FOR IMMEDIATE RELEASE
AT A GLANCE
First product, Bud Sentinel, ships 23 guardrail models that run 2.3X faster on a laptop CPU than competing systems on a $15,000 NVIDIA A100 GPU
8.39 ms: guardrail classification on a laptop CPU (incl. network overhead)
2.3× faster: than competitors running on a $15,000 NVIDIA A100 GPU
0.70 ms: p50 latency at 8K tokens with 10,000 concurrent requests (Xeon 8592+)
23 models, 33 variants: security, safety, toxicity, compliance, and quality guardrails
$0 GPU cost: production-grade guardrails on commodity hardware enterprises already own
Bud Ecosystem today revealed what the AI Safety industry has been getting wrong: you don’t need a $15,000 GPU to run production AI Safety. The company launched Resource Aware Attention, a new model architecture built from scratch for CPUs, alongside Bud Sentinel (23 guardrail models) and the Bud Guardrail Gateway, a production-grade gRPC serving infrastructure for deploying guardrails at scale — all running on commodity hardware.
The result: state-of-the-art AI safety at sub-millisecond latency and thousands of requests per second — on the CPUs organisations already own. No GPUs required.
The Problem: AI Safety Has a $15,000 Hardware Tax
Here’s what nobody talks about: every serious GenAI deployment needs guardrails — classifiers that catch toxicity, jailbreaks, and prompt injections on every request. The leading models (Meta’s Prompt Guard, ArchGuard, PIGuard) all assume you have a GPU. On an NVIDIA A100, they manage 18–19ms per classification. On the CPUs that 99% of organisations actually run? Latency explodes to 300–800ms. Unusable.
It gets worse. These models cap at 512 tokens. Real-world prompts routinely exceed 8,000. That means 16 parallel GPU workers per classification — turning a single guardrail check into a distributed computing problem that costs as much as the language model itself.
The Fix: Stop Porting GPU Architectures. Design for CPUs.
The conventional approach is to take GPU-native architectures and squeeze them onto CPUs. Diminishing returns every time. Bud Ecosystem’s research team took the opposite path: they built a new attention mechanism from scratch, designed around what CPUs actually excel at — cache locality, NUMA topology, and SIMD vector instructions.
Resource Aware Attention is none of the usual tricks — not compression, not quantisation, not a runtime hack. It computes attention natively within CPU memory hierarchies. GPU-class performance on commodity hardware, no accuracy loss. And it processes up to 65,536 tokens in a single pass — no chunking, no windowing, no parallel workers.
KEY INSIGHT : When you design attention for the hardware the world actually has — instead of porting architectures designed for hardware it doesn’t — you don’t just close the GPU gap. You eliminate it.
Bud Sentinel: 23 Models, Zero GPUs
Bud Sentinel is the first product built on Resource Aware Attention: 23 specialised models across 33 variants, covering five categories of content safety and compliance:
| Category | Models | Capabilities |
| Security | 3 | Jailbreak detection, prompt injection defence, secrets/credential detection |
| Safety | 4 | Content moderation, content safety, suicide and self-harm detection, drug enablement detection |
| Toxicity | 8 | Social media toxicity, toxic conversation detection, hate speech, abuse, profanity, obscenity, insults, threats, identity attacks, impoliteness |
| Compliance | 6 | PII detection across 11 regions (AU, US, ES, IN, FI, IT, KR, SG, PL, UK, General), illegal activity, political content, regulated advice, bias detection |
| Quality | 2 | Spam detection, domain-specific QA validation |
Every model runs on commodity CPUs at production scale. Zero infrastructure is required at any point in the pipeline — from inference to serving to scaling.
The Trade-Off Everyone Accepted Was a False Choice
The guardrail industry accepted a false trade-off: catch attacks and block real users, or let users through and miss attacks. Pick one. Bud Sentinel refuses to choose. It delivers the highest balanced accuracy of any guardrail system tested across four independent benchmarks: JailBreakBench, PIGuard, WildJailbreak, and Qualifire PI.
Overall Accuracy (Weighted Across All Benchmarks)
| Model | ASR ↓ | FRR ↓ | Balanced | Rank |
| Bud Sentinel | 15.97% | 14.92% | 84.56% | #1 |
| PIGuard | 25.01% | 25.86% | 74.57% | #2 |
| ProtectAI V2 | 36.35% | 24.39% | 69.63% | #3 |
| Prompt Guard 2 (86M) | 34.68% | 15.30% | 75.01% | #4 |
| ArchGuard | 5.40% | 81.65% | 56.48% | #5 |
| Prompt Guard (86M) | 5.83% | 89.35% | 52.41% | #6 |
ASR (Attack Success Rate): percentage of attacks that bypass the guardrail — lower is better. FRR (False Refusal Rate): percentage of legitimate requests incorrectly blocked — lower is better.
Prompt Guard and ArchGuard block attacks well (5–6% ASR) but reject 81–89% of legitimate users. Unusable in production. ProtectAI and Prompt Guard 2 keep false refusals low but let 34–36% of attacks sail through. Sentinel hits 15.97% ASR and 14.92% FRR simultaneously — the only model that performs on both metrics.
Performance: Your Laptop Beats a $15,000 GPU
Laptop CPU vs. $15,000 GPU
Per-classification latency at 512 tokens. Sentinel includes full gRPC round-trip overhead. Competitors benchmarked with raw inference only (no network overhead) — and they still lose.
| Model | Hardware | Latency | vs. Sentinel |
| Bud Sentinel | i7 Laptop CPU | 8.39 ms | — |
| Prompt Guard 2 | A100 GPU ($15K) | 18.52 ms | 2.2× slower |
| ArchGuard | A100 GPU ($15K) | 19.07 ms | 2.3× slower |
| PIGuard | A100 GPU ($15K) | 19.00 ms | 2.3× slower |
| ProtectAI V2 | A100 GPU ($15K) | 19.13 ms | 2.3× slower |
Same CPU, Different Architecture
Same chip. Same conditions. On an Intel Xeon 8272CL, competing models degrade from 18–19ms (GPU) to 334–402ms — a 56–67× performance collapse. Sentinel hits 5.99ms on the same hardware.
| Model | Xeon 8272CL | A100 GPU | CPU/GPU Ratio |
| Bud Sentinel | 5.99 ms | N/A (no GPU) | — |
| Prompt Guard 2 | 334 ms | 18.52 ms | 18× degradation |
| ArchGuard | 380 ms | 19.07 ms | 20× degradation |
| Prompt Guard (86M) | 402 ms | 18.92 ms | 21× degradation |
Production Throughput
On server-grade CPUs, Sentinel matches throughput that normally demands dedicated GPU clusters. Benchmarks on Intel Xeon Platinum 8592+ (256 vCPU) with the Bud Guardrail Gateway:
| Seq Length | Concurrency | p50 Latency | Throughput | Per 1K Tokens |
| 8,192 | 10,000 | 0.70 ms | 1,432 req/s | 0.09 ms |
| 16,384 | 10,000 | 1.31 ms | 761 req/s | 0.08 ms |
| 65,536 | 10,000 | 11.29 ms | 89 req/s | 0.17 ms |
At 8,192 tokens and 10,000 concurrent connections: 1,432 classifications per second, 0.70ms p50 latency. Entirely on CPU. One server. 124 million classifications per day. Estimated cost: $0.10 per million.
The Bud Guardrail Gateway
Bud Ecosystem also ships the Bud Guardrail Gateway — a single binary that serves all 23 Sentinel models with unified gRPC API, model loading, request routing, batching, health checks, and horizontal scaling. Deploy anywhere: public cloud, private cloud, on-premise, air-gapped, or edge.
- Single binary deployment — all 23 models served from one process with unified API
- Native long-context support — processes up to 65,536 tokens in a single pass without chunking or parallel workers
- Concurrency-optimised — p50 latency improves under load (0.70ms at 10K concurrent vs. 1.63ms at 100)
- Hardware agnostic — validated on Intel Xeon, AMD EPYC, and consumer-grade Intel Core processors
- Enterprise-ready — gRPC protocol, health checks, observability hooks, and integration with Bud SENTRY governance framework
The Bud Guardrails Dataset
No existing dataset covered the threat landscape these models needed. So Bud Ecosystem built one: over 4 million labelled rows spanning toxicity, jailbreak attacks, prompt injections, and adversarial perturbations — believed to be the world’s largest open guardrails dataset. It underpins all 23 Sentinel models.
Beyond Guardrails: Every GPU Primitive, Rebuilt for CPUs
Sentinel is first, not last. Bud Ecosystem plans to apply Resource Aware Attention across the GenAI stack:
- Embeddings — CPU-native embedding models for retrieval and similarity search
- Re-rankers — CPU-native cross-encoder re-ranking for RAG pipelines
- Routing — intelligent request routing and intent classification
- Compression — context compression and summarisation
- Caching — semantic caching for repeated query optimisation
The bet: every GenAI primitive that currently demands GPU acceleration can run on commodity hardware. Your infrastructure becomes your AI advantage.
What This Unlocks
Sub-millisecond guardrails on commodity CPUs don’t just cut costs. They make deployment patterns possible that were impossible before:
- Edge deployment: Guardrails on phones, IoT gateways, and edge nodes — for the first time. ~25ms on ARM/edge CPUs versus ~2,400ms for competing models on the same hardware.
- Always-on agent monitoring: Guardrails on every agent action, tool call, and output. 124 million classifications per day. One CPU server. $0.10 per million.
- Sovereign and air-gapped deployments: Government, defence, and regulated industries deploy full guardrail stacks on existing infrastructure. No GPU procurement. No external cloud dependency.
- 15–18× cost-performance improvement: CPU cloud instances at ~$0.50/hour versus A100 instances at ~$2–3/hour, with Sentinel on CPU running 2.3× faster than competitors on A100. The cost advantage compounds at scale.
About Bud Ecosystem
Bud Ecosystem builds sovereign AI infrastructure that runs on any hardware. The platform spans the full Enterprise AI lifecycle: model training (Model Foundry), deployment (AI Foundry), GPU virtualisation (FCSP), developer tools (Pod), AI Consumption (Studio), enterprise agents (Bud Agent), tool integration (MCP Foundry), and governance (SENTRY). Hardware-agnostic across Intel, AMD, Qualcomm, and NVIDIA silicon. Production-grade AI on without lock in.
MEDIA CONTACT
Bud Ecosystem Communications
marketing@budecosystem.com
All performance benchmarks cited in this release were conducted by Bud Ecosystem’s research team. Competitor model results reflect publicly available architectures evaluated under controlled conditions. Resource Aware Attention, Bud Sentinel, Bud Guardrail Gateway, and Bud Ecosystem are trademarks of Bud Ecosystem. All other trademarks are the property of their respective owners.
.png)
