The traditional FPGA market is abandoned

http://www.xess.com/blog/extinction-level-event/

by jamesbowman • 10 years ago

76 33 10 years ago

payne92• 10 years ago

I think it gets squeezed even more with general purpose GPU computing.

astrodust• 10 years ago

Intel's trying to hedge their bets here by having strong FPGA capability. GPGPU compute is not the answer to every problem.

exelius• 10 years ago

There's no reason you can't have general-purpose SoCs that combine traditional pipeline-driven CPU cores, GPGPU cores and FPGA cores with memory, bus chips, storage, etc. We long ago reached the point of diminishing returns for cramming more raw computing power on a chip; so chips of the future will likely consolidate specialized features for various workloads in the name of power savings / speed increases for certain workloads.

rjsw• 10 years ago

Plus Intel owns the IP to all the pieces and is happy to provide documentation for them.

Xilinx have Mali GPUs on their latest ARM+FPGA hybrids but that isn't much use to me if I can't find out how to program it.

ska• 10 years ago

The main problem with GPGPU is the "general" comes with a lot of caveats and costs. It's great for some problems, no so much for others.

ZenoArrow• 10 years ago

I may be wrong on this, but isn't one of the main drawbacks of the GPGPU approach the high power draw? High end GPUs do seem to be quite thirsty. I've no idea if any FPGAs offer better performance per watt, but I wouldn't be surprised if they did.

EDIT: This article seems to suggest that FPGAs currently have an edge over GPUs when it comes to energy efficiency:

http://app.scientificcomputing.com/articles/2015/03/optimizi...

lvh• 10 years ago

I'm not sure I agree with the idea that Intel is going to make these folks work on something else because FPGAs don't grow that much. That makes little sense; Intel isn't really hurting for more employees, and the FPGA engineers can't necessarily be made to be productive on the higher-growth sections cited.

Instead, it makes a lot more sense that Intel is going to try and make that market more lucrative and larger instead. As a major chip manufacturer, they're in a great position to ship FPGAs to tons of new customers. Chicken-and-egg problems (you don't FPGA until you need to because it's not available) have made FPGAs a niche element. However, OpenPOWER/CAPI have demonstrated that you can get huge benefits from slapping FPGAs to existing general purpose compute.

So; TL;DR: I don't think it makes sense to assume Intel is just going to assume that market will stay what it is. Instead, they think they can make that market better, and do more than Altera/Xilinx can do individually. In that light, a purchase makes perfect sense.

(Disclaimer: I work for a company involved in OpenPOWER/OpenCompute and has shipped hardware that does this.)

Retric• 10 years ago

FPGA's are always significantly worse than dedicated HW. To sell them in bulk you need to be able to replace a wide range of chips you don't need at the same time with a good enough solution. At the same time you need enough speed you can't just simulate things on a general purpose processor.

That's a fairly narrow band and the real issue with FPGA's. A synthesizer for example seems like a good fit, but well general processes are fast enough. If we had a wider range of video codec's that might work, but again general purpose processors are good and dedicated hardware can help for the mainstream codec's.

Any ideas?

dv_dt2• 10 years ago

Anything that is needed to run in a pipelined data flow - particularly if the cpu processing of that data would be non-trivial. One example might be video processing (the kind that might go in all those automated cars on the horizon). You might have one set of algorithms at initial release, but in future releases one might change them.

Other examples might be crypto or compression/decompression, transcoding ..

ae_keji• 10 years ago

Maybe easy open source implementations of GPUs/video hardware, for the libre everything types.

kqr2• 10 years ago

In case the link is down:

http://webcache.googleusercontent.com/search?q=cache:http://...

Also, I believe that cheap microcontrollers have been able to replace FPGAs in some cases.

LeifCarrotson• 10 years ago

Cheap microcontrollers can replace old CPLDs, which now are technically FPGAs but really don't have much of a place in modern hardware.

They can also replace FPGAs because even small microcontrollers are now really systems-on-a-chip, with integrated SPI, CAN, LVDS, or even USB/transceivers. To me, that's more like replacing an FPGA with dedicated hardware - it's just that microcontroller peripheral options are so thorough that dedicated hardware no longer needs to be custom.

ArkyBeagle3• 10 years ago

It's mixed. You can spin a fast loop in a PIC @ 1,2, 4 or 8 MSPS, and that eats into FPGA territory.

USB is uniformly terrible; multi-PHY CAN offerings are disappearing. SPI can be good if you have the right interface chip ( which you may or may not actually know until you're running against a prototype ).

LVDS looks interesting, but it seems to have been a lot supplanted by SPI/I2C and Ethernet.

I dunno; it's just different. Phones have warped the market in puzzling ways.

petra2• 10 years ago

Or there's another alternative. The main reason it was hard to make money from the industrial etc segment - was low volume combines with a lot of support costs.

But what if

a. xilinx built c-like tools that enabled embedded software engineers develop easily ?

b. they released those freely to some segment of the market ?

c. they've built an external support and IP ecosystem, either open or closed or both ?

Those actions can increase margins for xilinx, and they seem to be doing a,b.

As for the hardware, maybe the article is right. Also ,recent industrial chips are using 28nm, and going beyond that is extremely expensive and might not fit the industrial scenario anyhow, maybe there's not a lot of innovation left in the industrial FPGA market,

chriscool• 10 years ago

Also ,recent industrial chips are using 28nm, and going beyond that is extremely expensive

Yeah, so ASICs that are not produced in big enough volume will not move to 14nm or less. This means that FPGA that can move to 14nm or less, because of bigger volume, may become competitive against those ASICs.

petra2• 10 years ago

When i'm saying "industrial chips are at 28nm" i mean industrial FPGA's.

Will xilinx create industrial FPGA's at 14nm or beyond ?

First transistor cost will have to become meaningfully lower than 28nm transistor costs. That only happens at 10nm. But at that node, NRE costs are extremely expensive. Spread over low-cost low-end industrial chips - this requires a huge volume, which xilinx probably doesn't have yet.

Also couple that with 28nm being more much reliable(all the failure mechanisms increase at 10nm: electron migration from wires, thermal hot sports, transistor fin self heating), and since reliability is key for industrial - it would be hard to see industrial moving beyond 28nm.

nickpsecurity• 10 years ago

No, the industrial market is split between 28nm or below, 45-65nm, and 180-350nm. You can compete with fixed HW on performance, power, or flexibility. I think you're talking performance. Plenty of improvements a FPGA can get on lower nodes over 45-65nm or 180-350nm. It's actually part of how Xilinx and Altera got their market share: replacing older, low-volume parts with FPGA's on new nodes.

ArkyBeagle3• 10 years ago

I think embedded software engineers will need significant... "reeducation" to be facile in FPGA. Like close to starting over. "They" don't want the external support and IP ecosystem; their behavior looks like they're still using closed-ness as an accounting virtue.

FPGA tools have been so bad for so long that I just don't think there's enough executive bandwidth to truly embrace software. The economics of being an FPGA designer pretty much requires lockin. It is a lot to ask of a team.

PaulHoule2• 10 years ago

People are still scratching their head over this acquisition but here is my take.

Intel has failed at phones not because "x86 sux" but because Intel makes a small number of SKUs that are often different bins of the same design -- these are heavily tuned in every way. Phone chips are not so optimized in terms of process and design, but instead they are true "systems on chip" where a number of IP blocks are integrated to produce something which is optimizable by adding functional blocks.

Something FPGA based, whether it ends up in the datacenter or elsewhere, could have a highly optimized FPGA block which is field configurable, so this gives Intel something to compete with SoC vendors on their own terms.

One detail is the nature of the memory and I/O systems. FPGAs can move a huge volume of data, so unless you upgrade the the paths in and out of the CPU/FPGA area, the FPGA would be starved for bandwidth.

It would take one of two things for the FPGA market to expand based on these developments.

First, if there was some "killer app" that sold a huge number of units, that would help. The trouble here is that if you sell a huge number of units, you might be better off making an ASIC functional block an integrating it into a true SoC.

The other one is that (I think) FPGA development today is constrained by the awful state of tooling. Pretty soon you will be able to pop a FPGA hybrid chip into a high-end Xeon workstation and have one hell of a development kit in terms of hardware, but without a big improvement in the tooling, very few people will be able to really make use of it.

sixdimensional• 10 years ago

Not sure but these two articles/papers came to mind:

"Scalable Multi-Access Flash Store for Big Data Analytics" https://people.csail.mit.edu/wjun/papers/fpga2014-wjun.pdf

"FPGA-based hardware acceleration for a key-value store database (2014)" https://news.ycombinator.com/item?id=11374801

rwmj• 10 years ago

I think your analysis of Intel's problem is spot on. But FPGAs have something like (IIRC) 4% of the density of custom-designed ASICs. Can Intel's superior process create FPGAs at such small scales to overcome this rather large disadvantage?

Agree about the terrible tooling. 'Twas always thus, unfortunately, even when I was doing PLDs back in 1991.

PaulHoule2• 10 years ago

Well the economics get awful for designing chips as the feature size gets smaller, which is keeping 28mm fabs pretty busy and will just get worse at 10 or 8 nm or whatever comes next.

KSS42• 10 years ago

Another reason they bought Altera was to keep the their FAB utilization high. They have to "feed the beast.

sitkack• 10 years ago

Excellent insight, but it is bigger than that. By weaving FPGA fabric into their products, both internal facing and external facing it will enable a higher engineering velocity.

1) less SKUs, differentiation is in the firmware (bitstream). Can serve a market that wants 4 sata ports w/o having custom silicon.

2) infield bugfixes and updates that go deeper into the product. The whole TSX debacle could have been avoided.

Just like in a code, 80% doesn't need to be fast yet it takes up area. Using an FPGA fabric could reduce die sizes (profits up).

Intel has always wanted VLIW, adding FPGA to the mix could get them their dream.

Just how everyone wanted to add GPU compute to the CPU, the same thing is happening. Eventually a mix of scalar (cpu), bulk parallel (gpu) and dynamic (fpga) will be available on die. Your basic blocks might just be jitted into a bitstream.

dmytroi• 10 years ago

Ehm, FPGA/CPLD are everywhere, it's hard to say, but my educated guess is that hardware market consumes similar amount of SoC-like and FPGA/CPLD chips, look, even Pebble smartwatch uses Lattice iCE40 FPGA, MacBooks also usually have few FPGA's here and there. FPGA/CPLD's are essential to implement "glue-logic", which does what it says - glue different pieces of hardware together. Of course you can do "glue" in software, but than your customers will enjoy poor battery life time and random glitches here and there. And nobody gonna spend big-money (it's actually depends, sometimes it 1kk+$, sometimes you can get away by 30-50k$) by running their own ASIC's when you can just buy jellybean FPGA's.

Indeed phone/desktop market might move to more one-chip-for-everything solution, but even then we need glue-ish logic to control something like screen backlight DC-DC converters, charging IC's, etc, which is much more easier done from FPGA/CPLD-like devices. On the other hand FPGA/CPLD's are essential in some classes of devices, for example test instruments: modern oscilloscopes usually have 3+ FPGA's in them, companies like Keysight usually only run custom ASIC's when they hit limitation of current silicon tech, like their N2802A 25GHz active probe (starting from 23500$) uses amplifier IC made with indium phosphide process (InP), which is kinda far away from whatever current consumer product companies are doing, you can check the teardown of this beast here https://www.youtube.com/watch?v=jnFZR7UsIPE

So in my opinion FPGA/CPLD market will live long and strong, players might change, but demand is enormous. The only problem in my opinion is that whole market is more B2B-like (FPGA's are usually just a humble IC's inside end customers products, you don't see stickers "Altera inside" or anything on products themselves), so it's kinda hard to get grasp what's going on.

kevin_thibedeau• 10 years ago

XESS is biased toward the segment of the market they play in which is centered around computational applications where having a micro on board is usually desirable. There are many more places where simple glue logic is needed and a small, cheap FPGA/CPLD fits the bill. One would hope that Altera's (and possibly Xilinx's) new masters won't fuck this up and shutter their low end devices but that doesn't mean there won't be demand for them into the future. People with a need for these types of FPGAs aren't XESS customers so they are blinded by the limited scope of their own business.

JoachimS• 10 years ago

The article is based on the premise that Altera will stop developing FPGAs for general availibility. Nothing I've seen (and would love to be shown otherwise) so far indicates this.

FPGAs has been used as test designs when qualifying fabs for volume production since they are regular, but more complex and closer match to general ASICs than memories.

FPGAs are also often used more and more either with internal CPU cores (hard or soft) or as a companion to CPUs providing acceleration, esp data plane processing.

This means that the aquisition of Altera is mor of adding a business that complements the CPU business, not remove a competitor from another market segment. Intel can sell CPU+FPGA solutions for data center, big data. But it can also sell more chips and increase utilization in their fabs.

And for FPGA users the FPGAs coming out of these fabs will probably be better with higher density, lower power consumption that what Altera managed to design themselves. And getting a Stratix or a Cyclone SoC with an Atom core inside running Linux would be a very neat solution.

The traditional FPGA market is abandoned

Related Stories