AI summary by TwinMind
The History of the FPGA: The Ultimate Flex
Asianometry
18m 7s
Early Programmable Logic Challenges
- Pre-FPGA era: two main prototyping options
- TTL chips on printed circuit boards (limited by size/power)
- Custom ASICs (application specific integrated circuits)
- Substantial upfront costs for photo masks
- Months-long fabrication wait times
- No ability to edit/recompile after manufacturing
- Only economical for large use cases
First Programmable Logic Devices (1970s)
- 1970 Harris Semiconductor PROMs
- Programmable read-only memories implementing programmable logic gates
- Fixed AND gates, programmable OR gates
- Used fusible nichrome metal links (one-time programmable)
- Targeted combinational logic functions
- Pioneered burnable fuse fabrication with wafer test processes
- 1975 Signetics PLA (Programmable Logic Array)
- Both AND and OR gate sets programmable
- Wider range of logic functions (minimal sum of products)
- Cost: $25 each (1975 dollars) = ~$140 today
- Required larger chip size, raising costs
- Still used burnable fuse programming
- Never achieved commercial success despite advantages
PAL Development and Success (1977-1980s)
- 1977 Monolithic Memories PAL (Programmable Array Logic)
- Engineers John Berkner and H.D. Tua reversed PLA concept
- Programmable AND gates, fixed OR gates
- Better performance/cost than PLA
- Added sequential logic circuits via macro cells
- Included flip-flops for single-bit storage
- Enabled mix of combinational and sequential logic
- Early 1980s improvements:
- UV light erasable programming (no more fuses)
- Categorized as Simple Programmable Logic Devices
- Widely adopted by memory manufacturers
- Easy user learning curve
- PAL Scaling Problem:
- Structure didn't scale with Moore's Law
- Growing AND gate sets became slower
- Transistor count grew faster than input/output connections
Complex Programmable Logic Devices (1980-1984)
- 1980 Altera Formation
- Founded by ex-Intel/Signetics/Intersil engineers
- Started as Source 3 design consulting company
- Raised $750K VC funding
- July 1984 EP300 release:
- CMOS EEPROM process (erasable/reprogrammable)
- Had to use Ricoh Japan for CMOS (unavailable in US)
- Pioneered fabless design model
- Complex Programmable Logic Device Architecture:
- Multiple smaller PALs connected via crossbar connections
- More scalable than single large PAL structure
- Step toward modern FPGA architecture
Xilinx FPGA Innovation (1984-1985)
- Company Formation:
- Founded by Ross Freeman (Zilog engineer) in February 1984
- Original name "Logica" already taken
- Vision: Moore's Law would make universal silicon possible
- Partnership with Seiko Corporation for fabrication
- Exclusive Japan resale rights
- Seiko's 2.5 micrometer CMOS process
- XC2064 Development Challenges:
- Designer Bill Carter had no prior CMOS experience
- Cross-Pacific communication with Japanese fab
- Largest chip Seiko had ever attempted
- High yield risk for startup
- Repeatedly told to "keep it simple"
First FPGA Architecture (XC2064)
- Revolutionary Design Changes:
- Removed AND/OR gate sets entirely
- Array of Configurable Logic Blocks (CLBs)
- Programmable interconnect switches
- Device I/O placed around perimeter
- Configurable Logic Block Structure:
- Pair of lookup tables (LUTs)
- LUTs map inputs to outputs for arbitrary logic functions
- Results routable to flip-flops and other logic
- Technical Specifications:
- 85,000 transistors
- 64 configurable logic blocks
- 58 input/output blocks
- <1,000 gates equivalent
- Required external memory for programming storage
- First Fabrication Results:
- 25 wafers received from Seiko
- Only 1 wafer worked (the 11th)
- Successfully programmed as inverter
- Proved FPGA concept viable
Market Competition: FPGA vs PAL
- FPGA Disadvantages:
- Much higher cost (hundreds of dollars vs cheap PALs)
- Large die size = extremely low yields
- Complex, hard-to-use EDA software
- Unpredictable performance
- Xilinx proprietary design tools
- PAL Advantages:
- Memory-like structure = easy manufacturing
- Simple, widely available third-party EDA tools
- Easy user adoption
- Predictable performance
Technology and Business Convergence (1990s)
- Moore's Law Benefits:
- Maturing lithography enabled classical scaling
- FPGAs scalable while PALs were not
- Crashing transistor costs favored FPGAs
- Chemical mechanical polishing reduced interconnect costs
- Exponential growth in LUT counts and wire lengths
- Fabless Model Rise:
- Independent foundries emerged
- Altera/Xilinx pioneered fabless startup model
- Design teams could focus on innovation vs manufacturing
- Foundries used large FPGA dies to perfect processes
- FPGA Startup Ecosystem:
- Multiple new FPGA companies emerged
- Most eventually closed or exited
- Survivors absorbed IP/patents from failed startups
- Xilinx and Altera maintained market leadership
Software Control Advantage
- PAL Software Limitation:
- Third-party EDA vendors controlled capabilities
- PAL makers had no agency over hardware potential
- Market commoditized to lowest common denominator
- Similar to cheap Wintel PC maker dynamics
- FPGA Software Control:
- Xilinx/Altera controlled their own EDA tools
- Like NVIDIA controlling their own drivers
- Enabled dictating future roadmaps
- Software crafted to enable new features
- Added automation for productivity gains
Modern FPGA Industry
- Market Consolidation:
- Xilinx and Altera held majority market share
- Multi-billion dollar industry
- 2015: Intel acquired Altera for $16.7 billion
- 2022: AMD acquired Xilinx for ~$50 billion
- Current Applications:
- ASIC design/fab costs increasingly expensive
- Strong in aerospace, military, telecom
- Growing into AI processing and data centers
- Valuable gap-filling role in market
- Swiss army knife functionality across industries
