Tags
Introducing Runyard Compare: Head-to-Head GPU Benchmarks with TurboQuant
Runyard now has a Compare page. Pick Device A and Device B from a catalogue of 160+ GPUs — Apple Silicon, NVIDIA RTX 50/40/30, AMD RX 7000/9000, Intel Arc, cloud rigs — and see a live score for every LLM across quality, speed, fit, and context. Toggle TurboQuant on Device B with a single switch and watch the scores update in real time. No spreadsheet. No guesswork.
What the Compare Page Does
The Compare page scores every model in the Runyard catalogue — 100+ LLMs — simultaneously across two devices you configure. Each device has a composite score (0–100) built from four dimensions: model quality, inference speed, memory fit, and context window headroom. The winner for each model row is highlighted. Sort by score, speed, model size, or name. Filter by use case. Click any row to see a detailed side-by-side breakdown with dimension bars, donut score, max context, quantization, and cloud run options.
- ▸160+ GPUs in the dropdown — Apple M1 through M5 Ultra, RTX 5090 down to GTX 1050, AMD RX 9000, Intel Arc, mobile SoCs
- ▸Composite score updates live as you change VRAM, RAM, cores, backend, and bandwidth
- ▸Sort by score / speed / model size / name with ascending/descending toggle
- ▸Use-case filter pills — General, Coding, Reasoning, Creative, Multilingual, and more
- ▸Click any model row → full detail panel opens with donut chart, dimension bars, context bar, HuggingFace and Ollama links, cloud provider marquee
The TurboQuant Toggle — and Why It Matters
Device B has a TurboQuant toggle. Flip it on and every score recalculates immediately. TurboQuant (Zandieh et al., ICLR 2026) applies 4× KV cache compression — meaning Device B can run 4× longer context on the same VRAM. That context boost feeds directly into the composite score: higher context headroom raises the Context dimension score, which flows into the overall composite. Models that were ties become Device B wins. Models that were marginal become viable.
When TurboQuant is active, Device B's context score recalculates using the TQ-expanded window (up to 4× the hardware max, capped at the model's spec limit). The composite score reflects that gain. The green glow on Device B's column and the ✦ marker on winning rows make it immediately obvious which models flip from "tie" to "B wins" once TQ is on.
Device A shows a red "No TurboQuant" callout when TQ is active on Device B — so you can see exactly how much context headroom Device A is leaving on the table.
How to Use It
- 1.Go to runyard.dev/compare
- 2.Select your GPU from the Device A dropdown (or use auto-detected hardware)
- 3.Select the GPU you're comparing against in Device B
- 4.Toggle TurboQuant on Device B to see the score boost
- 5.Click any model row to open the full detail panel
- 6.Sort by score to see which device wins overall
Compare your GPU against any other device — with TurboQuant on and off.
Open Runyard Compare → →Tools
Find AI models that fit your exact hardware. Enter your specs and get a ranked list instantly.
Newsletter