someguy12

I work for a small trading firm. Our trading platform is on FPGA. I use SystemVerilog for both RTL and testbenches. I edit in Sublime with a SystemVerilog plugin. I use Sublime's build system to call Modelsim/Questa commands to catch errors inline. I run sims in Questa. I build with Jenkins calling Vivado Tcl scripts and use Perl to scrape the logs and timing reports to determine if a build was successful.

Indeed. Deathbird Stories was my intro to Ellison; it completely blew my 20 year old mind.

Yeah. The trick is getting your foot in the door. Equal amounts of expertise and luck are helpful. Most jobs you'll find online or through a recruiter aren't going to be the $300k+ jobs. I found a job opening at an investment bank on monster.com (of all places) after having worked for tech companies for 15 years. I started at $150k. However once I was in finance, I was able to make connections and acquire skills that made me more valuable. After almost 5 years I was able to land a job that pays (much) more than $300k.

In my experience these firms do not do true hardware design. We purchase FPGA boards from 3rd party vendors and then do the FPGA development in house.

I'm an FPGA dev making more than 300k in NYC working for a trading firm.

Although this is misplaced, I'll reply anyway.

There isn't more buzz around the Stratix 10 MX because it's a phantom chip. Intel's December press release clearly states that the chips are available but you can not buy one of these chips today. A blogger did some research and came to the conclusion that the press release claims were simply not true:

https://semiaccurate.com/2017/12/19/intels-claims-fpgas-hbm-...

In contrast, Xilinx released a video of their competing chip working with all channels running at the full 460GB/s on day 1 of silicon bring up:

https://www.xilinx.com/video/fpga/first-virtex-ultrascale-pl...

That quote is misleading. He's talking about perceived price. The actual price of an equity is the price at which the execution occurs. And for every execution there are two sides, hence the zero sum game is the only game in town.

They mention that the F1 instance is using Xilinx's UltraScale+ FPGAs and that they have approx. 6800 DSP slices.

You can see that they must be targeting the VU9P from the product selection guide (pdf) here:

http://www.xilinx.com/support/documentation/selection-guides...

From the guide:

  VU9P:
    6840 DSP slices,
    2.3M flip-flops,
    1.1M LUTs
    36Mb distributed RAM
    76Mb block RAM
    270Mb ultra RAM

Additional documentation is linked on the last page of the selector guide.

Note that these specs are per FPGA. An F1 instance comes with up to 8 of these devices.

It's a very expensive, very customizable appliance. Hence Amazon's AWS F1 instance and Nimbix cloud.

Altera/Intel and Xilinx have FPGAs coming out this year with HBM which should make them competitive with existing GPU memory bandwidth.

Can you explain your perspective on open source ideology as it applies to hardware design? I see people complain every so often about lack of open source tools for FPGAs and CPLDs and it doesn't make sense to me.

Most people seem happy to use a chip that performs the exact functions they wanted even though it's completely closed/proprietary. But they seem to balk at using a programmable chip where they can choose the functionality but the software that let's them do so is closed/proprietary.