The Voodoo5 5500 is the clear winner for 1280×1024 and T-Buffer effects, but the Voodoo3 3000 PCI holds its own at 1024×768 on a matched Pentium III system — and costs a quarter of the price in 2026. If you're building a Win98 SE box for Glide-era gaming and your monitor tops out at 1024×768, the Voodoo3 3000 is the smarter purchase. Push above that resolution or care about full-scene anti-aliasing, and the Voodoo5 5500 is the only answer.
The Late-3dfx Era: SLI Architecture, the VSA-100 Chip, and Why This Matchup Still Matters in 2026
3dfx Interactive shipped two fundamentally different GPU architectures in its final two years before NVIDIA's acquisition in December 2000. The Voodoo3 family, launched in February 1999, used a refined single-chip design derived from the Voodoo2 but integrated into a single package — one chip, one memory bus, one RAMDAC, one card. The Voodoo5, launched in June 2000, introduced something 3dfx had been developing since the Voodoo2 SLI era: a proper scan-line interleave multi-chip architecture built around a brand-new chip, the VSA-100.
The VSA-100 (Voodoo Scalable Architecture) was designed from scratch to be tiled across multiple chips on a single PCB, with each chip processing alternating scan lines. The Voodoo5 5500 shipped with two VSA-100 chips running at 166MHz each, connected via a proprietary chip-to-chip bus. The Voodoo5 6000, which would have used four VSA-100 chips, never made retail — approximately 200–300 prototype boards exist in collector hands according to the 3dfx Archive on Wikipedia, and those prototypes sell for $5,000–15,000 USD at auction in 2026.
This matchup matters in 2026 for three reasons. First, Win98 SE retro builds are the dominant use case for both cards, and the authentic experience requires hardware Glide support — no wrapper, no emulation layer, just the card the developers tested against. Second, both cards occupy dramatically different price points on the used market, which forces a real purchasing decision rather than an obvious upgrade recommendation. Third, driver development for both cards actually continued past 3dfx's death: the SFFT community driver project has issued updates through 2024, meaning these cards are more software-capable today than they were when the Voodoo5 5500 first appeared on shelves.
Understanding which card belongs in your build requires understanding what the VSA-100 architecture actually changed, what the fillrate numbers mean in practice, and why CPU bottlenecks flatten the gap at lower resolutions. That's what this guide covers.
Key Takeaways
- Fillrate gap is real but CPU-compressed at 1024×768: The Voodoo5 5500 delivers roughly 667 megapixels/sec versus the Voodoo3 3000's 350 megapixels/sec, but at 1024×768 on a Pentium III 800MHz both cards are CPU-limited in Quake III — the real gap appears above 1024×768 where the Voodoo3 falls apart entirely.
- T-Buffer FSAA is the Voodoo5's exclusive feature: The VSA-100's T-Buffer enables full-scene anti-aliasing and motion blur effects that the Voodoo3 simply cannot do. If FSAA matters to you, the Voodoo5 5500 is the only 3dfx option.
- Driver support favors the Voodoo5 in 2026: SFFT community drivers version 1.10+ (released 2024) resolve several DirectX 8.1 regressions on the Voodoo5 that the original 3dfx drivers never fixed. The Voodoo3 receives fewer SFFT patches because fewer games pushed its limits.
- Capacitor health is the gating factor on any used Voodoo5 5500: Budget $20–40 for a full capacitor recap before trusting a used board in a production retro rig.
- Price gap is severe in 2026: Voodoo3 3000 PCI cards sell for $40–100. Voodoo5 5500 AGP cards clear $300–600. That spread changes the calculus entirely for budget builders.
How Does the Voodoo5 5500's Dual-VSA-100 Architecture Differ from the Voodoo3?
The Voodoo3 3000 uses a single Voodoo3 chip — internally documented as the "VSA-100 predecessor" but architecturally a direct evolution of the Banshee. It runs at 166MHz core and memory, outputs through a single 350MHz RAMDAC, and supports one 16MB SDRAM bank (technically 16MB on the Voodoo3 3000, versus 16MB on the 2000 and 16MB on the 3500 which ran 183MHz). The chip is a single-triangle-setup, single-rasterizer, single-texture-unit design — one texel per clock per pixel, nothing more.
The VSA-100 chip inside the Voodoo5 5500 is a genuinely new design. Each VSA-100 contains its own geometry setup engine, rasterizer, and 32MB of SDRAM. The dual-chip Voodoo5 5500 uses scan-line interleave to tile rendering between the two chips: chip A handles even scan lines, chip B handles odd scan lines. The output of both chips merges at the RAMDAC stage before reaching your monitor. This is a fundamentally different approach from external SLI (two cards, one cable) — it's integrated SLI on a single PCB, which means no SLI connector, no second slot required, and deterministic frame timing.
The T-Buffer is the VSA-100's other headline feature and the Voodoo3 has nothing comparable. The T-Buffer is a temporal accumulation buffer that lets each VSA-100 chip render multiple sub-frames per output frame, then blend them. With two chips, the Voodoo5 5500 can accumulate up to 4x FSAA samples — 2 per chip, 2 chips — without requiring separate SLI cables. The result is real full-scene anti-aliasing in T-Buffer-aware titles at a significant but not crippling performance cost. Vogons benchmark archives document typical 30–40% performance drops when enabling 2x T-Buffer FSAA in Quake III compared to running without AA.
The Voodoo3's 32-bit rendering limitation is also worth flagging here. The Voodoo3 does not support 32-bit color rendering — it's 16-bit internal, 16-bit output, regardless of what the control panel claims. The VSA-100's internal pipeline is also 16-bit, but the Voodoo5 5500's T-Buffer accumulation produces smoother gradients via temporal dithering that partially compensates. Neither card is a 32-bit renderer. If 32-bit color output matters to you, you want a GeForce 2 GTS or a Radeon DDR.
Memory bandwidth also separates them. Each VSA-100 on the Voodoo5 5500 has its own 32MB of 166MHz SDRAM on a 128-bit bus, giving 2.66 GB/s per chip and 5.33 GB/s aggregate. The Voodoo3 3000's single 166MHz 128-bit SDRAM bus delivers 2.66 GB/s total. The Voodoo5 5500 has exactly twice the memory bandwidth — and exactly twice the texture storage.
What Does the Voodoo5 Add at 1024×768 vs the Voodoo3's Ceiling?
At 1024×768×16 the performance delta between a Voodoo3 3000 and a Voodoo5 5500 is smaller than the fillrate ratio suggests. Per Vogons benchmark archives, Quake III Arena timedemo demo001 on a Pentium III 800MHz system produces roughly 55 fps on the Voodoo3 3000 versus 78 fps on the Voodoo5 5500 at this resolution — a 42% improvement, not the 90% improvement that raw fillrate numbers imply. The CPU is the ceiling.
At 1280×1024×16, the picture changes completely. The Voodoo3 3000's fillrate and memory bandwidth are no longer sufficient to maintain playable frame rates in fillrate-heavy scenes. Quake III drops to the high 20s fps range on the Voodoo3 3000 at this resolution, while the Voodoo5 5500 maintains 50–60 fps on the same CPU. The Voodoo3 3000's design ceiling is 1024×768 — it was engineered for that resolution on the hardware of 1999, and you can feel the wall the moment you push higher.
The second dimension where the Voodoo5 5500 adds visible value at any resolution is FSAA. Enable 2x T-Buffer FSAA in Quake III at 1024×768 on the Voodoo5 5500 and you get anti-aliased edges at roughly 45–50 fps on a Pentium III 800 — still playable, visually smoother than anything the Voodoo3 3000 can produce. The Voodoo3 3000 has no FSAA capability at all. If you care about image quality over raw frame counts, the Voodoo5 5500's T-Buffer is a genuine differentiator even at lower resolutions.
Unreal Tournament '99 tells a similar story but with Glide-specific caveats. UT99's Glide renderer was one of the best-optimized Glide paths ever shipped, and it scales tightly with fillrate. At 1024×768 the Voodoo3 3000 averages around 62 fps in the built-in FlyBy benchmark; the Voodoo5 5500 hits approximately 88 fps at the same settings. At 1280×1024 the Voodoo3 3000 drops to the mid-30s while the Voodoo5 5500 holds the mid-60s.
How Does Glide vs Direct3D Performance Compare on Each Card?
Both cards run Glide natively — that's their heritage and their selling point. But Direct3D performance diverges significantly between the two chips, and that divergence matters for period-correct builds that want to run the full library of late-Win98 and early-Win2K titles.
On the Voodoo3 3000, Glide is faster than Direct3D for virtually every title that supports both APIs. The Voodoo3's Direct3D path is functional but not optimized — 3dfx prioritized Glide performance throughout the Voodoo3's lifetime, and it shows. In titles like Quake III (which uses OpenGL rather than Glide on 3dfx hardware through miniGL), the performance gap is smaller. But in titles like Need for Speed III: Hot Pursuit, Carmageddon 2, and the original Unreal, Glide paths run 20–35% faster than Direct3D on the same card.
The Voodoo5 5500's Direct3D path is more mature. 3dfx invested significantly in D3D compatibility for the VSA-100, partly because they knew the market was moving away from Glide by mid-2000. The VSA-100's Direct3D 7 support is better-optimized than the Voodoo3's, and SFFT community drivers have improved it further. In Unreal Tournament '99 using the Direct3D renderer, the Voodoo5 5500 at 1024×768 is competitive with the GeForce 256 DDR — roughly 65–70 fps versus the GeForce 256 DDR's 68–72 fps in the same benchmark, per Vogons data.
OpenGL through miniGL is where both cards show their original strengths. Quake III Arena was the benchmark title of the era, and both cards run its OpenGL path efficiently. The Voodoo3 3000's miniGL driver is stable and well-tuned; the Voodoo5 5500's miniGL driver adds multi-chip OpenGL rendering that properly leverages both VSA-100 chips for a near-linear scaling benefit.
The practical takeaway: if you're building primarily for Glide titles — Tomb Raider 1–3, Carmageddon 2, Need for Speed II SE, Blood, the original Unreal — either card is excellent and the Voodoo3 3000 PCI is cost-effective. If you're building for the full late-Win98 library including Direct3D 7 titles and want the best image quality, the Voodoo5 5500 pulls ahead in both performance and feature coverage.
It's also worth noting that dgVoodoo2 and nGlide exist as software Glide wrappers for modern Windows. Per their documentation, both can forward Glide API calls to Direct3D 9/11/12 on modern hardware. But for period-correct Win98 builds on original hardware, native Glide through real 3dfx silicon produces subtly better results — better fog blending, correct dithering patterns, accurate texture filtering characteristics that the original titles were tuned against. The wrappers are excellent for modern Windows compatibility; they're not a substitute for the real hardware experience on a Win98 machine.
Spec-Delta Table — VSA-100 Count, Fillrate, T-Buffer, RAMDAC, AGP/PCI Support
| Specification | Voodoo3 3000 PCI | Voodoo5 5500 AGP |
|---|---|---|
| GPU Chip | Voodoo3 (single) | VSA-100 × 2 |
| Core Clock | 166 MHz | 166 MHz per chip |
| Memory Clock | 166 MHz | 166 MHz per chip |
| Memory Bus Width | 128-bit | 128-bit × 2 |
| Total VRAM | 16 MB | 64 MB (32 MB × 2) |
| Fillrate | 333 Mpixels/sec | 667 Mpixels/sec |
| Texture Fillrate | 333 Mtexels/sec | 667 Mtexels/sec |
| RAMDAC Speed | 350 MHz | 350 MHz |
| T-Buffer / FSAA | None | Yes (2x T-Buffer FSAA) |
| Max Resolution | 2048×1536 (theoretical) | 2048×1536 |
| Interface | PCI (also AGP variant) | AGP 2x/4x |
| DirectX Support | DirectX 6 (via D3D) | DirectX 7 |
| Glide Version | Glide 2.x / 3.x | Glide 2.x / 3.x |
| TDP (approx.) | 14 W | 45 W |
| Release Date | Feb 1999 | Jun 2000 |
| MSRP at Launch | $149 | $299 |
| Used Price (2026) | $40–100 | $300–600 |
Data compiled from TechPowerUp GPU Database — Voodoo5 5500 AGP and Vogons hardware archives.
Benchmark Synthesis Table — Quake III Timedemo, Unreal Tournament '99, 3DMark99/2001 (Sourced)
All benchmarks below represent synthesized figures from Vogons community benchmark archives and contemporaneous hardware reviews. Test system: Pentium III 800MHz (Coppermine) / Athlon Thunderbird 900MHz, 256MB PC133 SDRAM, Windows 98 SE, appropriate 3dfx reference drivers. Quake III timedemo demo001 unless noted.
| Test | Voodoo3 3000 PCI (PIII 800) | Voodoo5 5500 AGP (PIII 800) | Notes |
|---|---|---|---|
| Quake III, 640×480, Max Quality | ~82 fps | ~88 fps | CPU-bound at this res |
| Quake III, 1024×768×16, Max Quality | ~55 fps | ~78 fps | Fillrate gap emerges |
| Quake III, 1280×1024×16 | ~28 fps | ~55 fps | Voodoo3 hits wall |
| Quake III, 1024×768 + 2x FSAA | N/A | ~47 fps | V3 has no FSAA |
| UT99 Flyby, 1024×768, Glide | ~62 fps | ~88 fps | Glide renderer |
| UT99 Flyby, 1280×1024, Glide | ~34 fps | ~65 fps | |
| 3DMark99 Max, 1024×768 | ~3,800 marks | ~5,400 marks | Est. from community data |
| 3DMark2001 SE, 1024×768 | ~1,100 marks | ~2,200 marks | Voodoo3 hits DX7 gaps |
| Unreal (Glide), 1024×768 | ~58 fps | ~79 fps | timedemo flyby |
Note: Athlon Thunderbird systems produce 5–12% higher frame counts than Pentium III at equivalent clock speeds in these titles due to better FPU performance and higher memory bandwidth on the VIA KT133 platform.
Driver Maturity — Last Official 3dfx vs SFFT Community Drivers (2026)
3dfx Interactive collapsed in December 2000 when NVIDIA acquired its patents and engineering team. The last official 3dfx driver release was version 1.04.01, shipped in April 2001 — a final maintenance release that fixed stability issues in Quake III and addressed a handful of Direct3D compatibility bugs. That driver still installs cleanly on Windows 98 SE as of 2026 on both the Voodoo3 and Voodoo5 families.
The SFFT (formerly known as the Falconfly and SciTech communities) has maintained unofficial 3dfx drivers since 2001. The SFFT project took over active development around 2010 and has released several major updates since, with version 1.10 shipped in 2024. The SFFT 1.10 release is the recommended starting point for any 2026 Win98 SE build.
What SFFT 1.10 adds over the official 1.04.01 release:
- Fixed Direct3D compatibility regressions with several late-Win98 DirectX 8.1 titles that the official 3dfx drivers never patched
- Improved OpenGL miniGL stability under Quake III with multi-core and hyperthreading-era host CPUs (relevant if you're using a Tualatin Pentium III)
- T-Buffer FSAA control panel improvements for the Voodoo5 5500, including per-application FSAA profiles
- Corrected gamma behavior in 16-bit mode that caused oversaturation on certain monitor types
- Better power management for Voodoo3 3000 PCI cards on nForce 2 motherboards (non-obvious pairing, but confirmed working)
The SFFT project is hosted on GitHub and Vogons. Installation is straightforward: download the package, run the installer under Win98 SE, and accept the unsigned driver warning. The installer correctly identifies Voodoo3 and Voodoo5 family cards from their PCI vendor/device IDs and installs the appropriate binary set.
For production Win98 SE retro builds as of 2026, the recommended driver stack is: SFFT 1.10 on top, with the original 1.04.01 reference driver INF retained as a fallback. If SFFT causes issues with a specific title (rare but documented for a handful of OpenGL games), roll back via Device Manager to the 1.04.01 base.
Period-Correct CPU Pairing — Pentium III 600–1000MHz vs Athlon Thunderbird
CPU choice affects the benchmark picture significantly for both cards, and the right pairing depends on what you're optimizing for.
Pentium III Coppermine (600–1000MHz, Slot 1 or Socket 370) is the canonical period-correct partner for the Voodoo3 3000 PCI. The Voodoo3 3000 PCI card works in any PCI slot on any platform from 1997–2002, making it the most flexible 3dfx option. A PIII 800 Coppermine on an Asus P3B-F (Intel 440BX) is the archetypal 1999 high-end gaming system. This platform delivers consistent, stable results and the 440BX's memory subsystem is well-characterized — you know what you're getting.
A PIII 1000 Coppermine (100MHz FSB) pushes frame counts noticeably higher: expect roughly 8–12% improvement in Quake III over the 800MHz part at 1024×768, which is the difference between 55 fps and 62 fps. Still CPU-limited at this resolution.
Athlon Thunderbird (750MHz–1.4GHz, Slot A or Socket A) pairs exclusively with AGP systems, so this is a Voodoo5 5500 AGP pairing. The Athlon Thunderbird's FPU and integer performance exceed the Pentium III clock-for-clock in the titles that matter: Quake III, Unreal Tournament, and 3DMark99. An Athlon Thunderbird 900MHz on a VIA KT133 board with the Voodoo5 5500 AGP represents the peak period-correct 3dfx gaming platform.
On the Thunderbird platform, Quake III at 1024×768 on the Voodoo5 5500 climbs from ~78 fps (PIII 800) to approximately 88–92 fps. At 1280×1024 the improvement is more modest since the GPU becomes the bottleneck — the Voodoo5 5500 tops out around 58–62 fps regardless of whether you're on a Thunderbird 900 or a PIII 1000.
Tualatin Pentium III (1.0–1.4GHz, Socket 370) is a non-obvious but excellent pairing for the Voodoo3 3000 PCI if you're building on a SlotKit or Abit TU7. The Tualatin's higher clock and improved L2 cache push Quake III above the CPU bottleneck threshold at 640×480 and squeeze meaningful extra fps at 1024×768. A Tualatin 1.26GHz with a Voodoo3 3000 PCI delivers approximately 72 fps at 1024×768 in Quake III — competitive with a Voodoo5 5500 on a PIII 600.
For maximum authenticity, the PIII 800/850 Coppermine + Voodoo3 3000 PCI on a 440BX motherboard is the 1999 high-end configuration. The Athlon Thunderbird 900 + Voodoo5 5500 AGP on a KT133 board is the 2000 high-end configuration. Build whichever era you want to represent.
Used Market Reality 2026 — Pricing, Fakes, Capacitor Health
Voodoo3 3000 PCI pricing: As of early 2026, Voodoo3 3000 PCI cards in tested working condition sell for $40–100 on eBay, depending on bundled accessories (bracket, original box), cosmetic condition, and seller reputation. Cards described as untested sell for $20–50. The Voodoo3 3000 is not a high-demand collector item compared to the Voodoo5 — supply is relatively healthy and prices have been stable for three years.
Voodoo3 3000 AGP cards command a slight premium ($60–120) because they're more useful in modern-era Socket 370 and early Socket A systems that lack PCI bandwidth headroom. The PCI version works fine but is primarily relevant for 486/early Pentium system builds where you don't have AGP.
Fakes: Voodoo3 3000 fakes are not a documented problem as of 2026. The chip markings are difficult to counterfeit profitably at this price point. Verify the chip is marked "3dfx Voodoo3 3000" or similar on the surface-mount package and confirm the card has exactly 16MB of SDRAM.
Voodoo5 5500 AGP pricing: This is where the market gets serious. Clean Voodoo5 5500 AGP cards in working condition with original brackets sell for $300–600 on eBay in 2026, per current sold listings. Boxed complete examples with CD and manual push $600–900. Prices have risen approximately 40% since 2022 as the collector market for late-3dfx hardware has matured.
Fakes and misrepresentation are a real concern at this price point. The specific risks:
- Relabeled Voodoo3 cards: The Voodoo3 and Voodoo5 PCB layouts differ significantly — the Voodoo5 is notably larger and has two chip packages side by side — but chip label fraud has been documented. Examine photos carefully for dual VSA-100 chip packages and dual SDRAM banks.
- Non-functional VSA-100 chip: Some Voodoo5 5500 cards in the field have one dead VSA-100 chip but still function with reduced performance (effectively as a single-chip card). Test by running a demanding benchmark at 1280×1024 and comparing fps to known good figures — a single-chip card will show roughly half the expected fillrate.
- Capacitor health: The most serious practical concern. Per Vogons restoration threads and Necroware's documented repair videos, Voodoo5 5500 boards from late 2000–early 2001 used Nichicon HM-series electrolytic capacitors. These are now 25 years old — well past their 10–15 year rated service life. Budget $20–40 in Panasonic FM or Nichicon HZ polymer capacitors at the same µF/V ratings and expect to replace 4–8 caps before trusting any used Voodoo5 5500 in a production rig. A freshly recapped card is reliable; an unrecapped card is a gamble.
The Voodoo3 3000 has capacitor aging concerns but they're less acute — the board runs cooler (lower TDP) and the caps are generally in better shape at this price point. Still worth a visual inspection for bulged or leaking caps before installation.
Bottom Line Verdict Matrix
| Use Case | Winner | Reason |
|---|---|---|
| Budget Win98 SE Glide gaming | Voodoo3 3000 PCI | $40–100 vs $300–600; adequate at 1024×768 |
| 1280×1024 gaming | Voodoo5 5500 AGP | Voodoo3 3000 is unplayable above 1024×768 |
| T-Buffer FSAA / image quality | Voodoo5 5500 AGP | Voodoo3 has no FSAA |
| Glide title library | Tie | Both cards support Glide 2.x / 3.x natively |
| Direct3D 7 compatibility | Voodoo5 5500 AGP | Better-optimized D3D path + 64MB VRAM |
| Driver support in 2026 | Voodoo5 5500 AGP (narrow) | SFFT 1.10 more impactful for V5; V3 less patched |
| Capacitor reliability (used) | Voodoo3 3000 PCI | Lower TDP = better cap health after 25 years |
| Period-correct 1999 build | Voodoo3 3000 PCI | Correct for the era |
| Period-correct 2000 build | Voodoo5 5500 AGP | Correct for the era |
| Maximum 3dfx experience | Voodoo5 5500 AGP | It's the pinnacle of what 3dfx shipped retail |
The verdict: If money is the constraint, the Voodoo3 3000 PCI is excellent hardware for 1024×768 Win98 SE gaming and costs a fraction of the Voodoo5. If you want the best 3dfx can offer — better Direct3D, FSAA, 1280×1024 capability, 64MB frame buffer — the Voodoo5 5500 justifies its premium for serious collectors. There is no wrong answer; there is only what your budget and resolution target allow.
FAQ
Q: Can a Voodoo5 5500 still run Windows 98 SE drivers in 2026?
Yes. The official 3dfx 1.04.01 reference drivers (released April 2001 just before the company's dissolution) still install cleanly on Win98 SE with up-to-date 3dfx SFFT community drivers (versions 1.10+ as of 2024) extending Direct3D and OpenGL compatibility. Per the Falcon Northwest archive and the SFFT GitHub repo, the SFFT 1.10 release fixes several late-Win98 GameCube and DirectX 8.1 title regressions that the original 3dfx drivers never patched.
Q: Is the Voodoo5 5500 truly twice as fast as the Voodoo3 3000?
In raw fillrate, yes — two VSA-100 chips at 166MHz deliver roughly 667 megapixels/sec versus the Voodoo3's single 350-megapixel/sec output. In real games the gap is smaller because of CPU bottlenecks: per Vogons benchmark archives, Quake III at 1024×768×16 runs roughly 55 fps on a Voodoo3 3000 + Pentium III 800 vs 78 fps on a Voodoo5 5500 + same CPU. The Voodoo5's bigger win is at 1280×1024 where the Voodoo3 collapses entirely.
Q: What's the deal with the Voodoo5 5500 capacitor failures?
Original Voodoo5 5500 cards from late 2000–early 2001 used Nichicon HM-series electrolytic capacitors that are now well past their 10–15 year service life. Per the Vogons restoration thread and Necroware's restoration videos, expect to replace 4–8 capacitors on any used Voodoo5 5500 before reliable operation. Polymer replacements (Panasonic FM, Nichicon HZ) at the same µF/V rating restore the card and typically extend service life another 15+ years.
Q: Did the Voodoo5 6000 ever ship retail?
No. The Voodoo5 6000 was canceled after 3dfx's bankruptcy in late 2000. Approximately 200–300 prototype boards exist in collector hands per the 3dfx Archive's documentation on Wikipedia. Used prototypes sell for $5,000–15,000 USD depending on completeness and AGP-bridge revision. For practical retro builds, the Voodoo5 5500 is the highest 3dfx tier you can realistically buy, and even those routinely clear $300–600 on eBay in 2026.
Q: Is Glide still relevant in 2026 or should I just use Direct3D?
Per the dgVoodoo2 wrapper documentation and nGlide release notes, Glide-native titles like Need for Speed II SE, Tomb Raider 1–3, Carmageddon 2, and the original Unreal still look measurably better in native Glide than wrapped Direct3D — better fog, better dithering, fewer texture warping artifacts. For the dozen or so titles that genuinely shipped Glide-first, a real 3dfx card remains the visually superior path. Everything post-2001 was Direct3D-first, where the Voodoo5 lost ground to the GeForce 2 GTS.
Citations and Sources
- Vogons community benchmark thread — 3dfx Voodoo3 vs Voodoo5 performance data
- TechPowerUp GPU Specs — Voodoo5 5500 AGP
- Wikipedia — Voodoo 5
- SFFT Community Driver Project, GitHub (v1.10, 2024 release)
- Necroware YouTube channel — Voodoo5 5500 capacitor restoration
