Watch completebenchmark runs.
Each of these is a real session captured end to end: same prompts, same models, same passing criteria, against a GPU or hosted-API baseline. Where a clip is sped up, the badge says so.
Repeatability, concurrency, and cost, in that order.
Start with repeatability. It is the one that surprises people, because most inference stacks quietly give you a different answer under load.
Deterministic, repeatable output in live Cursor sessions
The same coding session run repeatedly produces identical output, at full throughput. GPU output varies under higher load and batching.
- Hardware
- Score on Xeon versus a GPU baseline
- Workload
- Live Cursor coding sessions, repeated
8 concurrent coding sessions versus 2 on GPU
Score-SDK holds eight concurrent Cursor-like coding sessions on a single CPU server where the GPU baseline manages two.
- Hardware
- One CPU server versus a GPU baseline
- Workload
- Eight concurrent coding sessions
3× more concurrent knowledge-search sessions
Six concurrent RAG knowledge-search sessions on Score versus two on the GPU baseline.
- Hardware
- Score on Xeon versus a GPU baseline
- Workload
- Six concurrent knowledge-search sessions
Fast, cost-efficient RAG agents on Score Engine
Concurrent retrieval-augmented agents running against a self-hosted open-weight model.
- Hardware
- Score Engine, self-hosted
- Workload
- Concurrent retrieval-augmented agents
RAG cost: self-hosted gpt-oss-120B on Gen4 Xeon vs Amazon Bedrock
Direct cost comparison for the same retrieval workload, self-hosted against a hosted API.
- Model
- gpt-oss-120B
- Hardware
- Gen4 Xeon versus Amazon Bedrock
- Workload
- Retrieval-augmented generation, cost comparison
Coding agent: Gemma 4-31B on Gen6 Xeon vs OpenRouter
Multi-turn coding agent, self-hosted with Score-SDK, against the same model served by a hosted API.
- Model
- Gemma 4-31B
- Hardware
- Gen6 Xeon versus OpenRouter
- Workload
- Multi-turn coding agent, cost comparison
Want this run on your workload instead?
We'll record the same comparison using your models, your context lengths and your concurrency, and hand you the numbers either way.