Your opinions on the FPGA market in the next few years

[TimNJ]

Hi guys,

I've done some FPGA development in my university courses with Xilinx dev boards. I've found it pretty much pain-free so I am considering buying a Xilinx dev board from Digilent to get started developing at home.

Once you pick a vendor, it doesn't necessarily mean you're locked into it, but moving around between the major players (say from Xilinx to Altera/Intel to Lattice etc.) definitely wastes time.

The Intel/Altera acquisition seems interesting but you always have to wonder how an acquisition might negative affect a business. On the other hand, Intel's cutting edge processes could be a huge benefit. I've heard rumors of a possible Qualcomm or Broadcom takeover of Xilinx but as of now, Xilinx is still an FPGA maker making FPGAs.

I also wonder if there's any chance that smaller players (or traditionally non-FPGA makers) may take a stab at making FPGAs, though I think there's pretty high barriers to entry at this point.

Anyway, this question half me trying to figure out what FPGA maker/toolchain to go with and half me being curious about the industry and where it's going. If you have any insight, it would be cool to hear!

[rstofer]

First, I think you need to separate design from manufacturing.  TSMC does the manufacturing for Altera and they do it for a lot of players.  There has been a tendency over the last 20 years to get out of the wafer fab end of production and to farm it out to foundaries, like TSMC.

http://www.tsmc.com/english/default.htm

I can't see the Intel/Altera venture being helpful to Altera.  Intel makes CPUs.  I'm a cynic but I don't see a happy future for Altera.

I'm hoping that Xilinx can maintain their independence.  Their stock price has steadily risen over the last 5 years and I hope they have taken measures to prevent a hostile takeover (poison pill).  Getting involved with Qualcomm or Broadcom would be the end of Xilinx.

As to Digilent:  I have been buying boards from them for the last 10 years or so.  I enjoy using their products but the takeover by National Instruments is something I'm wary of.  Nevertheless, I use their boards almost every day and I'm a huge fan of their Analog Discovery gadget.  Outstanding company!

[nctnico]

Definitely go with Altera. It is very likely that in a few years, Intel will integrate fabrics in their own processors.

You seem to have forgotten that Intel has failed many times trying to get into new markets. They have bought and sold numerous companies in the past because they where unable to turn them into a commercial succes. Another problem with Intel is that their x86 cores aren't as low power as ARM can be which basically prevents people from using x86 in mass produced embedded platforms.

[rstofer]

Hi guys,

I've done some FPGA development in my university courses with Xilinx dev boards. I've found it pretty much pain-free so I am considering buying a Xilinx dev board from Digilent to get started developing at home.

Once you pick a vendor, it doesn't necessarily mean you're locked into it, but moving around between the major players (say from Xilinx to Altera/Intel to Lattice etc.) definitely wastes time.

There is a learning curve for different devices and different toolchains but logic is logic.  The vast majority of VHDL or Verilog code is device/toolchain independent.  In the end, both products wind up doing the same thing.  If I were working in the field, I would spend time with both families and both toolchains.

[nctnico]

It depends. I have been 'Xilinx only' for over 15 years and never had the need to learn Altera or another brand although people say Xilinx' tools aren't the nicest ones to work with.

[TimNJ]

Definitely go with Altera. It is very likely that in a few years, Intel will integrate fabrics in their own processors.
Though Xilinx may complain (just like Mozilla complained when Bill Gates decided to offer IE for free with Windows), big companies like Windows/Intel have demonstrated several times that money can make illegal things legal (also, refer to Windows Media Player banned in EU, then EU quickly realized enforcing laws can be so expensive and how stupid they did).
So, unless Xilinx sells itself to a big processor player (maybe ARM or AMD), otherwise they face huge challenge from Intel.

PS. This is happening. Intel is said planning to put fabrics to Skylake or Kabylake dual socket chips.

Yeah from my perspective, it seems like they have an opportunity to dominate the data center/high performance computing market with tight integration of fabrics and processors. If they can make this all come to fruition (which I have my doubts), I wonder if they will leave behind the standalone FPGA business. i.e. there might be more money in data center hardware than selling FPGAs to say Rigol.

[TimNJ]

First, I think you need to separate design from manufacturing.  TSMC does the manufacturing for Altera and they do it for a lot of players.  There has been a tendency over the last 20 years to get out of the wafer fab end of production and to farm it out to foundaries, like TSMC.

http://www.tsmc.com/english/default.htm

I can't see the Intel/Altera venture being helpful to Altera.  Intel makes CPUs.  I'm a cynic but I don't see a happy future for Altera.

I'm hoping that Xilinx can maintain their independence.  Their stock price has steadily risen over the last 5 years and I hope they have taken measures to prevent a hostile takeover (poison pill).  Getting involved with Qualcomm or Broadcom would be the end of Xilinx.

As to Digilent:  I have been buying boards from them for the last 10 years or so.  I enjoy using their products but the takeover by National Instruments is something I'm wary of.  Nevertheless, I use their boards almost every day and I'm a huge fan of their Analog Discovery gadget.  Outstanding company!

Yes, I misspoke when calling a Xilinx an "FPGA maker". It would seem to me that Intel (as a fab) would be more expensive than TSMC and from what I've read, TSMC is outpacing Intel in terms of technology. I believe that they will get to 10nm before Intel. How much does that matter? I don't know.

Regarding Digilent: I was listening to a podcast (was it the Amp Hour?, I don't remember) and the CEO of Digilent explained that NI was letting them run almost completely independently as their markets don't quite overlap.

[TimNJ]

Definitely go with Altera. It is very likely that in a few years, Intel will integrate fabrics in their own processors.

You seem to have forgotten that Intel has failed many times trying to get into new markets. They have bought and sold numerous companies in the past because they where unable to turn them into a commercial succes. Another problem with Intel is that their x86 cores aren't as low power as ARM can be which basically prevents people from using x86 in mass produced embedded platforms.

http://www.theverge.com/2016/8/16/12507568/intel-arm-mobile-chips-licensing-deal-idf-2016

Looks like Intel is going to build ARM chips. But who's going to buy them?

[TimNJ]

and it seems that TSMC is outpacing Intel.

Not exactly. Intel's 14nm can accommodate more transistors per area than similar TSMC 16nm or GF 14 nm or Samsung 14 nm.
Not because TSMC has demonstrated low volume production of 10nm that makes it better than Intel -- maybe their 10nm is only as dense as Intel's 14nm.

Interesting. Is that what is referred to as pitch? I guess it's not apples-to-apples, as I have been assuming.

[TimNJ]

There is a learning curve for different devices and different toolchains but logic is logic.  The vast majority of VHDL or Verilog code is device/toolchain independent.  In the end, both products wind up doing the same thing.  If I were working in the field, I would spend time with both families and both toolchains.

I interned at a place this summer which used Altera and Xilinx about 50/50, so yeah it seems like a good engineer should be flexible enough to switch between vendors depending on what part meets your requirements. But I guess when it comes to home development, I'd like to stick with just one for now.

[rstofer]

Definitely go with Altera. It is very likely that in a few years, Intel will integrate fabrics in their own processors.

You seem to have forgotten that Intel has failed many times trying to get into new markets. They have bought and sold numerous companies in the past because they where unable to turn them into a commercial succes. Another problem with Intel is that their x86 cores aren't as low power as ARM can be which basically prevents people from using x86 in mass produced embedded platforms.

http://www.theverge.com/2016/8/16/12507568/intel-arm-mobile-chips-licensing-deal-idf-2016

Looks like Intel is going to build ARM chips. But who's going to buy them?

Chips are chips, cost is the determinant.  This is particularly true for ARM chips.  Who cares about the CPU, it's all about the on-chip peripherals.  The thing is, ARM chips are jelly bean parts.  They are made by the hundreds of millions by a dozen different companies.  It is very likely we will be down to two architectures:  The Intel x64 and ARM.  I don't see a happy future for Microchip or Atmel (owned by Microchip) as ARM prices drop to the point it is cost effective to use the ARM for low pincount applications.  We'll see...

[TimNJ]

and it seems that TSMC is outpacing Intel.

Not exactly. Intel's 14nm can accommodate more transistors per area than similar TSMC 16nm or GF 14 nm or Samsung 14 nm.
Not because TSMC has demonstrated low volume production of 10nm that makes it better than Intel -- maybe their 10nm is only as dense as Intel's 14nm.

Interesting. Is that what is referred to as pitch? I guess it's not apples-to-apples, as I have been assuming.

xx nm is not pitch, it is minimum feature size. You can not make everything down to minimum feature size.
So, some fabs will specify a transistor to have x times minimal feature size to be reliably made possible, while some other can push further down.
That's why you see although TSMC has reliable 16nm process, high end GPUs still face yield issue -- because chip designers have decided to risk yield to get higher density and performance, by negotiating a new process with reduced x for the same xx nm process.

Right, yeah I wasn't saying that 16nm was the pitch, but nonetheless thanks for good info. So many variables! Who know's where the industry is headed!

[rstofer]

Right, yeah I wasn't saying that 16nm was the pitch, but nonetheless thanks for good info. So many variables! Who know's where the industry is headed!

Smaller feature size to reduce capacitance and shorten line lengths resulting in higher speeds and less heat.  That's where it has been heading for decades.  I was told by a VP at National, way back in the late '70s, to "never underestimate technology!".  We were embarking on a project to build semiconductor RAM for IBM mainframes.  The kind of thing that would occupy a bay or two of mainframe cabinetry.  Of course, today, we all have 10 times that much memory on our tablets...

ETA:  Also use lower switching voltages.

[ehughes]

I have been using Xilinx for about 15 years.   Altera of and on.


Buy a 7 series Digilent board.     It is no big deal to switch when needed.     A job that would look at your Xilinx experience and not hire you because they are using Altera is someone you do not want to work for.

Just be flexible and be willing to switch. 

Xilinx has ruled the roost since the 80's.          They aren't going anywhere.

[TimNJ]

Right, yeah I wasn't saying that 16nm was the pitch, but nonetheless thanks for good info. So many variables! Who know's where the industry is headed!

Smaller feature size to reduce capacitance and shorten line lengths resulting in higher speeds and less heat.  That's where it has been heading for decades.  I was told by a VP at National, way back in the late '70s, to "never underestimate technology!".  We were embarking on a project to build semiconductor RAM for IBM mainframes.  The kind of thing that would occupy a bay or two of mainframe cabinetry.  Of course, today, we all have 10 times that much memory on our tablets...

ETA:  Also use lower switching voltages.

I meant "who knows who's going to get acquired or who's going to dominate the market in say 10 years". But technically speaking, yes of course, lower the capacitance and lower the switching time!

[TimNJ]

I have been using Xilinx for about 15 years.   Altera of and on.


Buy a 7 series Digilent board.     It is no big deal to switch when needed.     A job that would look at your Xilinx experience and not hire you because they are using Altera is someone you do not want to work for.

Just be flexible and be willing to switch. 

Xilinx has ruled the roost since the 80's.          They aren't going anywhere.

I've been looking around. Do you mean an Artix 7 board? At college, we used Spartan 3E with ISE but I think that might not be recommended for new designs. I do want to try out Vivado.

[rstofer]

For Vivado, the Artix 7 board is the place to start.
http://store.digilentinc.com/arty-board-artix-7-fpga-development-board-for-makers-and-hobbyists/

Notice that the board comes with a Vivado Design Suite voucher.  There are only a couple of guys on the forum that understand the licensing of Vivado but I think the voucher is important.  There is a WebPack version of Vivado but I think you get more with the voucher.  It's all a guess...

[filssavi]

The design vaucher was necessary before the introduction of the vivado webpack HLx, in fact the free vivado back then didn't include the ILA (debugger core) so the licenze war required, with vivado webpack HLx (HLS) it has changed and now the only difference between desygn and webpack licenses is the FPGA/SoC supporter, in fact the design versione can work with any product in the Xilinx lineup (series 7 and above as before) the webpack version however is limite to the lower end products (i think all the artix family and the zynq's that are based on a artix part, no or very limite kintex range and no virtex)

In fact Digilent is not selling the Vivado voucher any more, they now give you (for 10$ or thereabout) the licenze to SDSoC the new Xilinx ide to use the FPGA (the PL part of the zybo for sure, i don't know if ut works with non zybo parte) to accellerate software (openCL if i'm not mistaken but don't quote me on this)

[Scrts]

For automotive: BMW partnered with Intel for the next generation advanced-driving systems and pushing hard the Tier1 suppliers to implement Altera SoC FPGAs into the next designs. They see it very big in terms of Intel CPU+FPGA in the future.

[Xenoamor]

I'm personally finding Vivado to be much more enjoyable to use than ISE. These Artix 7 series chips really plow through DSP

[danadak]

Cypress has a line of parts with standard peripherals + a fabric that can be
programmed in Verilog or using a GUI tool of basic and expanded elements.
Series is called PSOC.

Parts have either an 8051 or ARM M0 or ARM M3 core. DMA so you can run
multiple tasks outside MCU. DSP engine to implement filters.

Plus lots of analog, OpAmps, SAR, DelSig, reference, mixer, comparator,
analog switch, DAC.......

Memory, RAM. FLASH, Emulated EE.

Lots of videos, cheap tools (SW free)........

http://www.cypress.com/products/microcontroller-mcu-and-programmable-system-chip-psoc-families



Regards, Dana.

[Muxr]

Intel's fabs historically and as it stands currently have the highest density. They are ahead of everyone else, the other major players are Samsung/GlobalFoundries and TSMC. Intel's 14nm FinFet process is denser than Samsung's 14nm.

Running fabs is expensive and the decline of PC means Intel has to expand their portfolio in order to keep those fabs making wafers. Their failure to break into mobile has forced their hand a bit, and I suspect is part of the reason they bought Altera.. allowing 3rd party ARM makers to use their fabs for the first time is also very telling. AMD just demoed what looks to be possibly a very competitive x86 core (Zen) as well. So basically Intel is under a lot of pressure.

So I would expect Altera to have an edge when it comes to gate density on their future high end FPGAs. I am not convinced about the wide adoption of CPU+FPGA hybrids until the software stack improves. The main audience for this stuff would be the data center folks (HPC), which have no FPGA experience. Nvidia has spent a long time developing the CUDA ecosystem to bridge this gap so I suspect a similar effort would be required to provide an easier API to tap into FPGA resources. The current FPGA workflow is also very static.. you write an HDL, compile it into bitstream and you rarely change it. In datacenter you want dynamism, you want to run different bitstream depending on need, unless you have a very specific acceleration in mind not being addressed by ASICs or CUDA/OpenCL.

Chipzilla (another name for Intel) also seems to be loosing their fab edge, the competition seems to be catching up for the first time I can remember: http://www.eetimes.com/document.asp?doc_id=1327087

In either case it will be interesting to see what happens.

[gardner]

Cypress ... PSOC

Cypress PSoC is certainly interesting, but the capabilities of its digital fabric is nowhere near that of an FPGA.  I'm not sure it's even on par with what you would do with CPLDs.  I have not seen any provision for synthesizing anything from HDL.  The toolchain is highly specific to the Cypress parts.

The analogue fabric in the Cypress PSoC is interesting though, and I thought I heard that more FPGAs were including analogue features.  I would have thought that would be a significant market trend.

[frogmaster]

Anyway, this question half me trying to figure out what FPGA maker/toolchain to go with and half me being curious about the industry and where it's going. If you have any insight, it would be cool to hear!

I don't really have any insight  , but I'll dump my thoughts here.

First of all, moore's law - most likely- is coming to an end and we're processing more and more stuff, e.g. Augmented reality stuff, Pokemon Go, Speech, AI, more video processing etc. So going for FPGAs is the right thing and I see all those ARM cores in mobiles, tablets, etc. getting 'FPGA-extensions' very soon (especially ZYNQ-like SoCs).

Not sure about how Intel will handle the situation with Altera, but they will have to come up with a toolchain that is pretty similar to what Xilinx is trying to push with their HLS and PYNQ stuff now- and I think this is because they're targeting all those ubiquitous gadgets more and more.
(See https://media.readthedocs.org/pdf/pynq/latest/pynq.pdf and https://translate.google.com/translate?hl=en&sl=zh-CN&u=http://xilinx.eetrend.com/news/10408&prev=search ).

So clearly the intent is to get regular programming code, i.e. MATLAB, OpenCL, C++ and other stuff compiling/synthesizing to HDL or PL directly, as fast and as seamlessly as possible. Getting to interact PL with the CPU is also quite interesting, so Xilinx have switched to AXI (ARM standard) pretty early and they seem to have a clear path in mind.

I feel that Altera tools are a little behind in this regard, but Intel + Altera is a different beast now, of course. Personally, I think I'll stick with Xilinx, although, if you know your Verilog/VHDL you shouldn't have any problems just working with both. It'll probably waste time if you really should have to learn both, but wth.

[nctnico]

Anyway, this question half me trying to figure out what FPGA maker/toolchain to go with and half me being curious about the industry and where it's going. If you have any insight, it would be cool to hear!

I don't really have any insight  , but I'll dump my thoughts here.

First of all, moore's law - most likely- is coming to an end and we're processing more and more stuff, e.g. Augmented reality stuff, Pokemon Go, Speech, AI, more video processing etc. So going for FPGAs is the right thing and I see all those ARM cores in mobiles, tablets, etc. getting 'FPGA-extensions' very soon (especially ZYNQ-like SoCs).

For high volume consumer devices an FPGA is about the worst choice. An ASIC is much more cost effective and consumes less power. An FPGA is a niche solution for projects where you can't use a SoC alone.