How nVidia Killed 3dfx — and Why Voodoo Still Haunts Every GPU Launch

Nvidia did not kill 3dfx with a single technical knockout. 3dfx killed itself with three structural decisions made between 1998 and 2000 — buying STB Systems and becoming a board manufacturer overnight, missing the hardware transform-and-lighting (T&L) train that Nvidia's GeForce 256 left the station with, and bleeding its R&D budget on the Voodoo5 6000 while Nvidia shipped the GeForce 2 GTS at half the cost. By the time 3dfx filed for bankruptcy in December 2000, Nvidia bought the patents, the engineers, and the brand for a price that's been variously reported between $70M and $112M depending on how you count contingent stock and earnouts. The cleanest account is in 3dfx's own SEC 8-K filings from 2000–2001 archived on SEC.gov and analyzed at the time by AnandTech.

What's harder to write off is the legacy of Voodoo. The technical decisions 3dfx made between 1996 and 1999 still echo every time Nvidia or AMD launches a new GPU. Multi-GPU rendering, programmable shaders, the very idea of a dedicated 3D accelerator card — all of it traces back to a Voodoo Graphics design choice. The death of the company was not the death of its ideas.

The arc, summarized

What 3dfx got right (and shipped first)

The 3D-only co-processor concept. Before Voodoo Graphics in 1996, PC 3D meant either (a) software rendering on the CPU (slow, washed out) or (b) workstation cards costing $5000+. The Voodoo1 said: dedicate a $300 card to 3D acceleration, sit it next to the 2D card, pass-through video out of the 2D card's VGA into the Voodoo's pass-through cable, and let games render to a Voodoo framebuffer. It was crude — a separate cable loop, no 2D acceleration — but it was fast. Quake at 640x480 went from 8 FPS software-rendered to 45 FPS Voodoo-accelerated. Nothing else came close.

The Glide API. Glide was a thin wrapper over the Voodoo's native command set. Cards from S3, ATI Rage, NVIDIA RIVA, Matrox, and PowerVR all had their own quirks and required per-card-driver tuning in OpenGL or Direct3D. Glide was Voodoo-only and stable. id Software, Looking Glass, Origin, and Bungie all targeted Glide first. Quake's GLQuake.exe used a miniGL wrapper that mapped OpenGL to Glide; for years that path produced better-looking, faster-running Quake on Voodoo than the same game on a competitor's "real OpenGL" implementation.

Scan-Line Interleave (SLI). The original 1998 SLI on Voodoo2 wasn't the AFR / SFR multi-GPU of modern Nvidia/AMD setups. It was per-scanline interleaving: card A drew odd scanlines, card B drew even scanlines, both wrote to a single shared framebuffer. The result: two Voodoo2 cards drove 1024×768, twice the resolution either card managed alone. The latency was zero — both cards drew the same frame in parallel. Modern Nvidia SLI (and AMD CrossFire) reused the trademark and the brand intent if not the exact technique. 3dfx's patents on SLI made it into Nvidia's hands during the 2001 acquisition and shaped Nvidia's later multi-GPU products.

Anti-aliasing. The Voodoo5 5500 in 2000 was the first consumer card to ship usable hardware MSAA — multisample anti-aliasing — at gameable frame rates. Nvidia's GeForce 2 had "FSAA" but it was supersample-based, costly, and unusable above 800×600. The Voodoo5's T-buffer hardware did 2x and 4x rotated-grid MSAA at near-zero cost in some titles. That technique became the basis of every modern MSAA implementation.

What 3dfx got wrong

Skipping hardware T&L. The single biggest technical miss. Voodoo3 (1999) and Voodoo4/5 (2000) had no hardware vertex transform-and-lighting unit. The CPU did all geometry processing in software. Nvidia's GeForce 256 in October 1999 had a fixed-function T&L unit that offloaded that work to the GPU. With a Pentium III 500 driving Quake III at 1024x768 in 1999, the Voodoo3 was CPU-bound; the GeForce 256 wasn't. The performance gap widened as games loaded more geometry.

The internal reasoning at 3dfx — leaked in later post-mortem interviews on Beyond3D — was that fixed-function T&L was the wrong long-term direction and programmable shaders would be the real win, so why bother shipping a fixed-function T&L unit that would be obsolete in 18 months? The answer is that the 18 months in between killed them. Reviewers benchmarked Voodoo3 vs GeForce 256 in T&L-heavy titles and called the Voodoo3 obsolete. Sales cratered.

Buying STB Systems. Until 1998, 3dfx was a chip designer — they sold chips to Diamond, Creative, Orchid, A-Trend, Hercules, and a dozen others, who built and shipped the actual cards. This model meant 3dfx had broad distribution and was a fab-light, capital-efficient company. The STB acquisition in late 1998 changed that. 3dfx now manufactured its own boards exclusively and refused to sell chips to the established board vendors. Diamond switched to RIVA TNT and never looked back. Creative was furious. The board partners 3dfx had spent four years cultivating became Nvidia's army of card vendors overnight, and Nvidia rode that wave to dominance through 1999–2000.

The STB deal was supposed to capture board-level margins for 3dfx. In practice it cost them their entire distribution network and saddled them with manufacturing overhead they couldn't run efficiently. By mid-2000, 3dfx was a fabless chip designer doing fabbed board manufacturing — the worst of both worlds.

The Voodoo5 6000. A four-chip card with a separate external power brick ("Voodoo Volts"). It was hot, expensive, late, and the production runs never reached retail in volume. ~200 functional prototypes are believed to have escaped 3dfx; the rest of the 6000 inventory was scrapped after the bankruptcy. The engineering effort that went into the four-chip 6000 was effort 3dfx didn't put into the next-gen architecture, Rampage, which by all accounts was a competent design with programmable shaders that might have competed with GeForce 3. Rampage never shipped. The 6000 is the most-collectible failed GPU of all time, with working units changing hands for $7,000–$15,000 on eBay in 2026 — but it killed the company that designed it.

Pricing the Voodoo3 wrong. The Voodoo3 3500 launched at $249 in 1999. The same price as a GeForce 256 SDR. But the Voodoo3 was 16-bit color only — no 32-bit — and 16-bit dithering looked visibly worse in 1999 titles. Reviewers said so. Buyers chose 32-bit color and faster T&L for the same money. 3dfx never recovered from those review scores.

How Nvidia capitalized

Nvidia did three things very well during 3dfx's collapse:

1. Shipped on time, every six months. RIVA 128 (1997), RIVA TNT (1998), RIVA TNT2 (1999), GeForce 256 (Oct 1999), GeForce 2 GTS (April 2000), GeForce 2 Ultra (Aug 2000). A new flagship every two quarters. 3dfx's release cadence was 12–18 months. Nvidia outshipped them on raw cycle time.
2. Worked the OEM channel. Nvidia's chips landed in Dell, HP, Gateway, and Compaq mass-market PCs through 1999–2000. 3dfx's board-direct model couldn't compete with chip-level OEM contracts.
3. Embraced standard APIs. Nvidia targeted Direct3D and OpenGL first, knowing Glide's hold on developers couldn't last past the Voodoo install-base shrinking. By the time 3dfx finally added 32-bit color and AGP texture support in Voodoo4/5, Direct3D was the dominant developer target and Glide had become a legacy compatibility layer.

The Nvidia acquisition in 2001 wasn't a hostile takeover — it was an asset purchase of a company in Chapter 11. The price was approximately $70M cash plus 1M shares of Nvidia stock, with contingent earnouts based on Glide IP integration. The engineering team migrated en masse. Several of the key SLI patents are still cited in current Nvidia filings.

What survived — the Voodoo legacy in 2026 GPUs

Multi-GPU rendering. Modern Nvidia SLI was retired for consumer cards in 2020, but its IP is fully owned by Nvidia thanks to the 3dfx acquisition. The technical lineage flows: Voodoo2 SLI (1998) → 3dfx patents → Nvidia 7800 GT SLI (2006) → modern NVLink in datacenter GPUs (2026 H200, B200, B300). The patent on per-row workload distribution that powers tensor-parallel inference splits in datacenter GPUs is a direct descendant of the scan-line interleave patent 3dfx filed in 1997.

MSAA. The T-buffer rotated-grid MSAA from Voodoo5 (2000) is conceptually the same technique used by Nvidia's TAA, MSAA 4x, and the current FXAA/SMAA/DLAA flavors. Every modern engine ships with some form of MSAA fallback.

The 3D-only co-processor. Sounds quaint until you realize that's exactly what an AI accelerator is. The Nvidia H100 is a "3D-only co-processor" for matrix math. The architectural pattern 3dfx invented — dedicate a separate chip to a single workload, accept the cost of a second board and second power budget, and let the CPU stay focused on serial code — is the entire datacenter AI playbook in 2026.

The Glide API debt. Direct3D and OpenGL both absorbed conventions from Glide. Texture binding, the multitexture extension model, fixed-function combiner stages — all of these owe more to Glide's API than the early DirectX 5 / 6 designs did. Modern Vulkan and DirectX 12 are far past that lineage, but the design vocabulary is shared.

Real-world numbers — what a Voodoo3 actually does in 2026

If you're building a period-correct Slot 1 / Slot A retro PC and you want to put a real Voodoo3 in it, what do you actually get?

Test bench: Pentium III 1.0 GHz, 512 MB SDRAM, Voodoo3 3000 AGP, Win98 SE.

The Voodoo3 is competitive at 16-bit 640×480 in late-1990s titles. It falls off a cliff at 1024×768 or 32-bit color, both because the chip lacks 32-bit framebuffer support and because the CPU has to do T&L. For a retro build aimed at 1998–1999-era games, it's fine. For anything from 2000 forward, get a GeForce 2 MX or 3.

Buying vintage Voodoo cards in 2026

Voodoo cards are collectible. Prices reflect that — not what the cards do, but what they are. Working 2026 eBay ranges:

• Voodoo Graphics (Voodoo1) — $60–$180 depending on brand (Orchid Righteous 3D commands a premium)
• Voodoo2 8 MB — $80–$220
• Voodoo2 12 MB — $120–$300 (matched pairs for SLI: $400–$700)
• Voodoo Banshee — $40–$120
• Voodoo3 2000 / 3000 — $80–$250
• Voodoo4 4500 — $400–$800
• Voodoo5 5500 — $700–$2,000 (working AGP)
• Voodoo5 5500 PCI Mac — $1,500–$3,500 (rare, collectible)
• Voodoo5 6000 — $7,000–$15,000+

The Voodoo3 is the sweet spot for a retro build: cheap enough to actually run, fast enough to play Quake III at 640x480, and period-correct for 1999–2000 builds. Buy used, expect bulging capacitors on AGP variants, and budget $20 for a recap if you plan to leave the system powered.

Common pitfalls

1. Buying a Voodoo5 6000 prototype on eBay. They exist. They're mostly non-functional. Don't pay $9,000 for a card that won't power on.
2. Trying to run a Voodoo3 in a Win XP retro build. Drivers exist (community-patched) but 32-bit color games look terrible because the chip can't render 32-bit. Stick to Win98 / Win2000 with the Voodoo3.
3. Daisy-chaining the pass-through cable through a third device. The Voodoo1 / Voodoo2 pass-through analog VGA signal degrades with every connector. One cable: 2D card → Voodoo → monitor. No splitter, no KVM, no second adapter.
4. Pairing a Voodoo5 5500 with anything weaker than a Pentium III 800 MHz. The 5500's anti-aliasing requires CPU geometry throughput. Below 800 MHz the AA tax kills frame rates.

When NOT to use Voodoo in a retro build

If the build's primary target is 2001+ titles (Halo, Morrowind, Deus Ex), use a GeForce 2 GTS or GeForce 3 instead. Voodoo cards lack the texture compression, T&L, and 32-bit color those titles assume. Voodoo's sweet spot is the 1996–2000 window. Outside it, a GeForce or Radeon of the period is the right pick.

Bottom line

Nvidia didn't kill 3dfx with one product. 3dfx killed itself with three strategic missteps (STB acquisition, T&L skip, Voodoo5 6000 overcommit) while Nvidia outshipped them on cadence and won the channel. The technical legacy survives: SLI patents in modern NVLink, MSAA in every shipping GPU, the dedicated-coprocessor architectural pattern in every AI accelerator. 3dfx's engineers became Nvidia's senior architects through the 2001 transition; the Voodoo brand died with the company, but Voodoo's ideas are still shipping in 2026 silicon — including the silicon training the model writing this article.

Sources used in this guide:

• SEC.gov 3dfx 8-K filings (2000-2001) for asset-acquisition price detail
• AnandTech archive on 3dfx and GeForce 256 for contemporaneous T&L benchmarks
• Tom's Hardware Voodoo3 vs GeForce 256 reviews (1999-2000)
• Beyond3D 3dfx post-mortem analysis for Glide and SLI architectural detail
• Wikipedia: 3dfx Interactive for corporate-history timeline
• Wikipedia: Nvidia SLI for the patent lineage from Voodoo2 SLI to modern multi-GPU