Welcome to the Silicon Graphics Web Navigation Console. Below you can configure the browsing arsenal that powered the artists, physicists, and dreamers of the ’90s—and pit those teal towers against today’s tensor-farm fantasies. Select a workstation, align an IRIX release, and see how the price/performance story evolved from SGI glam to H100 grind. Warning: by the time you scroll to the price banner you will want an SGI, an H100 rack, or both.
CPUs: MIPS R2000/R3000/R4000 | RAM: 8–256 MB
Browsers: None worth mentioning — this was pre-web
The Personal IRIS (1988) brought SGI graphics to individual desks for the first time. The IRIS Crimson (1992) was the bridge between the 4D series and the Indigo² — a deskside tower with RealityEngine graphics that could fill a room with polygons. Browsing the web? Sir, the web barely exists yet. You're running fm and reading man pages.
CPUs: MIPS R3000/R4000 | RAM: 24–128 MB
Browsers: NCSA Mosaic (Motif), early Arena builds
Static HTML 1.0, inline GIFs, no frames. Perfect for rendering teapots while you read a CERN press release.
R4400 at a blistering 250 MHz. Netscape Navigator 2.x arrives: tables, frames, JavaScript 1.0, and the ability to crash spectacularly on malformed `
Unified Memory Architecture meets Netscape 3.01 Gold. JavaScript 1.1, animated GIFs, and livestreaming QuickTime if you sacrificed enough RAM. Great for coding VRML scenes at 3 a.m.
CPUs: R4400 → R10000/R12000 | RAM: 256 MB – 16 GB
Graphics: RealityEngine2, InfiniteReality, InfiniteReality2
Browsers: Netscape 3–4, eventually Mozilla ports
The visualization supercomputers. The Onyx (1993) powered flight simulators, VR caves, and Hollywood pre-viz. The Onyx2 (1996) with InfiniteReality could push 10M textured polygons/sec and drive multiple high-res displays simultaneously. These were the machines ILM, Pixar, and NASA used when "real-time" meant something. You didn't browse the web on these — you rendered the future.
CPUs: Up to 512× R10000/R12000/R14000 | RAM: Up to 1 TB (ccNUMA)
Topology: Hypercube interconnect, NUMAflex
Browsers: Whatever your terminal server could handle
SGI's answer to "what if we connected 512 MIPS processors with fat pipes and pretended it was one machine?" The Origin 2000 (1996) introduced ccNUMA to the masses — if by "masses" you mean national labs and oil companies. The Origin 3000 pushed it further with R14000 chips. Nobody browsed the web on these; they modeled nuclear stockpiles and weather systems. But somewhere, someone probably ran Lynx on one, just to flex.
Dual-channel graphics, IMPACT rails, Netscape 4.05–4.78. CSS 1, DOM Level 0, and the occasional browser freeze whenever DHTML gets ambitious. Still, you can monitor your render farm and browse Slashdot simultaneously.
Community ports of Mozilla 1.0/1.2 via Nekoware. CSS 2 partial support, better PNG handling, but startup times counted in coffee breaks. Comparable to what a G3 Mac could pull off with Mozilla 1.0.
R16k or Itanium horsepower. Firefox 1.5 ports exist, but most SGI veterans switch to x86/Linux for modern browsing and keep the SGI purely for Maya, Flame, or massive OpenGL sims.
NVIDIA RTX 2080 Ti, RTX 3090. Chrome, Firefox Quantum, and the rise of CUDA. Hobbyists build “AI rigs” that could vaporize an Indigo PSU.
Google said "NVIDIA? We'll build our own." TPU v1 (2016) was inference-only and powered AlphaGo. TPU v2/v3 added training. TPU v4 (2022) formed the backbone of PaLM and Gemini. TPU v5e/v5p (2023) brought cost efficiency, and TPU v6 Trillium (2024) pushes 4.7× the peak compute of v5e. These ASICs live in massive pods — up to 8,960 chips per cluster — and you can't buy them, only rent them on Google Cloud. The silicon equivalent of "nice chip, shame you can't have it."
Amazon saw Google and NVIDIA printing money and said "we have fabrication partners too." Inferentia (2019) handled inference workloads. Trainium (2021) tackled training. Trainium2 (2024) is where it gets serious — powering Anthropic's Claude models across massive UltraClusters of 100,000+ chips connected via custom NeuronLink fabric. You're literally reading the output of these chips right now. The silicon behind the chatbot you're probably using right now.
H100, DGX Spark, eight-way HGX trays, Google TPU pods, and AWS Trainium UltraClusters. PyTorch/LLAMA/Claude on demand. Three silicon empires — NVIDIA, Google, and Amazon — all competing to burn the most electricity per token. You can serve GPT-4 sized models from your garage if your power company approves, rent a TPU pod, or just call an API and let someone else worry about the thermals.
No Internet Explorer. Microsoft never ported IE to IRIX. Netscape dominated, with late-era relief from community Mozilla/Firefox builds. Expect IE5-grade content to render better on your G3 Mac than on an SGI; the Octane2 with Mozilla 1.0 will match a Mac running the same vintage build, but nothing beyond that.
JavaScript Reality. Netscape 4′s JS 1.2 is roughly IE4-level. Dynamic HTML will break. The Retro AI Proxy solves this by rendering psuedo-live UIs that target IE5 features but degrade to iframe polling—exactly the kind of trick an SGI could use to stay useful in 2025.
Use Cases. Monitor SGI news feeds, order visual effects middleware, render check your latest Maya scene, and yes—run your Retro AI Proxy in a teal window, because you can.
Wonder what that teal tower would cost in today’s AI arms race? Adjust the dials below and compare the sticker shock.
$44,900 (1988) / ≈ $112,000 today
R2000 36 MHz, 8–32 MB RAM, GT graphics
AI perf: ≈ 0.000005 tokens/sec — it can barely grep
$80,000+ (1992) / ≈ $170,000 today
R4000 100–150 MHz, 256 MB RAM, RealityEngine
AI perf: ≈ 0.00008 tokens/sec — but those polygons though
$12,995 (1991) / ≈ $26,000 today
R4000 100 MHz, 64 MB RAM, GR2-XZ graphics
AI perf: ≈ 0.00005 tokens/sec — suitable for ASCII palm trees
$74,995 (1999) / ≈ $130,000 today
R12000 400 MHz, MXE graphics, 2 GB RAM
AI perf: ≈ 0.0003 GPT-4 tokens/sec (mostly renders teapots)
$250,000+ (1996) / ≈ $470,000 today
4–8× R10000 195 MHz, InfiniteReality2, 16 GB RAM
AI perf: ≈ 0.001 tokens/sec — but it could render the Matrix
$500,000–$5M+ (1996) / ≈ $1–10M today
Up to 512× R10000/R12000, 1 TB ccNUMA
AI perf: ≈ 0.01 tokens/sec total — nuclear stockpile math, not chatbots
$110,995 (2004) / ≈ $170,000 today
4× R16000 700 MHz, V12 graphics, IRIX 6.5.22
AI perf: ≈ 0.0007 tokens/sec if you bribe IRIX with coffee
≈ $30,000 (2025 street)
80 GB HBM3, 3.5 PFLOPS FP8
AI perf: ≈ 2.5M GPT-style tokens/min
≈ $250,000 barebones
8× H100 NVLink, 28 PFLOPS FP8
AI perf: ≈ 20M tokens/min (and 20k BTU/hr)
$3,999 (desktop kit, 2025)
1× Spark AI SoC, 32 GB HBM3e, 400 W PSU
AI perf: ≈ 400k tokens/min — think “personal GPT lab”
≈ $3,800 (MSRP×2)
48 GB combined GDDR7, 1 kW draw
AI perf: ≈ 1.2M tokens/min with mixed precision
≈ $3,000 (MSRP×2)
48 GB combined GDDR6X, 900 W
AI perf: ≈ 750k tokens/min, still cheaper than therapy
≈ $??? (Cloud rental only, ~$4.20/chip/hr)
459 TFLOPS BF16 per chip, ICI mesh interconnect
AI perf: Trained Gemini Ultra — enough said
≈ $??? (Cloud rental, 2024+)
4.7× peak compute vs v5e, HBM, liquid-cooled
AI perf: Google won't tell you, but it's training the next Gemini right now
≈ $??? (AWS rental, 100k+ chips per cluster)
Custom NeuronLink interconnect, HBM, liquid-cooled
AI perf: Trains Claude. You're reading its homework right now
≈ $0.76/hr per chip (AWS on-demand)
190 TFLOPS BF16, 32 GB HBM per chip
AI perf: Budget inference king — serves models while your wallet survives