A High-Throughput, Low Cost, Low Power Edge Inference Co-Processor

Attend the Linley Spring Processor Conference to listen to Cheng’s presentation on Monday afternoon April 10th in Santa Clara:
click here to see the full agenda. Or come back here April 11th to see the slides on-line.

Learn How the Winograd Transform Accelerates Inference Models

2019 03 11 Cheng Wang on Winograd Transformation.png

Watch Cheng Wang’s video interview by SemiEngineering - he describes how convolution can be accelerated by the Winograd Transformation without sacrificing precision.

Here is the link to the video.

We Are Hiring Smart & Motivated People Who Want Growth+Challenge

See our Careers Page for our openings for RTL Design and Physical Design.

The Four Characteristics of an Optimal Inference Accelerator

EE Times article: CLICK HERE.

MorningCore Technology Licenses EFLX eFPGA for TSMC12FFC


MorningCore Technology, the semiconductor subsidiary of China telecommunications giant Datang, is licensing EFLX4K eFPGA for TSMC’s 12nm FinFET Compact technology (12FFC) and the EFLX Compiler for Programming eFPGA. MorningCore joins a growing list of major companies and organizations adopting EFLX eFPGA to make their logic flexible. Here is the press release with more info.

Learn about Options for Real Time Object Recognition

Attend Cheng Wang’s talk on Real Time Object Recognition from Backup Cameras to Highway Speeds March 27th at the Autonomous Vehicle Hardware Summit in Silicon Valley. Did you know that thousands of children are seriously injured every year worldwide by cars backing up over them and hundreds of them die? Real time object recognition on a backup camera can save lives. Here is the AGENDA. To register, go to Autonomous Vehicle Hardware Summit.

Lies, Damn Lies or TOPS/Watt?


How to measure and compare neural accelerator performance and power efficiency? Unfortunately TOPS has become an inflated marketing metric with too many unstated assumptions that vary between vendors. The right comparison is to determine throughput, for example in images per second of a given size if it is for a CNN, for your model, for your batch size, for your process and for the exact process corner, voltage and temperature conditions you care about. Then do the same for throughput per watt while measuring power under the exact same conditions as the throughput. This will give you a much better metric. See our TOPS/Watt page for more info.

New NMAX Technical Disclosure at the Edge AI Summit

Cheng Wang, Co-Founder and Sr VP, Engineering/Architecture/Software, disclosed substantial new technical info at the Edge AI Summit in San Francisco on December 11. You can see his full slide deck on the NMAX page. NMAX excels at getting high hardware utilization at batch=1, which is critical for edge applications. Most of the new information is in the 2nd half of the slide deck.

NMAX delivers from 1 to 100+TOPS, as you need, at 10x lower cost and 3-5x lower power than existing solutions because we get 50-80% MAC utilization on tough models, meaning we need less silicon area, and we achieve it with 10x less DRAM bandwidth, meaning less DRAM cost and system power. See the AI Edge Summit slides to learn how the architecture works and benchmarks.

2018 12 every tile is reconfigured every layer.png
2018 12 cheng presenting at ai edge summit.jpg

Flex Logix in EE Times’ Silicon 60 Class of 2018

EE Times list of 60 startups worth watching including Flex Logix: https://www.eetimes.com/document.asp?doc_id=1333985

Microprocessor Report: Flex Logix Spins Neural Accelerator

Download their article HERE.

eFPGA Performance Modeling & Silicon Correlation

We generate timing for all process corners like you do for ASICs. See our slides from TSMC OIP HERE.

Boeing Licenses Flex Logix eFPGA for GF14LPP

Beoing has licensed Flex Logix’ eFPGA for GlobalFoundries’ 14LPP process. GlobalFoundries manufactures 14LPP in Malta, near Albany, in upstate New York State. Boeing joins a growing list of companies designing with EFLX eFPGA. See the press release on our NEWS PAGE.

There are other Aerospace companies and US government organizations now designing or planning SoCs for GF14LPP using EFLX eFPGA. Get more information on EFLX eFPGA for GF14LPP HERE.

The availability of eFPGA on an advanced FinFET process manufactured in the USA is of major interest to US Aerospace companies who until now have not had access to American-manufactured FPGA technology. Market research reports show that almost 10% of FPGA sales by $ value are to Aerospace. And DARPA and others report that >30% of the IC content of Aerospace systems are FPGA. Integrating eFPGA enables SoCs to be made smaller, lighter, low power; and often integration breaks chip-to-chip I/O bottlenecks thus improving performance. Now eFPGA is available in an advanced FinFET process manufactured in the USA for those projects that require domestic manufacturing.

Flex Logix has been working for years with US Aerospace companies and government organizations: DARPA funded Flex Logix to demonstrate the viability of large, high performance eFPGA with a 200K LUT4 eFPGA evaluation chip on TSMC16FFC (which is available now on an evaluation board for customers to run their designs at GHz speeds). Sandia Labs contracted with Flex Logix to develop eFPGA for their proprietary 180nm wafer fab to develop multiple SoCs with eFPGA: the first SoC, Dragonfly, was described by Sandia at DAC 2018 (see below for slides) and that SoC is now out of fab and being validated.

Our support of US Aerospace companies does not limit our ability to support our customers worldwide: we have customers in Israel, China and other parts of Europe/Asia who are also benefiting from EFLX eFPGA. For customers in China who prefer domestic manufacture, EFLX eFPGA is available on multiple TSMC processes including 16FFC which is expected to be manufactured at TSMC’s Nanjing fab.

More Customers Have Working Silicon & Are in Design

Harvard and Sandia Labs both have working silicon using EFLX eFPGA and made presentations (DAC, HotChips); others have not made public disclosure. More customers are in design such as Boeing, HiPer, SiFive - other customers are in design but are not disclosing publicly at this time. See are Working Silicon Web Page for more details on these chips. More customers are in contract, planning and evaluation. The adoption of eFPGA has begun and is accelerating.

Harvard Silicon Shows eFPGA is Lowest Energy Flexible DNN

2018 08 20 Harvard NN energy efficiency HotChips.png

Monday August 20, Harvard University presented a paper at Hot Chips: "SMIV: A 16nm SoC with Efficient and Flexible DNN Acceleration for Intelligent IoT Devices".  The design was done in TSMC16FFC.  They implemented several methods for neural network efficiency including a 2x2 array of EFLX eFPGA with two EFLX 4K Logic cores and two EFLX4K DSP cores with 40 22x22 MACs each for 80 total.  The best solution of four was AON which will not benefit from algorithm advances: there is a tradeoff between performance and future flexibility.  Of the more flexible solutions the eFPGA solution was similar in area efficiency to an ARM A53 and a quad-accelerator, but on energy efficiency the best of the flexible solutions by far is eFPGA (see chart at right).

Harvard's presentation is available to Hot Chips attendees and will be posted on the Hot Chips site in the future.

Harvard is working now on a follow-on neural accelerator using EFLX eFPGA again.

Working with Harvard, and with other customers, has helped us understand the needs for neural networks: on November 1st at the Linley Fall Process Conference we will unveil our Neural Inferencing Architecture called NMAX™. NMAX uses eFPGA for reconfigurable state machines and XFLX/ArrayLinx/RAMLinx interconnects for efficient and low power data movement. NMAX achieves scalable, high performance inferencing with very low DRAM bandwidth. Low DRAM bandwidth means fewer DRAMs, less DRAM $ and less power. See our NMAX page HERE on November 2nd for more details.

Benchmarking eFPGAs: Performance, Area, Physical Design Issues

Flex Logix has now published an application note on benchmarking eFPGA performance showing details of performance with exact measurement conditions for a range of applications.  Unfortunately there is no competitive information but you can download the RTL files to use when you benchmark other eFPGA.  Go to the web page with the details HERE.

Also, there is public information from Achronix and Menta allowing a detailed comparison of area with Flex Logix on common process nodes. Go to the web page with the details HERE.

Tony Kozaczuk, Director Solutions Architecture, has written an article discussing all of these factors in benchmarking eFPGA: see HERE.for

Screenshot 2018-05-21 09.59.47.png


eFPGA Timing Signoff Methodology

2018 04 efpga timing signoff methodology.png

Click HERE for a video discussion by Senior VP Engineering Cheng Wang or HERE for a written summary.


EFLX Compiler Demonstration Video

Click on the image to the right for a ~10 minute video demo of EFLX Compiler.




EFLX200K 16nm Evaluation Board Demonstration Video

Click on the image to the right for a ~5 minute video demonstration.



EFLX4K eFPGA IP Core Validated on TSMC16FFC
EFLX200K Evaluation Boards Now Available

2018 04 Eval Board Demo Video.png

The EFLX4K eFPGA IP core, both the Logic and DSP versions, have been fully validated on TSMC16FFC. The GDS is also compatible with TSMC16FF+ with no change, just retiming.

Evaluation boards are available now that integrate the EFLX200K validation chip (a 7x7 array of EFLX 4K cores: 182K LUT4, 560 MACs, 1.4Mbit attached SRAM, PLL & PVT) for customers to test their RTL on real silicon.  Customers can either use the boards at no cost for short-term evaluation or can purchase boards.  The evaluation board comes with documentation, examples, and test benches.

Contact info@flex-logix.com for more information on obtaining an evaluation board for evaluation or purchase.


2017 09 TSMC Partner of the Year 2017 -- New IP.JPG


Six Game Changer Companies to Watch in 2018: click HERE.
TSMC OIP Partner of the Year 2017: press release HERE.
Flex Logix joins TSMC IP Alliance: press release HERE.
EE Times Top 60 to Watch : SEE HERE.

Follow us on Twitter @efpga or Linked-in for the latest news on Flex Logix.


Flexible, Proven Building Blocks: We Deliver More, Sooner 

Our management team has the experience to grow Flex Logix rapidly while meeting commitments.  Our engineers have worked at many of the world's best chip companies on processes from 180nm to 7nm designing high volume ICs.  We can meet your needs and we will meet our commitments.  

We prove our Hard IP in Silicon before you use it.  We execute effectively and quickly: in 2.5 years we have designed EFLX™ in TSMC 40, 28 and 16FF.  Our lead customers have their chips working with EFLX arrays embedded.

EFLX150 1x1 to 5x5.jpg

We will port to TSMC 7/22/55/65/90/180nm on demand.  We can port to any CMOS process, foundry or captive, in ~6 months from receipt of PDK and standard cell library -- we only port with a committed customer to ensure we have a partner to help us optimize our digital architecture for the target process.  For example, in 40ULP we optimized for numerous power management modes with rapid transitions, 0.5V state retention and even 0.5V operation if needed.  Whereas in 16FF customers want performance, performance, performance.  See EFLX Implementation Options for a full list. In all nodes it is the same digital architecture with the same programming software.

EFLX4K 1x1 to 7x7.jpg

In each process node we implement up to 4 IP cores: EFLX150 and EFLX4K; each has an all-logic version and a version with embedded Multiplier-Accumulators for DSP Acceleration.  Each IP core is a stand-alone FPGA, but incorporates additional top-level interconnect which allows automatic connection to adjacent IP cores turning them automatically into larger EFLX arrays.

Our innovative building block approach means we can deliver any size embedded FPGA from 100 LUTs to 200K LUT4s, with about 75 sizes in between: and each size can be further customized with the mix of DSP acceleration and/or RAM that you might also need.  We have even fabricated an EFLX200K array which is in characterization now.

IF 200K LUTs is too small for you, inquire about our roadmap to ~800K LUT4 arrays which we can implement quickly: info@flex-logix.com.

EFLX® Compiler - try it yourself!

Our software, the EFLX Compiler,  converts your RTL into a bit stream which you load into the EFLX FPGA array to program it to run your RTL including DSP and Block RAM.  A GUI is available for many phases of eFPGA design, timing, I/O placement and more.  The EFLX Compiler is in use at our lead customers and is now available for evaluation for qualified customers: contact us at info@flex-logix.com.

Learn more about EFLX Compiler here.

Need a summary to circulate to the team?

Download our 4-page PDF overview of NMAX neural inferencing: click here to get the PDF.

Download our 4-page PDF overview of EFLX eFPGA:  click here to get the PDF.
Download our 2-page product brief on the EFLX100 logic/DSP cores for TSMC 40ULP/LP: click here to get the PDF.
Download our 2-page product brief on the EFLX4K logic/DSP cores for TSMC 28HPC/HPC+: click here to get the PDF.
Download our 2-page product brief on the EFLX150 logic core for TSMC 16FF+/FFC/12FFC: click here to get the PDF
Download our 2-page product brief on the EFLX4K logic/DSP cores for TSMC 16FF+/FFC/12FFC: click here to get the PDF.
Download our 2-page target spec on the EFLX4K logic/DSP cores for GF 14LPP:  click here to get the PDF.

Download our 2-page DSP Architecture brief: click here to get the PDF.

Copyright © 2015-2018 Flex Logix Technologies, Inc. EFLX and Flex Logix are Registered Trademarks of Flex Logix Technologies, Inc.  
NMAX, ArrayLinx, RAMLinx and XFLX are Trademarks of Flex Logix Technologies, Inc.