Skip to main content

LPU (Language Processing Unit) Architecture

What is an LPU

LPU (Language Processing Unit) is Groq's in-house language processing unit, founded in 2016 by former Google TPU team members. Purpose-built for extreme low-latency LLM inference, featuring 228MB on-chip SRAM per chip (vs GPU 80GB HBM), achieving deterministic latency through on-chip SRAM + compiler scheduling.

Core innovation: Compiler-Defined Hardware — no caches, no out-of-order execution, all latencies are predictable.

Core Architecture: TSP

Tensor Streaming Processor (TSP)

  • Functional units: Matrix Multiply, ReLU, Add, Multiply, Transpose, Shuffle
  • Compiler schedules all operations in advance
  • Data streams through TSP units, no intermediate storage

Compiler

  • GroqWare SDK (Python / C++)
  • Fully offline compilation
  • Dataflow graph maps directly to hardware

On-Chip SRAM

  • 228 MB SRAM (GroqChip v1)
  • 80 TB/s bandwidth (vs HBM 3 TB/s)
  • Deterministic access latency (no cache misses)

LPU vs GPU vs TPU

DimensionLPU (Groq)GPU (H100)TPU (v4)
ArchitectureCompiler-defined streamingSIMT general parallelSystolic array
Memory228MB SRAM80GB HBM32GB HBM
Bandwidth80 TB/s3.35 TB/s1.2 TB/s
LatencyDeterministic, sub-millisecondAffected by memory/schedulingMedium
Model scale supportedSmall (multi-chip aggregate)Large (80GB VRAM)Large (Pod aggregate)
CompilationFully offlineJust-in-time (JIT)XLA offline
EcosystemSmall (GroqWare)CUDA matureJAX/TF

Use Cases

  • Ultra-low latency LLM inference (GroqCloud API supports Llama 3 70B, Mixtral 8x7B)
  • ✅ Real-time conversational AI (first token latency < 100ms)
  • ✅ Batch LLM inference (high throughput)
  • ✅ Multimodal real-time inference
  • ❌ Large model training (not applicable)
  • ❌ General GPU computing

Groq Commercialization

  • GroqCloud (API service, from 2024)
  • GroqRack (8 GroqChip servers, $1.8M/rack)
  • Customers: Meta (Llama inference), Anthropic, Instagram, Substack

2026-Q1 NVIDIA Acquires Groq (Major Event)

DateEventDetails
2025-12InvestmentNVIDIA invests $250M in Groq
2026-Q1Full acquisitionNVIDIA acquires Groq for ~$20B
2026 H2Product integrationGroq 3 LPU rebranded as NVIDIA Groq 3 LPX, integrated into Vera Rubin platform
2026 H2+Synergistic ecosystemLPX rack as ultra-low latency inference co-processor for Rubin GPU

💡 Strategic significance of the acquisition:

  • NVIDIA, already leading in GPU compute, completes its "ultra-low latency inference" capability via LPU
  • Rubin GPU + LPX co-processing = full-scenario AI compute coverage (training + inference + extreme low-latency inference)
  • Customers: OpenAI, Anthropic, Meta, Mistral, etc.
  • GroqCloud continues operating (OpenAI-compatible API)

Groq 3 LPX Rack (2026 H2)

ItemSpecification
Chip count256 Groq 3 LPUs / rack
On-chip SRAM (rack)128 GB aggregate
SRAM bandwidth (rack)40 PB/s
InterconnectGroqSync + NVLink-Network, 640 TB/s
INT8 compute (rack)~640,000 TOPS
TDP (rack)~80 kW
perf/W35× H100 (official)
TTFT (Time to First Token)< 20ms
TPOT (Time per Output Token)< 5ms

Groq 3 LPX = currently the only rack-scale LPU system designed for Agentic AI. 40 PB/s SRAM bandwidth ≈ 5,000× H100 HBM bandwidth (80GB HBM3 = 3.35 TB/s).

Detailed Product Pages

Groq (Independent)

Groq (Under NVIDIA)

  • Groq LPU v2 - 2024-Q3, 4nm 80GB SRAM 200W GroqCloud service, last generation before acquisition
  • NVIDIA Groq 3 LPX - 2026 H2 256 LPU rack, 128GB aggregate SRAM 40 PB/s, post-acquisition integration into Vera Rubin platform