Groq LPU
SRAM-based language processing units sold as a token-priced inference cloud.
Full Groq LPU record →Head to head
Groq LPU (Groq) and NVIDIA Blackwell (NVIDIA) both sit in Chips & Compute. NVIDIA Blackwell shipped more tracked changes in the last 30 days (9 vs 4). Every value below comes from the latest crawl of the vendors' own pages.
| Attribute | Groq LPU Groq | NVIDIA Blackwell NVIDIA |
|---|---|---|
| Segment | Chips & Compute | Chips & Compute |
| Version | LPU | Blackwell |
| Entry cost | Pricing not published. Contact sales or check account console for usage-based rates. | Pricing is not publicly listed (enterprise hardware architecture/system) |
| Pricing tiers | GroqCloud Console Free Tier / Pay-As-You-Go free | — |
| Model stack | canopylabs/orpheus-arabic-saudi · canopylabs/orpheus-v1-english · kimi-k2-instruct-0905 · Llama / Kimi / GPT-OSS · Llama 3.1 8B Instant · Llama 3.3 70B Versatile · LPU v2 · meta-llama/llama-4-maverick-17b-128e-instruct · meta-llama/llama-4-scout-17b-16e-instruct · minimax-m2.5 · minimax/minimax-m2.5 · minimaxai/minimax-m2.5 · moonshotai/kimi-k2-instruct-0905 · openai/gpt-oss-120b · openai/gpt-oss-20b · openai/gpt-oss-safeguard-20b · Orpheus English · orpheus-arabic-saudi · orpheus-v1-english · Qwen 3.6 27B · qwen/qwen3-vl-32b-instruct · qwen3-vl-32b-instruct · Whisper V3 Large · whisper-large-v3 | Blackwell · Blackwell B200 · Grace CPU |
| Context window | 131k | — |
| Public API | ||
| Routes models | ||
| Changes / 30d | 4 | 9 |
| Origin | Global | Global |
| Capabilities | deterministic latency, OpenAI-compatible API, open-model catalogue, batch API, Fast LLM inference, OpenAI Compatibility, Prompt Caching, Speech to Text | 72-GPU NVLink domain, FP4 inference, HBM3E 192GB, CUDA ecosystem, Second-generation Transformer Engine, Fifth-generation NVLink, FP4 AI Support, Tensor Cores with microscaling formats |
SRAM-based language processing units sold as a token-priced inference cloud.
Full Groq LPU record →NVL72 rack-scale coherent GPU domains, the default substrate for frontier training and inference.
Full NVIDIA Blackwell record →