I hardly even understand FPGA's but I don't think TI's newest tech will obsolete it. Intel just purchased Altera for $16.7B and I've already seen them release updates to their line including fabrication on 14nm or less.
Xilinx invested heavily in TSMC's TSV interposer and likely their IN-FO fan-out packaging. These allow them to integrate more silicon area in the same package space. It can potentially offset the technology node gap, and maybe for lower cost. Yes, TSV is expensive, but a process node is even more expensive.
Multi-die FPGA packages are a giant PITA to deal with, compared to monolithic. It's a lot harder to resolve timing problems.
There's a small customer base that's already used to partitioning their designs across multiple FPGAs, and taking the huge performance hit from it (e.g., ASIC emulation). But for most, yeah... giant PITA.
There was some press material around that, as well as some preliminary support in the Sourcery toolchain back around the 2010-2012 timeframe. But I haven't heard anything about that effort lately, and I think that it has stalled.
In the article, a TI rep is quoted saying, "We believe customers will be able to develop their devices three times faster than when using FPGAs — in days rather than weeks"
I'm kind of wondering about that. Maybe if you were already up and running with syslink/codecengine/xdctools, etc. IMHO, working with TI's c6X DSPs was just awful, and their software suite was one of the only dev environments that has made me pine for Xilinx's stuff.
Syslink? Perhaps the worst-behaved linux kernel module I've ever encountered. Xdctools? Did anyone ever really want yet another build system? And in javascript? When we were working with the toolchain, even TI considered it inscrutable - their own image loader was a hacked example buried deep in the xdctools examples dir. And support? Every time we sent an email to TI, all we got was, "have you asked this question in the TI forums?" It was such a ridiculously poor standard of support that eventually it became a joke on our team.
But I know that in embedded, the quality of the software tools is rarely considered. Oh well.
Yeah, the headline on eetimes is a bit...sensationalistic. It may be useful for cellular signal processing, but there are quite a few other things that fpgas are used for. Now, were this chip to have a proper PCIe x16 bus dangling off it, and elements useful for video encoding/decoding, it may be more intriguing. Right now, it's just an interesting toy for telcos.
This is additional information that what I suspected was happening, is, which is that the combination of some hard "cores" + FPGA fabric is becoming useful. The original Xilinx run at this used Power PC cores but the current stuff (Zync) using ARM cores has had much better traction.
Looking at the TI site they want $328 each[1] in singles, not a great price point but not horrible given their performance but I don't think it crosses into the 'revolutionary' category.
That said, one of these and a couple of those 4 GSPS 12-bit ADC's and you'd have a pretty sweet front end for an oscilloscope. My new 200Mhz scope only has 2.5GSPS front ends on it and it cost $6K.
[1] This compares to $60 for singles of the Zync chip.
Thanks, I am not a fan of phonetically obscured names :-). Xilinx's first foray into this space was the upper end Virtex series (the Virtex 4 and Virtex 5), I went back and looked at the pitch they gave me in 2005 around this. So here we are 10 years later and it seems like this stuff is finally getting closer to mainstream.
12 comments
[ 2.4 ms ] story [ 37.0 ms ] threadThere's a small customer base that's already used to partitioning their designs across multiple FPGAs, and taking the huge performance hit from it (e.g., ASIC emulation). But for most, yeah... giant PITA.
In the article, a TI rep is quoted saying, "We believe customers will be able to develop their devices three times faster than when using FPGAs — in days rather than weeks"
I'm kind of wondering about that. Maybe if you were already up and running with syslink/codecengine/xdctools, etc. IMHO, working with TI's c6X DSPs was just awful, and their software suite was one of the only dev environments that has made me pine for Xilinx's stuff.
Syslink? Perhaps the worst-behaved linux kernel module I've ever encountered. Xdctools? Did anyone ever really want yet another build system? And in javascript? When we were working with the toolchain, even TI considered it inscrutable - their own image loader was a hacked example buried deep in the xdctools examples dir. And support? Every time we sent an email to TI, all we got was, "have you asked this question in the TI forums?" It was such a ridiculously poor standard of support that eventually it became a joke on our team.
But I know that in embedded, the quality of the software tools is rarely considered. Oh well.
Looking at the TI site they want $328 each[1] in singles, not a great price point but not horrible given their performance but I don't think it crosses into the 'revolutionary' category.
That said, one of these and a couple of those 4 GSPS 12-bit ADC's and you'd have a pretty sweet front end for an oscilloscope. My new 200Mhz scope only has 2.5GSPS front ends on it and it cost $6K.
[1] This compares to $60 for singles of the Zync chip.