Why NVIDIA Killed 3dfx: How the Voodoo Brand Still Haunts Modern GPU Launches

NVIDIA bought 3dfx Interactive in December 2000 for approximately $70 million in stock. They acquired the Glide API patents, the SLI trademark, and most importantly a group of engineers whose design lineage traces from SGI's RealityEngine through the Voodoo Graphics chip and into the GeForce FX. 3dfx wasn't killed by NVIDIA — it had already lost the GPU war before the acquisition closed. NVIDIA was buying what remained.

The 1996-2002 GPU war: context

The GPU industry between 1996 and 2002 went through a consolidation that mirrors what happened to the CPU market in the 1990s. Dozens of accelerator makers — 3dfx, NVIDIA, ATI, Matrox, S3, Rendition, PowerVR, Trident — competed in what was effectively a new market segment. By 2002, three remained competitive: NVIDIA (top), ATI (second), and a fading Matrox serving the workstation niche.

3dfx's trajectory from 1996 to 2000 is one of the best-documented collapses in tech history — covered at length in AnandTech's 3dfx acquisition report and the Tom's Hardware Voodoo Graphics history pictorial. The speed of the collapse — from market leader to bankruptcy in under four years — was driven by a single strategic decision made in 1998.

Key takeaways:

• 3dfx invented consumer 3D acceleration in 1996 with the Voodoo Graphics chip
• The STB Systems acquisition in 1998 cut off the AIB channel that had driven Voodoo volume
• NVIDIA shipped the GeForce 256 in August 1999, beating the Voodoo5 to market by eight months
• The SLI trademark, key patents, and ~100 engineers transferred to NVIDIA in December 2000
• 3dfx's Glide API lives on in dgVoodoo2, still maintained as of 2026

What made the original Voodoo Graphics card revolutionary in 1996?

The Voodoo Graphics chip (internally code-named "Obsidian"), designed by Gary Tarolli at 3dfx Interactive in San Jose, was the first consumer product to deliver hardware-accelerated 3D rendering at a price point regular gamers could afford. Before Voodoo, 3D acceleration was either absent (software rendering on a CPU) or locked to $2,000+ professional workstations running OpenGL.

The Voodoo card was a pass-through device: it sat between the VGA card and monitor, intercepting the VGA signal and substituting its own 3D output only when a game called the proprietary Glide API. This design had two clever properties: the Voodoo didn't need to handle 2D at all (saving significant silicon budget), and it was board-vendor agnostic — Diamond, Orchid, Canopus, Creative, and others all shipped Voodoo-based products.

Glide was the key software layer. id Software's Quake and Quake II, the first Need for Speed, MechWarrior 2 Mercenaries, and dozens of other 1996-1997 releases had Glide render paths that ran 2-4x faster and at higher quality than the same games' software-renderer paths. For a PC gamer in 1996, buying a Voodoo card was the single largest performance upgrade possible.

The Voodoo Graphics launched at $299 in 1996 and within 12 months had captured an estimated 80% of the add-in 3D accelerator market.

How did 3dfx blow its lead between Voodoo2 and Voodoo3?

The Voodoo2 (1998) extended the lead — dual-chip SLI configurations (Scan-Line Interleave) pushed framerates to 60fps at 800×600, unmatched by anything else available. But the seeds of 3dfx's collapse were planted in that same year.

The STB acquisition (December 1998): 3dfx paid approximately $141 million in stock to acquire STB Systems, a major add-in board manufacturer. The rationale was vertical integration — by making their own cards, 3dfx would capture the margin that had been going to Diamond, Creative, and Hercules. The execution destroyed the channel.

Per multiple GPU-history postmortems including the Tom's Hardware retrospective, the AIB partners who had been 3dfx's volume drivers — Diamond (Viper line), Creative (3D Blaster), Hercules — immediately pivoted to NVIDIA's TNT and TNT2 chips. NVIDIA kept its fabless AIB-partner model, scaled volume across a dozen board vendors simultaneously, and in 1999 outshipped 3dfx roughly 4:1 at retail.

The Voodoo3 positioning mistake: The Voodoo3 (launched February 1999) was technically inferior to NVIDIA's TNT2 in one measurable dimension: it lacked 32-bit color rendering (limited to 16-bit). 3dfx's marketing position was that human eyes couldn't distinguish the difference in motion — arguably true, but the spec sheet loss handed reviewers and competitors a clear narrative: "Voodoo3 is last-gen."

What was the Voodoo5 6000 and why did it never ship?

The Voodoo5 6000 was 3dfx's terminal project: a single AGP card carrying four VSA-100 chips (3dfx's final GPU design), each at 183 MHz, targeting $600 at retail in late 2000. The card required an external "Voodoo Volts" DC power brick because the AGP slot couldn't supply enough current for four active chips.

Per TechPowerUp's Voodoo5 5500 spec entry and documented 3dfx engineering history, the Voodoo5 6000 faced two compounding problems:

1. Manufacturing yield. VSA-100's quad-chip yield was low — only a small fraction of production wafers produced four working chips at spec. The engineering sample program burned through inventory before retail quantities were achievable.
2. Market timing. NVIDIA launched the GeForce 2 Ultra in August 2000 at $499, delivering DirectX 7 T&L hardware acceleration that the Voodoo5 architecture (fixed-function, no T&L) fundamentally couldn't match. Reviewers benchmarking pre-release Voodoo5 6000 samples against GeForce 2 Ultra found the 4-chip Voodoo5 competitive in raw fill rate but trailing significantly in titles using T&L features.

Approximately 200 Voodoo5 6000 engineering samples exist in collector hands as of 2026. Working units clear $8,000-15,000 on eBay — the single most valuable consumer GPU collectible.

Era spec comparison: Voodoo5 5500 vs GeForce 2 GTS vs Radeon DDR (1999-2000)

The T&L (Transform and Lighting) column was decisive. As game engines adopted DirectX 7's T&L pipeline from 2000 onward, titles that used it ran substantially faster on GeForce 2 and Radeon than on the Voodoo5. 3dfx's excellent FSAA implementation became a minor footnote against a missing fundamental feature.

What did NVIDIA actually buy in December 2000?

The December 2000 acquisition transferred four categories of asset from 3dfx to NVIDIA:

Patents. 3dfx held several multi-GPU rendering patents from the Voodoo2 SLI era. NVIDIA later deployed these patents defensively and in the foundation of their own SLI implementation (introduced with the GeForce 6800 in 2004).

The SLI trademark. "SLI" stood for Scan-Line Interleave in 3dfx's implementation — an approach where one card rendered odd scan lines and the second rendered even scan lines, halving per-card workload at the cost of some synchronization complexity. NVIDIA repurposed the trademark for "Scalable Link Interface," their alternate-frame rendering implementation that became SLI for GeForce 6000 and later.

~100 engineers. Including key architects Gary Tarolli (3dfx co-founder, previously at SGI's hardware architecture team) and Scott Sellers (3dfx CTO). Both joined NVIDIA post-acquisition. The SGI-to-3dfx-to-NVIDIA pipeline is one of the most-cited engineering lineages in GPU history — the teams that built SGI's RealityEngine geometry pipeline and 3dfx's Glide rendering pipeline contributed to NVIDIA's shader architecture through the G80.

The Glide source code and documentation. NVIDIA released the Glide source code to the public domain after the acquisition — the basis for the open-source Glide wrappers in use today.

Which Voodoo engineers shaped the modern GeForce?

Per documented engineering credits, Gary Tarolli (3dfx co-founder, ex-SGI) contributed to the GeForce FX (NV30) architecture team at NVIDIA following the acquisition. The GeForce FX's programmable shader design, while criticized at launch for thermal and performance issues relative to ATI's R300, set the architectural foundation for the G80's unified shader model.

Scott Sellers (3dfx CTO) moved into broader graphics research at NVIDIA. The specific projects he contributed to post-acquisition aren't publicly documented in the same granularity as Tarolli's, but his presence in the GPU architecture group through the G80 era is cited in multiple industry retrospectives.

The SGI connection matters: SGI's RealityEngine (1992) was the GPU architecture from which multiple 3dfx founders drew their design intuitions. Those intuitions — particularly around parallel rasterization pipelines and SIMD arithmetic units — are traceable through the Voodoo architecture into NVIDIA's pre-unified-shader pipeline design.

How does NVIDIA's modern SLI / NVLink lineage trace back to 3dfx?

3dfx's SLI (Scan-Line Interleave) connected two Voodoo2 cards via a 50-pin ribbon cable on a shared PCB connector, splitting the framebuffer horizontally across cards. One card rendered the top half of the screen; the other rendered the bottom half. The composite was merged in the analog domain before the monitor connection.

NVIDIA's 2004 SLI ("Scalable Link Interface") used the same trademark but a different implementation: alternate-frame rendering (AFR), where each card renders an entire frame in sequence and the output alternates. The SLI bridge connector (later NVLink) handles inter-GPU communication and framebuffer synchronization.

NVLink, introduced with Volta for professional applications and later extended to GeForce with Turing-era NVLink bridges, evolved from the PCIe lane allocation concepts that SLI used since 2004. The 3dfx trademark was the seed; the technical lineage is indirect but continuous.

Could 3dfx have survived? A counterfactual

The STB acquisition is the hinge point. Without it:

• 3dfx retains the AIB channel — Diamond, Creative, Hercules continue shipping Voodoo3 and Voodoo4 cards
• NVIDIA's volume advantage in 1999-2000 is reduced (though not eliminated)
• 3dfx reaches retail with Voodoo5 6000 in Q1 2001, competing against GeForce 2 Ultra and early Radeon 8500

However, the T&L gap remains. Even a channel-intact 3dfx shipping Voodoo5 6000 faces titles that favor GeForce 2's hardware T&L. 3dfx would have needed a VSA-200 with T&L support — a 12-18 month engineering cycle — to remain competitive past 2001.

Per GPU-history analysis, the most optimistic counterfactual is a 3dfx that survives to ship Rampage (its T&L-capable successor chip, canceled in the bankruptcy) in 2002, then competes in the DirectX 8 era. A less optimistic reading is that the STB acquisition accelerated an inevitable outcome by 12-18 months.

What's the collector market for Voodoo cards in 2026?

The Voodoo5 6000 market is illiquid — fewer than 10 working units sell per year publicly. Most pricing data comes from forum transaction logs rather than completed eBay sales.

When 3dfx won an era — and when it didn't

Related guides

• /reviews/1999-voodoo3-3500-pentium-iii-period-correct-build-2026 — Build a period-correct Voodoo3 rig
• /buying-guide/retro-gaming — Retro gaming hardware buying guide

Citations and sources

• AnandTech: 3dfx acquired by NVIDIA (2000)
• TechPowerUp: Voodoo5 5500 GPU specs
• Tom's Hardware: 3dfx / Voodoo Graphics history pictorial

This piece is editorial synthesis based on publicly available information. No independent first-party benchmarking is reported.

SpecPicks Editorial · SpecPicks · Last verified 2026-05-18