Short answer: Microsoft's Direct3D won the late-90s API war not by being technically superior — Glide was faster and easier for developers through 1998 — but by being cross-vendor, OS-integrated, and improving generation-over-generation while 3dfx's Glide stayed proprietary. The 1999 launches of NVIDIA's GeForce 256 and the Direct3D 7 spec collapsed the performance gap, the 3dfx STB acquisition alienated AIB partners, and the Voodoo5 6000's vaporware reputation finished the company off. 3dfx filed for bankruptcy in 2002; NVIDIA bought the assets.
The 1996–2002 API war, and why it matters in 2026 for retro builders
For six years, choosing a graphics card meant choosing an API. A 3dfx Voodoo gave you Glide-exclusive titles — Unreal, Need for Speed III, Quake 2's mini-driver Glide port — that ran faster and looked better than their Direct3D equivalents. An NVIDIA TNT2 or a Matrox Millennium gave you broader compatibility through Direct3D and OpenGL but lower performance in the Glide titles you actually wanted to play.
By 2001 the game was over. NVIDIA had absorbed 3dfx's engineering team. Glide was a wrapper-only relic. Direct3D 8.0 with programmable shaders had launched. Every game ship since uses Direct3D or OpenGL/Vulkan as its primary API; Glide is preserved only through community wrappers like nGlide and dgVoodoo2.
In 2026 this matters for two reasons. First, retro-PC building is having a moment — the Voodoo era is the apex of "computer hardware as object" for collectors, and understanding what made Glide special is required context for any period-correct build. Second, the historical pattern (proprietary vendor API beats open standard for 3–5 years, then ecosystem gravity flips it) is rhyming in 2026 with CUDA vs ROCm, where the same dynamics are playing out at a different scale. Knowing how Glide lost is useful pattern-matching for guessing how the next API war ends.
This piece tells the story, lands on the technical and business reasons Glide failed, and walks through what a Glide-era retro rig looks like in 2026 — including the storage adapters and period-correct audio hardware that make a Voodoo build feel right.
Key takeaways
- Glide won 1996–1998 on raw performance — 30–50% faster than Direct3D on the same scene
- Direct3D 6 and 7 closed the gap; Direct3D 8's programmable shaders ended Glide's relevance
- NVIDIA's GeForce 256 (Oct 1999) introduced hardware T&L, which Glide couldn't expose
- The 1998 STB acquisition turned 3dfx into its own retail competitor, alienating Diamond, Creative, and ELSA
- Voodoo5 6000 became vaporware; Voodoo4/5 ran hot and lacked T&L; 3dfx bankrupt in 2002
- Modern Glide wrappers (nGlide, OpenGlide, dgVoodoo2) handle 95%+ of Glide-era titles on current hardware
- Building a Glide-era rig today: Voodoo2/3 PCI cards, period-correct CompactFlash storage, Sound BlasterX G6 for audio
What was Glide, and why was it 3dfx's competitive moat?
Glide was 3dfx's proprietary 3D graphics API, shipped alongside the original Voodoo Graphics card in 1996. It exposed a thin layer over the Voodoo's hardware capabilities — texture mapping, alpha blending, perspective correction, mip-mapping — with minimal CPU overhead and a small API surface. A game developer targeting Glide needed to learn maybe 30 functions; the equivalent Direct3D code at the time required 150+ function calls with significant boilerplate.
The performance moat came from three sources. First, Glide bypassed Windows entirely — it talked directly to the Voodoo card through a custom driver, avoiding GDI overhead. Second, the API was tightly fitted to the hardware — every Glide function mapped to a small, well-understood sequence of hardware operations, where Direct3D had to support many vendor-specific quirks. Third, the developer experience was clean — id Software's John Carmack publicly preferred Glide's simplicity, and his endorsement carried weight.
Per Quake 2 timedemo benchmarks from the 1998–1999 PC Gamer and Maximum PC archives, Glide on a Voodoo2 delivered roughly 30–50% higher framerates than Direct3D 5 on a Matrox or early Riva 128 at 800×600. The gap was perceivable in side-by-side comparisons — Glide felt smoother, was more consistent, and shipped with better default visual quality settings.
Direct3D 5/6/7: Microsoft's catch-up curve
Direct3D launched in 1995 as part of DirectX 2, and the early versions were technically inferior to Glide on every metric that mattered. Direct3D 3 (DirectX 3, late 1996) was widely panned by developers — the API surface was confusing, the driver model required deep vendor cooperation, and OpenGL via id's mini-driver approach was often easier to ship against.
Direct3D 5 (DirectX 5, mid-1997) introduced a cleaner pipeline model and improved texture management. Direct3D 6 (DirectX 6, mid-1998) added multitexturing support and standardized vertex buffers, putting Direct3D within 10–15% of Glide's performance on a Riva TNT or original Voodoo2. Direct3D 7 (DirectX 7, late 1999) launched alongside the GeForce 256 and exposed hardware T&L — and that's the point where Direct3D started winning on capability, not just closing the gap on performance.
The crucial dynamic: Direct3D improved each version because Microsoft funded the work and could lean on multiple silicon partners to shape the spec. Glide stayed roughly constant — 3dfx added features incrementally (Glide 2 in 1998, Glide 3 in 1999) but couldn't keep pace with the spec's evolution. Glide 3 added 32-bit color and large textures, but the same features had already shipped in Direct3D 6/7 with better tooling and broader hardware support.
By 2000 the question wasn't "does Glide outperform Direct3D" — it was "does the title I want to play require Glide, and is the title's developer still actively supporting it?" The answer for most titles was no.
OpenGL's quiet third-place role
OpenGL was the third API in the race, and id Software's John Carmack carried it almost single-handedly through the Voodoo era. Quake (1996) shipped with OpenGL support via id's "miniGL" driver — a subset of OpenGL specifically chosen so that 3dfx could implement it as a wrapper over Glide. The miniGL approach let Glide-class hardware run an "OpenGL" path without needing a full OpenGL implementation, which was important because full OpenGL drivers were the province of professional workstation cards (3DLabs, Intergraph) that cost more than a Voodoo did.
By Quake 3 (1999), Carmack had pushed the industry toward full OpenGL drivers, and NVIDIA's Riva TNT2 and GeForce 256 shipped with proper OpenGL implementations. The Glide-as-OpenGL-substitute approach died because the hardware caught up.
OpenGL never became the consumer-gaming dominant API on Windows — Microsoft made sure of that through Direct3D investment and OS integration — but it remained the cross-platform default for engines that needed Linux or Mac support. id Tech engines used OpenGL through Doom 3 (2004); the Direct3D shift happened later for them than for most studios.
The 1999–2000 turning point: GeForce 256, T&L, and Voodoo3 obsolescence
NVIDIA's GeForce 256 launched in October 1999 and is the inflection point in this whole story. It was the first consumer card with hardware T&L (transform and lighting) — geometry processing handled on the GPU rather than the CPU. For games that used T&L, the GeForce 256 freed up huge amounts of CPU budget that previously went to vertex math. For the Voodoo3, released six months earlier, this was catastrophic.
The Voodoo3 — and the contemporaneous Voodoo2 SLI configurations — couldn't expose T&L through Glide because the hardware didn't have a T&L unit. Adding it would have required a new silicon revision, and 3dfx was already overdrawn on the Banshee/Voodoo3 development cycle. The technical gap meant that any developer targeting hardware T&L (Quake 3 Arena's vertex-shaded mode, Unreal Tournament's higher-detail settings) preferred the GeForce 256, and the Voodoo3 felt suddenly slow despite its higher fill rate.
The market noticed. By Q2 2000, NVIDIA's market share in consumer 3D crossed 3dfx's for the first time. Per industry retrospectives, the GeForce 2 GTS launched in April 2000 widened the gap further — twice the pixel pipelines of the GeForce 256, T&L on a refined process, and a price point that overlapped with Voodoo3/Voodoo4. The Voodoo4 was DOA at launch, and 3dfx's roadmap visibly slipped.
Spec-delta table: Voodoo era cards
| Card | Year | Memory | Fill rate | T&L | DirectX |
|---|---|---|---|---|---|
| Voodoo Graphics | 1996 | 4 MB EDO | 50 MPixel/s | no | DX5-class |
| Voodoo2 | 1998 | 8/12 MB EDO | 90 MPixel/s | no | DX6-class |
| Voodoo2 SLI (pair) | 1998 | 12+12 MB | 180 MPixel/s | no | DX6-class |
| Voodoo3 3500 | 1999 | 16 MB SDRAM | 366 MPixel/s | no | DX6-class |
| Riva TNT2 Ultra | 1999 | 32 MB SDRAM | 300 MPixel/s | no | DX6/DX7 |
| GeForce 256 DDR | 1999 | 32 MB DDR | 480 MPixel/s | yes (HW) | DX7 |
| Voodoo5 5500 | 2000 | 64 MB SDRAM | 367 MPixel/s | no | DX6-class |
| GeForce 2 GTS | 2000 | 32 MB DDR | 1,600 MPixel/s | yes (HW) | DX7 |
The fill rate column tells half the story. The other half is that the GeForce 256 and GTS spent their pixel-pipeline budget on T&L geometry that the Voodoo cards couldn't process — the per-frame complexity ceiling for the NVIDIA cards was dramatically higher.
How games shipped: Quake 2, Unreal, and the three-renderer era
The 1998–1999 era was the only period when major games regularly shipped three rendering backends — software, Glide, and Direct3D — and let the player choose. Quake 2 shipped with software, glquake (Glide via miniGL), and a Direct3D renderer. Unreal famously shipped with software, Glide, Direct3D, and an experimental OpenGL renderer; the Glide path was easily the best-looking and best-performing through 1999.
By 2000, the three-renderer pattern was dying. Direct3D had caught up; supporting Glide added test matrix complexity without a meaningful audience. Unreal Tournament (1999) shipped with Glide support but the Direct3D path matched it within 5–10% on equivalent hardware, and the marginal extra users on Voodoo cards didn't justify maintaining the renderer.
The last meaningful Glide-exclusive titles were a handful of late-1999 racing games and the Unreal-engine derivatives that retained Glide support through patches. By 2001, shipping a game with Glide support was a niche move, mostly for the small but loyal Voodoo collector base.
The 3dfx STB acquisition and Voodoo5 6000 vaporware
3dfx's 1998 acquisition of STB Systems was supposed to vertically integrate manufacturing — Glide and silicon under one roof, sold direct to retail. Instead it alienated every AIB partner (add-in board vendor) that had previously built Voodoo cards. Diamond, Creative Labs, ELSA, and other AIB partners had built their businesses on Voodoo Graphics and Voodoo2; the STB deal turned 3dfx into their direct retail competitor overnight.
Diamond Multimedia, the largest Voodoo board partner, immediately pivoted to NVIDIA. Creative Labs slow-walked away. ELSA went all-in on NVIDIA workstation parts. The retail channels 3dfx tried to own through STB couldn't absorb the volume that Diamond and Creative had previously moved, and the company was now competing on retail execution with no track record there.
The Voodoo5 6000 — a four-chip flagship that was supposed to demonstrate 3dfx's continued silicon ambition — never shipped. Engineering samples existed; reviewers got brief looks; full production was repeatedly delayed. The card became symbolic of 3dfx's 2000–2001 decline: ambitious silicon design held back by manufacturing problems the STB-integrated model couldn't fix.
Per 3dfx industry retrospectives, 3dfx filed for Chapter 11 in December 2000. NVIDIA acquired the intellectual property and a substantial portion of the engineering team in 2001. By 2002 the brand was dissolved.
Modern Glide wrappers — what's actually playable in 2026
The good news for retro builders is that Glide-era titles are highly playable on modern hardware through wrappers. The three relevant projects:
- nGlide: Pure Glide-to-Direct3D wrapper, Windows-only, supports most Glide 2/3 titles. Best for vanilla Unreal, original Tomb Raider, Need for Speed III.
- OpenGlide: Glide-to-OpenGL wrapper, cross-platform, older but still functional. Useful on Linux retro setups.
- dgVoodoo2: Glide + DirectX 1–9 wrapper to Direct3D 11/12, Windows-only, the most actively maintained option. Handles 95%+ of Glide-era titles per its compatibility list.
The aesthetic isn't perfectly authentic — modern GPUs render the same scenes with higher precision than the Voodoo silicon ever could — but compatibility is excellent and performance is essentially unlimited. For preservation and play, the wrappers are entirely sufficient. For period-correct nostalgia where you want to see the games as they originally rendered, real hardware is the only path.
Building a Glide-era rig today
A 2026 period-correct Voodoo build targets the 1998–2001 era. The core silicon choices:
- GPU: Voodoo2 12 MB (single or SLI pair) for late-90s feel; Voodoo3 3000/3500 for 1999–2000 era; Voodoo5 5500 PCI for the apex Voodoo experience
- CPU: Pentium III 600–933 MHz or Athlon Thunderbird 700 MHz–1 GHz for late-90s realism
- Motherboard: Asus P3B-F, Abit BE6, or any Slot 1 / Socket A board with stable BIOS
- RAM: 256–512 MB SDRAM PC-133
- Storage: Transcend CF133 CompactFlash 32GB on a passive CF-to-IDE 44-pin adapter — silent, period-correct, instant-boot
- Storage adapter (workbench): FIDECO SATA/IDE to USB 3.0 for imaging CF cards from your modern PC, or Vantec CB-ISATAU2 for USB 2.0 hosts
- Audio: Creative Sound BlasterX G6 as a modern-era period-feel external DAC; or original Sound Blaster Live! / Audigy for pure authenticity
- Monitor: Period CRT 17" or 19" — gives true scanline aesthetics that LCDs can't replicate
The Voodoo2 SLI configuration deserves a specific note: pairing two Voodoo2 12 MB cards with the proprietary SLI ribbon cable doubles fill rate and remains one of the most distinctive period-correct configurations available. Working Voodoo2 cards run $80–$200 each on eBay; matched SLI pairs run $200–$500.
Per detailed retrospectives on Vogons, Voodoo3 cards are the value pick at $60–$150 for a working 3000-series card. The Voodoo5 5500 PCI is the apex consumer pick at $400–$700. The unicorn Voodoo5 6000 prototype runs into the thousands when one surfaces.
For practical period gaming, a GeForce 4 Ti 4600 or Radeon 9700 Pro in the same retro-rig era delivers better Direct3D performance with broader title support. The Voodoo chase is collector-driven, not performance-driven — but the collector value is real.
Bottom line: the API war's lessons for every modern API battle
Glide lost the API war not because it was technically inferior — for three of the six years of the war, Glide was the better API — but because the structural dynamics favored Direct3D's cross-vendor, OS-integrated, generationally-improving model. Microsoft funded ongoing Direct3D development across multiple silicon partners; 3dfx funded Glide development as a marketing expense for one silicon line.
This pattern rhymes in 2026 with CUDA vs ROCm. CUDA, like Glide, is a proprietary API with strong developer tooling and a hardware company controlling the spec. ROCm, like early Direct3D, is the open competitor catching up generation by generation while a third party (AMD, here) drives ecosystem investment. The Vulkan layer and Intel's oneAPI play similar roles to OpenGL's late-90s position.
The historical lesson: closed APIs win the race for 3–5 years, then ecosystem gravity flips. CUDA's dominance through 2026 looks insurmountable today, but Direct3D's looked equally remote to Glide developers in 1998. Watch ROCm 7+, Intel's oneAPI, and the increasingly capable Vulkan compute paths over the next 2–4 years. The 2028–2030 inflection in AI silicon may rhyme more closely with the 1999–2001 inflection in 3D than anyone currently expects.
The Voodoo cards are beautiful objects of computer history. The lesson is bigger than the silicon.
Common pitfalls
- Pairing a Voodoo with a modern CRT-replacement. LCDs can't reproduce CRT-era scanline aesthetics; if you care about period look, source a real CRT.
- Running a Voodoo5 6000 prototype. The few in circulation are unstable and not worth the collector premium for actual use; chase a 5500 instead.
- Trusting eBay "tested working" claims. Voodoo cards from 25 years ago often have failing electrolytic capacitors. Budget for recap work or pay extra for documented refurbishment.
- Mixing AGP and PCI Voodoo cards. AGP Voodoo2 doesn't exist; Voodoo2 is PCI-only. Voodoo3 had AGP and PCI variants. Verify slot type before buying the wrong card for your motherboard.
- Forgetting the 3.3V vs 5V PCI question. Late-generation Voodoo cards (Voodoo3 and later) are 3.3V-keyed; older Pentium-II motherboards have 5V slots only. Slot keying isn't always obvious — research before purchase.
FAQ
Was Glide really faster than Direct3D in its prime? Per Quake 2 timedemo benchmarks from the 1998–1999 PC Gamer and Maximum PC archives, Glide on a Voodoo2 delivered roughly 30–50% higher framerates than Direct3D on a TNT2 at 800×600 in glquake. The gap closed dramatically with Direct3D 6.0 and was effectively erased by Direct3D 7.0 in 2000. The performance moat that justified Glide-exclusive titles evaporated within two driver generations.
What killed 3dfx specifically — the API loss or the STB acquisition? Per industry retrospectives, both. The STB acquisition in 1998 forced 3dfx to compete with its own former AIB partners, alienating Diamond, Creative, and ELSA. Simultaneously, NVIDIA's TNT2 and GeForce 256 closed the performance gap while supporting Direct3D and OpenGL natively. 3dfx's last cards (Voodoo4 and 5) shipped late, ran hot, and lacked T&L. The combined hit was unrecoverable; bankruptcy followed in 2002.
Can I still run Glide-exclusive games on a modern PC? Yes, with nGlide or dgVoodoo2 wrappers. These intercept Glide API calls and translate them to Direct3D or Vulkan on modern hardware. Per the dgVoodoo2 compatibility list, the wrapper handles 95%+ of Glide-era titles including Unreal, Wing Commander Prophecy, and the original Tomb Raider. The aesthetic isn't perfectly authentic — modern GPUs render the same scenes with higher precision — but compatibility is excellent.
Is it worth chasing a real Voodoo card for a period-correct build in 2026? For pure nostalgia and a 1996–2001 build target, yes. Voodoo2 cards still go for $80–$200 on eBay; Voodoo5 5500 PCI cards command $400–$700; the unicorn Voodoo5 6000 prototype runs into the thousands. For practical gaming, a GeForce 4 Ti or Radeon 9700 Pro in the same period-rig era delivers better Direct3D performance with broader title support. The Voodoo chase is collector-driven, not performance-driven.
How does the Glide vs Direct3D fight rhyme with today's CUDA vs ROCm? Closely. CUDA, like Glide, is a proprietary API with a strong developer-tooling moat and a hardware company controlling the spec. ROCm, like early Direct3D, is the open competitor catching up generation by generation while a third party (AMD here, Microsoft then) drives ecosystem investment. The historical lesson: closed APIs win the race for 3–5 years, then ecosystem gravity flips. Watch ROCm 7+ and Intel oneAPI for the 2028 inflection.
