Groq LPU (GroqChip): The Complete Guide

The LPU keeps an entire model in fast on-chip memory (SRAM) rather than fetching it from separate memory chips for every word, which Groq says gives it “memory bandwidth upwards of 80 terabytes/second, while GPU off-chip HBM clocks in at about eight terabytes/second” 1. As of this check, the chip carrying the Groq brand today is still this original 2020-era design — not a newer generation (see Lineage below).

Groq built the LPU around a “Tensor Streaming” design, where the compiler — not the chip itself — decides in advance exactly when every calculation happens 3. This same generation underpins the specific systems Groq has documented — the GroqCard accelerator card and the GroqNode/GroqRack configurations used in the Argonne Leadership Computing Facility’s testbed 4 5 — and, as of the check date, Groq has not announced a newer chip generation for its own branded hardware (see Lineage).

Academically, this is a Tensor Streaming Processor: one very large core built from small, deterministic functional units connected by a software-scheduled on-chip network, rather than the many independent cores and hardware schedulers found in a GPU 3.

The GroqCard has up to nine chip-to-chip connectors, letting many cards link directly without a separate networking chip 4. Groq’s successor chip, sold today as the NVIDIA Groq 3 LPX (see Lineage), reports 500 MB of SRAM and 150 TB/s of bandwidth — roughly double this chip’s memory and nearly double its bandwidth 6.

An analyst estimate puts the chip’s wafer cost at “likely less than $6,000,” well below leading-edge chips of the same era 8. Groq’s 2023 announcement of a planned next-generation chip likewise described its “already completely North American-based operations for engineering and manufacturing” 9 — though that next-generation chip did not ship as a Groq product (see Lineage).

An independent analyst estimated that serving one open-weight Mixtral model on GroqCloud took 576 of these chips — 8 racks of 9 servers with 8 chips each — though Groq has not published an official chip count for any specific model 8.

This matches the pattern across Groq’s public materials generally: it publishes cloud API prices but has never listed a hardware price for any of its own systems 4.

Groq’s own marketing claims “up to a 10X speed advantage” over GPUs and similar energy savings for the LPU generally — a Groq claim, not an independently verified, chip-specific benchmark 1.

Groq’s 2023 release said the Samsung deal would let it build “systems ranging from 85,000 to more than 600,000 chips, without any external switches” — a claim about how many chips could be networked into one system, not about factory output 9. No Groq-branded chip from that plan ever shipped under Groq’s own name. Instead, on 24 December 2025 Groq licensed its inference technology to NVIDIA in a non-exclusive agreement, and founder Jonathan Ross and other Groq leaders joined NVIDIA 12. NVIDIA’s resulting rack — which states “Groq and LPU are used under license from Groq, Inc.” — reached full production in August 2026, built by Samsung on what press reports describe as a 4 nm process 13 14. Because it is sold, branded and manufactured under NVIDIA’s name rather than Groq’s, this site treats it as an NVIDIA product, not a second Groq-branded chip, and this Groq product-page series ends with the original LPU. Groq’s own home page describes the arrangement plainly: “We pioneered the LPU. Now, with LPX, it works alongside NVIDIA’s next-generation GPUs” 15.

An analyst’s estimated wafer cost of “likely less than $6,000” suggests each chip itself is inexpensive to produce; the real cost of a Groq deployment comes from needing hundreds of them working together to match one GPU’s memory capacity — for example, an estimated 576 chips to serve a single Mixtral model 8.