known market. However the TAM (total addressible market) is much larger and being held back due to high cost of initial learning (a high end development board is several thousand dollars). Reducing this cost will greatly expand the market (look at Raspberry Pi for an example).
Point 4b: Economics: FPGA programming in Verilog is not hard to learn so anyone with a clever idea could build the circuits. Making it "nearly free" (under $100) would open the TAM. Rather than the artificial paywall the company makes money licensing the designs which is how ARM got rich. Another revenue path is consulting on board design.
Point 5: Fabrication: A low-cost large LUT chip costs nearly the same as a low-cost trivial LUT chip when factoring all of the costs. Since LUTs are replicated once designed adding more LUTs to the design is a minimal cost.
Point 6a: Design: The key issue in design comes down to timing and routing. The timing tree gets larger so clock skew becomes an issue. But this can be overcome in a large FPGA by arranging clock domains so processes can run on parallel clocks, in my case one per robot finger.
Point 6b: Design: Current FPGA designs use a lot of chip real estate with wires and switching. I suggested a novel design that would first layer the routing wires and then put the LUTs in a second layer. While this will make the design and fabrication more expensive it will shrink the chip (thus more per wafer) and improve the crossbar issues. Indeed, Intel has been creating a "back of wafer" power distribution. Something similar for wire routing would be a serious breakthrough. Or even better, use the new "gate all around" technology to create "a vertical LUT" with minimal space.
Point 7: Creativitiy: Musk is putting hardware in space. You could easily create replicated FPGAs and leave them unpowered. If an FPGA fails due to radiation you switch it off and use a dark hardware replacement. If the failure is "we found a faster algorithm" you can do "over the 'air'" updates.
Point 8: As Musk would say: this is only a matter of moving atoms and electrons. There is no physical reason this can't be done.
etc. Now that I've done my part perhaps you could refute it point-by-point?
1. This is a software-domain issue that can be accelerated by off-the-shelf SIMD solutions, CUDA, CPUs and ASICs faster than an expensive FPGA.
2. There is not enough silicon in the world to make FPGA-only SOCs for the majority of smartphone customers.
3. Yes. Your calculation is not accounting for the size that would be necessary to make the FPGA competitive with CUDA/RISC and binning.
4A. Addressable market is not equivalent to demand.
4B. ARM gets screwed over by their biggest customers, and their smaller customers get away with using unlicensed clone chips. Their IP is not CUDA-level lucrative.
5. LUTs are not a bottleneck for mass-producing FPGAs, logic is.
6A. At extreme cost, which your "addressable market" is not interested in recouping. Again, look at the current demand for giant FPGAs.
6B. Even with Gates-All-Around design, you are still creating FPGAs that are gigantic and redundant compared to cheaper ASIC, GPGPU or CPU/SIMD chips.
7. Why do you think Elon Musk is choosing GPUs instead of FPGAs, then? Do you think he forgot that FPGAs existed when evaluating space-bound hardware?
8. Moving atoms and electrons is the easy part, doing it cheaply/efficiently/wastelessly is an asymptote challenge. Itanium and Xeon Phi died for a reason.
ok. I made a suggestion not a pitch deck. All ideas start somewhere.
There are always reasons ideas won't work but nothing seems to be a fundamental show-stopper. Silicon, for example, is the second most abundant element in the earth's crust so your point 2 seems questionable. (Yes, I know about Spruce Pine and the wafer chain, etc.). I did not suggest using FPGAs in a cell phone.
Running an FPGA seems less power hungry than running VLA chains to generate fine motor actions based on a dozen finger sensor tactile feedback signals arriving every few nanoseconds and responding in real time, not to mention the lag going from finger to brain.
Offloading that task to intelligent local loops seems less power hungry as the VLA only needs to know about exceptions, some of which can even be "learned local handling" without the VLA (or so goes my research goal).
Anyway either some person will see an opportunity or they won't. I'm quite happy pointlessly hacking "Learned Hands" for robots.
So this is the end of the debate on my part. Thanks for the feedback.
I misread 2), but my answer for the "robotic hand" application is the exact same as my response to 1).