iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
Nvidia’s Groq 3 LPU is designed to complement, not replace, Vera Rubin GPUs: the GPUs handle work that benefits from high throughput and large HBM capacity, while LPUs add a low-latency, on-chip-SRAM path for inference token generation. Nvidia announced Groq 3 LPX was in full production on August 24, 2026, and said Nebius planned to be its first AI cloud adopter. The often-cited $20 billion figure is a reported valuation, not a price disclosed in Groq’s official announcement. Samsung’s 4nm role is also narrower than the title may suggest: Groq announced Samsung would manufacture a next-generation LPU on its SF4X 4nm process in 2023, but the available announcements do not explicitly confirm that Groq 3 uses that node.
What Nvidia’s Groq agreement was—and what the $20 billion refers to
Groq announced on December 24, 2025 that it had entered a non-exclusive licensing agreement with Nvidia covering Groq inference technology. Groq founder Jonathan Ross, president Sunny Madra, and other team members were to join Nvidia to advance and scale the licensed technology. That announcement did not describe an acquisition or disclose a transaction price.
The $20 billion figure reported in secondary coverage should therefore be described as a reported deal valuation, not as an official purchase price. The distinction matters: the public announcement establishes a technology license and personnel changes, but does not provide terms that would establish a $20 billion sale.
Recommended Free Tools
Groq also said it would remain an independent company, with Simon Edwards becoming CEO, and that GroqCloud would continue operating without interruption. Ross said in the December 24 announcement: “The inference opportunity is growing, and we’re excited to partner with Nvidia to bring Groq’s technology to more people around the world.”
#1 Best Overall
- AI Performance: 767 AI TOPS
- OC mode: 2632 MHz (OC mode)/ 2602 MHz (Default mode)
- Powered by the NVIDIA Blackwell architecture and DLSS 4
- Axial-tech fan design features a smaller fan hub that facilitates longer blades and a barrier ring that increases downward air pressure
- A 2.5-slot design maximizes compatibility and cooling efficiency for superior performance in small chassis
How Groq 3 LPX fits alongside Vera Rubin GPUs
Nvidia presents the combination as a heterogeneous system: different chips handle different parts of inference according to their strengths. Vera Rubin NVL72 supplies GPU compute and large HBM capacity; Groq 3 LPX adds an inference accelerator rack built around low-latency access to on-chip SRAM. The point is to divide work, not to substitute an LPU for every GPU.
| Compute component | Memory profile | Role Nvidia describes | Deployment level |
|---|---|---|---|
| Vera Rubin GPU system | Large HBM capacity; a specific capacity is not stated in the cited Nvidia descriptions. | High-throughput work, including attention over the accumulated KV cache and other memory-intensive operations. | Vera Rubin NVL72 GPU rack. |
| Groq 3 LPU | On-chip SRAM; each LPU accelerator has 500 MB SRAM and 150 TB/s SRAM bandwidth, according to Nvidia’s product page. | A low-latency path for token decode and responsive token generation. | Groq 3 LPX inference rack, specified by Nvidia as a 256-chip rack. |
In practical terms, prefill and attention over a growing context can place substantial demands on memory and aggregate compute. Token decode is a different phase: an interactive service may care greatly about how quickly the next token arrives for an individual request. Nvidia’s architecture aims to use the Rubin GPU and LPX where each is better suited, rather than treating one headline throughput measure as the whole story.
What SRAM changes for inference
HBM and SRAM serve different roles. Rubin GPUs provide large HBM capacity, useful when a workload must hold and process substantial model and context data. LPUs place SRAM on the chip, providing a high-bandwidth path for inference operations where low latency and fast token generation are priorities. SRAM’s value here is not that it replaces all other memory; it is that the system can use it for the part of inference Nvidia assigns to the LPU.
Rank #2
- Powered by the NVIDIA Blackwell architecture and DLSS 4
- Powered by GeForce RTX 5070 Ti
- Integrated with 16GB GDDR7 256bit memory interface
- PCIe 5.0
- WINDFORCE cooling system
Nvidia’s technical blog gives LPX rack-level figures of 128 GB aggregate SRAM, 40 PB/s on-chip SRAM bandwidth, 640 TB/s scale-up bandwidth, and 315 PFLOPS. Its product page separately specifies 500 MB SRAM and 150 TB/s SRAM bandwidth for each LPU accelerator. The per-LPU and per-rack figures describe different system levels; they should not be compared as though they were alternate measurements of one chip.
What Nvidia’s performance numbers do—and do not—show
Nvidia’s claims are tied to particular workloads and comparison conditions, not proof that LPX is universally faster or more efficient than another system.
- Gemma 4 benchmark: In its August 24, 2026 release, Nvidia reported 3,400 output tokens per second for Gemma 4 31B at a 100,000-token context in Artificial Analysis benchmarking, and called it the fastest performance then recorded for that model. Nvidia also said the system enabled four-times-faster responsiveness than the nearest alternative platform. These are Nvidia-reported results and comparisons under the stated model and context conditions.
- Throughput per megawatt: Nvidia’s technical materials say Vera Rubin NVL72 paired with LPX can achieve up to 35x higher throughput per megawatt than GB200 NVL72 for models exceeding 2 trillion parameters at long context and high interactivity. “Up to” and the specified operating point are essential qualifications; this is not a general average across AI workloads.
- Power claim: Nvidia says LPX deterministic scheduling can reduce power for a given workload by a potentially low-double-digit percentage compared with a similarly specified nondeterministic system. This is a company claim, not an independently established result.
The available cited materials do not establish an independent study statistic for these claims. Readers evaluating a deployment should treat them as vendor-published results and assess them against their own model, context length, traffic pattern, latency target, and system configuration.
Rank #3
- Powered by the NVIDIA Blackwell architecture and DLSS 4
- Powered by GeForce RTX 5060
- Integrated with 8GB GDDR7 128bit memory interface
- PCIe 5.0
- WINDFORCE cooling system
Production status and the first announced cloud adopter
Nvidia said on August 24, 2026 that Groq 3 LPX was in full production and identified Nebius as the first AI cloud planning to adopt it for its Token Factory inference platform. “Plans to adopt” describes the announced intent; it does not by itself confirm that Nebius had deployed the system or made it generally available to customers by that date.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →The rollout sits within Nvidia’s broader Vera Rubin platform announcements. Nvidia included Groq 3 LPU among the platform’s seven chips in its March 16, 2026 announcement. On May 31, 2026, Nvidia said Vera Rubin was ramping into full production and named system builders and supply-chain partners. These milestones indicate platform and production progress, but they do not establish delivery timing or availability for every cloud or customer.
Does Groq 3 use Samsung’s 4nm process?
The confirmed process-node statement is from Groq’s August 16, 2023 announcement: Samsung Foundry would manufacture Groq’s next-generation LPU on its SF4X 4nm process. Samsung’s GTC 2026 blog later identified Samsung as a manufacturer for the Groq LPU, but did not specify the process node for Groq 3.
Rank #4
- Powered by the NVIDIA Blackwell architecture and DLSS 4. System Requirements: Minimum 850W PSU with 16-pin 12V-2x6 (12VHPWR) connector required. Verify before purchasing.
- Military-grade components deliver rock-solid power and longer lifespan for ultimate durability. Compatibility: 348mm (13.7") length, 3.6 slots, 4.3 lbs. Confirm case clearance and slot spacing. GPU bracket included.
- Protective PCB coating helps protect against short circuits caused by moisture, dust, or debris
- 3.6-slot design with massive fin array optimized for airflow from three Axial-tech fans
- Phase-change GPU thermal pad helps ensure optimal thermal performance and longevity, outlasting traditional thermal paste for graphics cards under heavy loads
Accordingly, Samsung’s manufacturing relationship is supported, and Samsung’s 4nm process was announced for a next-generation Groq LPU. The cited announcements do not explicitly establish that Groq 3 itself is fabricated on SF4X 4nm. Calling Samsung 4nm the confirmed “bedrock” of Groq 3 goes beyond what those statements say.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What this means for AI infrastructure buyers
Groq 3 LPX is rack-scale infrastructure for organizations building or operating inference services, not a consumer accelerator. The architecture is most relevant when a deployment needs interactive token generation and can justify a specialized rack alongside GPU systems. Nvidia’s named system builders and Nebius’s stated plan offer deployment context, but customers should distinguish announced plans and vendor claims from confirmed, generally available services.
- Match the architecture to workload behavior: context length, prompt-processing demand, decode latency, concurrency, and throughput targets all matter.
- Compare systems at the same model, context, traffic pattern, and service-level target; a peak throughput-per-megawatt claim at a demanding operating point is not a like-for-like answer for every application.
- Confirm actual access, deployment timing, and pricing with the provider or system builder. The cited announcements do not establish universal customer availability or a customer price.
Bottom line
Nvidia’s Groq agreement creates a licensed, talent-supported path to combine Groq’s SRAM-oriented inference technology with Vera Rubin GPU systems. The strategic change is the addition of a specialized low-latency decode option, not the replacement of Rubin GPUs. The $20 billion amount remains a reported valuation rather than a disclosed official price, and Samsung’s announced SF4X 4nm manufacturing role should not be presented as a confirmed Groq 3 process node.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
