Skip to main content

NVIDIA Groq 3 LPX (LPU Rack-Scale)

Overview

NVIDIA Groq 3 LPX is a rack-scale LPU (Language Processing Unit) system launching 2026 H2, the flagship LPU product following NVIDIA's acquisition of Groq. Each rack contains 256 Groq 3 LPUs, delivering 40 PB/s on-chip SRAM aggregate bandwidth, 640 TB/s interconnect bandwidth, and a 35× perf/W advantage (vs. H100 inference).

The Groq 3 LPX serves as an inference acceleration co-processor for the NVIDIA Vera Rubin platform — when users require ultra-low-latency trillion-parameter model inference (such as agentic AI, real-time code generation), they can deploy an LPX rack as a co-processor alongside Rubin GPUs.

Core Specifications

ItemSpec
ArchitectureGroq 3 LPU (Tensor Streaming Processor v3)
Form FactorRack-scale (256 per rack)
On-Chip SRAM (per LPU)512 MB
On-Chip SRAM (Rack)128 GB aggregate
On-Chip SRAM Bandwidth (Rack)40 PB/s
Interconnect (Intra-Rack)GroqSync + NVLink-Network (640 TB/s)
INT8 Compute (per LPU)2,500 TOPS (estimated)
FP8 Compute (Rack)~640 PFLOPS (estimated)
BF16 Compute (Rack)~320 PFLOPS (estimated)
TDP (Rack)~80 kW
perf/W (Inference)35× H100 (official)
Launch2026 H2 (alongside Rubin R200)

40 PB/s on-chip SRAM bandwidth ≈ 5,000× H100 HBM bandwidth (H100 80GB HBM3 = 3.35 TB/s). This is the core secret behind the Groq LPU's ultra-low latency.

NVIDIA Groq Acquisition

EventDateDetail
Initial partnership2025-12NVIDIA invests $250M in Groq
Full acquisition2026-Q1NVIDIA acquires Groq outright (~$20 billion)
Product integration2026 H2Groq 3 LPU rebranded as NVIDIA Groq 3 LPX
Integration into Vera Rubin platform2026 H2LPX rack as Rubin GPU co-processor

Acquisition significance: On top of NVIDIA's GPU compute leadership, the LPU fills the "ultra-low-latency inference" gap. Rubin GPU + LPX co-processing = full-spectrum AI compute coverage (training + inference + extreme low-latency inference).

Groq 3 LPU Single Chip vs Rack

ItemSingle LPUGroq 3 LPX Rack
Chip count1256
On-Chip SRAM512 MB128 GB
SRAM Bandwidth160 TB/s40 PB/s
InterconnectGroqSync 1 TB/s640 TB/s
INT8 Compute2,500 TOPS~640,000 TOPS
TDP~300 W~80 kW
UseSingle model inferenceMulti-model / agentic

128 GB SRAM aggregate ≈ 32× H100 80GB memory aggregate, but with 100× lower latency (nanosecond vs microsecond HBM).

Groq 3 LPX vs NVIDIA H100 / Rubin R200 (Inference Comparison)

MetricH100 (SXM)Rubin R200Groq 3 LPX
Memory / SRAM Aggregate80 GB HBM288 GB HBM4128 GB SRAM
Bandwidth3.35 TB/s22 TB/s40 PB/s
LatencyMicrosecondsMicrosecondsNanosecond (1000× better)
FP8 Compute (Rack / Card)~3,958 TFLOPS50 PFLOPS~640 PFLOPS
TTFT (Time to First Token)~200ms~100ms< 20ms
TPOT (Time per Output Token)~30ms~15ms< 5ms
perf/W1× (baseline)~3×35×
UseTraining + inferenceTraining + inferenceUltra-low-latency inference

LPX's core advantage is latency (not absolute compute). For agentic AI (1000+ calls/sec), TTFT < 20ms is critical.

Use Cases

ScenarioRecommended Configuration
Agentic AI inferenceLPX rack (1000+ calls/sec)
Real-time Code Generation (Cursor / Copilot)LPX rack
Trillion-parameter model inferenceLPX + Rubin GPU co-processing
Multi-model concurrency (RAG, function calling)LPX rack
Low-latency search (Perplexity, You.com)LPX rack

Software Ecosystem

  • GroqWare (rebranded post-acquisition as NVIDIA Groq SDK)
  • NVIDIA NIM + LPX backend
  • vLLM 0.8+ Groq backend (estimated)
  • OpenAI API compatibility layer (compatible with existing LLM applications)
  • LangChain / LlamaIndex integration

Pricing (Estimated)

ItemPrice
LPX Rack (256 units)$8M-10M / rack (estimated)
Monthly operating cost~$300K-500K (including power, cooling)
Per-dollar inference cost50-70% lower than H100 (based on 35× perf/W)

LPX is not a GPU replacement, but a GPU complement: In latency-sensitive scenarios like agentic AI, LPX is the only choice; for cost-sensitive large-scale inference, the Rubin R200 is more economical.

Vendor Information

ItemDetail
Original VendorGroq Inc. (acquired by NVIDIA 2026-Q1)
Current VendorNVIDIA Corporation (subsidiary)
First Release2026 H2 (Vera Rubin platform generation)
Product Pagehttps://www.nvidia.com/en-us/data-center/lpx/
API ServiceNVIDIA GroqCloud (merged from GroqCloud)
Acquisition Amount~$20 billion