Why Early GPUs Integrated TMDS Transmitters

In the late 1990s and early 2000s, graphics processing units (GPUs) transitioned from relying on external chips to handle digital display outputs to integrating the Transition-Minimized Differential Signaling (TMDS) transmitter directly onto the graphics silicon. This architectural shift was primarily driven by the need to reduce manufacturing costs, improve high-frequency signal integrity, simplify printed circuit board (PCB) design, and meet the escalating bandwidth demands of higher-resolution digital displays.

The Original Approach: Discrete Transmitters

When the Digital Display Working Group introduced the Digital Visual Interface (DVI) specification in 1999, graphics cards primarily used analog VGA outputs generated by internal RAMDACs. To support DVI, graphics card manufacturers added dedicated third-party TMDS transmitter chips—such as those manufactured by Silicon Image—onto the PCB.

In this discrete setup, the GPU outputted wide, parallel digital pixel data across PCB traces to the external transmitter chip. The external transmitter then serialized this parallel data into high-speed, low-voltage differential signals sent through the DVI cable to the monitor. While this allowed GPU makers to offer digital output without redesigning their core processor dies, it introduced major hardware limitations.

Preserving Signal Integrity at Higher Frequencies

The most critical engineering factor behind integrating TMDS onto the GPU silicon was signal integrity. As display resolutions shifted from standard XGA (1024x768) toward UXGA (1600x1200) and beyond, the pixel clock speeds required by the TMDS protocol rose toward the single-link ceiling of 165 MHz.

Routing a wide, parallel digital bus (often 24 bits or wider) across a PCB from the GPU to an external transmitter created severe timing and electromagnetic interference (EMI) challenges:

  • Trace Length Matching: High-speed parallel signals must arrive at the destination chip simultaneously. Discrepancies in trace lengths resulted in signal skew that corrupted video output.
  • Crosstalk and Parasitic Capacitance: Multiple high-frequency parallel lines running alongside each other generated inductive and capacitive interference.
  • Electromagnetic Radiation: High-speed parallel traces on the board surface acted as miniature antennas, making it harder to pass regulatory standards like FCC compliance.

By moving the TMDS transmitter directly onto the GPU die, the wide parallel bus remained entirely within the silicon, where distances are measured in micrometers rather than inches. The GPU only needed to output the serialized, noise-resistant differential pairs directly to the external display connector.

Reducing Manufacturing Costs and Bill of Materials

Integrating the transmitter delivered immediate economic advantages for GPU designers and board partners. An external TMDS transmitter was an expensive discrete component that inflated the total Bill of Materials (BOM) for each graphics card.

Moving the logic into the GPU silicon reduced costs in several ways:

  • Component Elimination: Board partners no longer had to purchase separate transmitter ICs and the supporting passive components (resistors, decoupling capacitors, and voltage regulators) required to operate them.
  • Simplified PCB Routing: The elimination of the dense parallel bus between the GPU and the transmitter freed up valuable PCB real estate. This often allowed board designers to use PCBs with fewer copper layers, significantly lowering the fabrication cost of the card.
  • Assembly Efficiency: Fewer components on the surface-mount line translated to faster pick-and-place times and fewer potential solder-joint defects during manufacturing.

As display technology advanced, single-link DVI hit a bandwidth wall at 1920x1200 at 60 Hz. Supporting larger panels—such as the 30-inch 2560x1600 displays introduced in the mid-2000s—required dual-link DVI, which doubled the number of data pairs.

Implementing dual-link DVI using external chips required either routing two massive parallel buses across the PCB or implementing multiple transmitter chips, which was expensive and mechanically unwieldy. On-die integration allowed GPU designers like NVIDIA and ATI to easily incorporate multiple dual-link TMDS engines into the core silicon. This architecture paved the way for GPUs capable of driving multiple high-resolution digital displays directly from a single graphics processor.