AI Chips · Product page

Intel Gaudi 2: The Complete Guide

The chip that tripled Gaudi’s memory and quadrupled its rental customers’ choices — still rented today, in India and through independent clouds. Simple to expert, every number links to its source.

Stuck at any point? Ask an AI about this page →
Ask an AI about this chip:CAGPX
Intel Gaudi 2 10 sections · 2 levels 11 linked sources Checked September 2026
How to read this page. Each section starts Simple, then goes to a Deep dive: stop wherever you have what you need. The small numbers are sources: click one to open the original document. Where the maker has not published a figure -- a price, a die size, a factory address -- this page says so rather than estimate it.

1.At a glance

SimpleStart here

Gaudi 2, launched at Intel Vision on 10 May 2022, tripled the memory of the first Gaudi chip — from 32 GB to 96 GB — and Intel’s own launch materials claimed “2x throughput performance over Nvidia A100 for training” 1.

Deep diveThe technical detail

Intel backed that claim with public numbers fast: MLPerf training results published 29 June 2022, just over a month after launch, showed Gaudi 2 with a “36% reduction in time-to-train as compared to Nvidia’s [time-to-train] for A100-80GB” on ResNet-50 2. Intel’s Sandra Rivera, then an executive vice president, called the speed of that submission notable in itself — “just one month after launch” 2.

2.Launch and history

SimpleStart here

Gaudi 2 is the generation that made Habana’s chips look like a real alternative to NVIDIA’s GPUs rather than a curiosity: Intel’s launch materials led with a head-to-head comparison against NVIDIA’s A100, not just a price argument 1.

Deep diveThe technical detail

Habana Labs, bought by Intel for about $2 billion in December 2019, designed Gaudi 2 the same way it designed the original chip — from its Caesarea, Israel team 3. Stability AI later used Gaudi 2 as the anchor for an AI supercomputer sized at about 4,000 chips, announced at Intel Innovation in September 2023 4.

3.What’s inside it

SimpleStart here

Twenty-four Tensor Processor Cores, up from eight on the first Gaudi chip, plus double the on-board SRAM (48 MB) and 24 integrated 100 Gigabit RoCE ports for chip-to-chip networking 1.

Deep diveThe technical detail

Intel’s fact sheet gives a server-level networking figure too: “2.4 terabits of networking throughput in the 8-card Gaudi server” 1. Core count would grow again to 64 on Gaudi 3 5 (see that page on this site), but the jump from 8 cores on the first Gaudi chip 6 to 24 here was the larger relative increase.

4.Full spec table

SimpleStart here
SpecIntel Gaudi 2
Memory96 GB HBM2e 1
Memory bandwidth2.45 TB/s 1
On-chip SRAM48 MB 1
Cores24 Tensor Processor Cores 1
Networking24 integrated 100 Gigabit RoCE ports 1
Process node7 nm 1
AI computeNot stated in Intel’s launch fact sheet
Deep diveThe technical detail

Intel’s fact sheet gives relative performance claims (below) rather than an absolute TFLOPS figure — the same gap this site found in Gaudi 2’s entry on the Intel brand page’s own specs table 1.

5.Where it’s made

SimpleStart here

Designed by Habana Labs in Caesarea, Israel 3. Intel’s fact sheet names the process node (7 nm) but not the specific foundry 1.

Deep diveThe technical detail

Gaudi 3, the next generation, is explicitly named by Intel’s own white paper as made by TSMC (see that page on this site) — no equivalent foundry name was found published for Gaudi 2, treated here as genuinely unreported.

6.Which systems use it

SimpleStart here

Sold mainly through system builders and cloud partners rather than one flagship cloud (compare the original Gaudi’s near-exclusive AWS launch, on that page on this site). Rented through Denvr Dataworks and listed on India’s government-run IndiaAI Mission compute pool 7 8.

Deep diveThe technical detail

Stability AI’s roughly 4,000-chip Gaudi 2 supercomputer, announced September 2023, is the largest named single Gaudi 2 deployment this site found sourced 4. The IndiaAI Mission’s shared compute pool, which includes Gaudi 2, listed 34,333 GPUs in total as of May 2025 9.

7.Official pricing

SimpleStart here

Intel published a list price for a kit of eight Gaudi 2 chips on a baseboard: $65,000, offered to system makers as of June 2024 — about $8,125 per chip by our arithmetic 10.

Deep diveThe technical detail

Rental prices vary by provider. Denvr Dataworks charges $1.25 per chip-hour, reserved, with 8 chips per machine 7. On India’s IndiaAI Mission compute portal, one Gaudi 2 card (96 GB) lists at ₹57.60 an hour and a full 8-card server at ₹460.80, before GST 8.

8.Real-world performance

SimpleStart here

Intel’s own launch claim: Gaudi 2 “demonstrates 2x throughput performance over Nvidia A100 for training” 1.

Deep diveThe technical detail

Intel followed up with an independently-checkable benchmark, not just a marketing multiplier. Its MLPerf Training 2.0 submission, published 29 June 2022, reported a 36% cut in time-to-train versus an A100-80GB on ResNet-50, and a 45% cut versus an A100-40GB configuration on both ResNet-50 and BERT 2. Generation over generation, Intel reported Gaudi 2 reaching “3x speed-up in training throughput for ResNet-50 and 4.7x for BERT” compared with the first Gaudi chip 2. MLPerf is a third-party industry benchmark, though the specific figures above are Intel’s own summary of its submitted results, not an outside reviewer’s independent test.

9.What came before, what came next

SimpleStart here
Came before: the first Gaudi chip, covered on its own page on this site. Came after: Gaudi 3, covered on its own page on this site.
Deep diveThe technical detail

Gaudi 2 shipped about seven months after the first Gaudi chip reached AWS general availability — a faster generational turnaround than the roughly two years between Intel’s Habana acquisition and that first chip’s own AWS launch (see that page on this site) 1. Gaudi 3 followed almost two years later, in April 2024 11.

10.Hidden in plain sight

SimpleStart here

Intel still hadn’t put its own name on the chip when this one launched: the fact sheet is titled “Habana Gaudi2 Launch Fact Sheet,” not “Intel Gaudi 2” — that rebrand only arrived with the next generation, Gaudi 3, in April 2024 1 11.

Deep diveThe technical detail

The 2x-over-A100 claim compares Gaudi 2 to NVIDIA’s A100, which by Gaudi 2’s 2022 launch was already two years old (2020) and would itself be followed by NVIDIA’s H100 later that same year — the comparison chip, not the newest one available at launch 1.

11.Sources

11 sources, checked September 2026. Where NVIDIA or another maker has not published a figure, this page says so rather than estimate it.

  1. Intel: Habana Gaudi2 launch fact sheet (10 May 2022)Official
  2. HPCwire (press): Intel Announces MLPerf Results for Habana Gaudi2 (29 Jun 2022)Press
  3. Intel Investor Relations: Intel acquires artificial intelligence chipmaker Habana Labs (16 Dec 2019)Official
  4. Intel Investor Relations: Intel Innovation 2023 (19 Sep 2023)Official
  5. Intel: Intel Gaudi 3 AI Accelerator white paperOfficial
  6. Intel: Intel® Gaudi® AI Accelerator (first generation) product pageOfficial
  7. Denvr Dataworks: PricingOfficial
  8. IndiaAI Mission (Government of India): IndiaAI compute price calculatorGovernment
  9. Press Information Bureau, Government of India: IndiaAI compute capacity of 34,333 GPUs (30 May 2025)Government
  10. Intel Investor Relations: Intel accelerates AI everywhere at Computex 2024 (4 Jun 2024)Official
  11. Intel Investor Relations: Intel unleashes enterprise AI with Gaudi 3 (9 Apr 2024)Official

Ask an AI about this page

Opens your assistant with this page as the source, and a question rather than a summary. It will ask what you are building before it answers.

ChatGPTClaudeGeminiPerplexityGrok

Nothing is sent from here. The link carries only this page’s title and address.