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
The first Gaudi — later called Gaudi 1 once a second generation existed — was Habana Labs’ AI training and inference chip, folded into Intel when it bought Habana for about $2 billion in December 2019 1. It reached most customers through Amazon Web Services: AWS unveiled instances built on it in December 2020 2 and made them generally available on 26 October 2021 3.
Deep diveThe technical detail
AWS’s own claim at general availability: EC2 DL1 instances deliver “up to 40% better price performance for training deep learning models” than comparable GPU-based EC2 instances 3 4. Intel’s current product listing calls it the “Intel® Gaudi® AI Accelerator” — the same brand later used for Gaudi 2 and Gaudi 3 — even though both the 2020 preview and the 2021 general-availability announcement used the “Habana Gaudi” name, not “Intel” 5 2 3.
2.Launch and history
SimpleStart here
Habana Labs, based in Caesarea, Israel, built this chip before Intel owned it. Intel’s roughly $2 billion purchase of Habana, announced in December 2019, gave Intel its first real AI-training silicon rather than building one from scratch 1.
Deep diveThe technical detail
AWS was Gaudi’s first big customer. Habana’s own chief executive, David Dahan, said “we are proud that AWS has chosen Habana Gaudi processors for its forthcoming EC2 training instances” 2. Intel’s Remi El-Ouazzane, chief strategy officer of its Data Platforms Group, said the deal would “help them reduce the cost of training AI models at scale” 2. At general availability, AWS’s Chetan Kapoor said “Habana’s Gaudi AI accelerators powering our new EC2 DL1 instances ensure the best price/performance for AI training in the cloud today” 6.
3.What’s inside it
SimpleStart here
Eight programmable Tensor Processor Cores (TPCs) plus a matrix-multiplication engine built for deep learning, with networking built onto the chip itself 5.
Deep diveThe technical detail
Intel’s product listing specifies 10 integrated 100 Gigabit Ethernet ports on the chip — the same on-chip-Ethernet approach Intel scaled up on later Gaudi generations: 24 ports on Gaudi 2 7, 24 ports of faster 200GbE on Gaudi 3 (see those pages on this site) 8.
4.Full spec table
SimpleStart here
| Spec | Intel Gaudi (first generation) |
|---|
| Memory | 32 GB onboard HBM2 5 |
|---|
| On-chip SRAM | 24 MB 5 |
|---|
| Cores | 8 Tensor Processor Cores 5 |
|---|
| Networking | 10 integrated 100 Gigabit Ethernet ports 5 |
|---|
| Process node | 16 nm 5 |
|---|
| AI compute | Not published by Intel in the sources we found |
|---|
Deep diveThe technical detail
Memory nearly tripled in one generation: this chip’s 32 GB of HBM2 became 96 GB of HBM2e on Gaudi 2, and TPC core count jumped from 8 to 24 (see the Gaudi 2 page on this site) 5 7.
5.Where it’s made
SimpleStart here
Designed by Habana Labs, based in Caesarea, Israel, both before and after Intel’s acquisition 1. Intel has not published which foundry manufactured this chip in the sources we found.
Deep diveThe technical detail
Unlike Gaudi 3, which Intel’s own white paper names as built by TSMC on a 5 nm process (see that page on this site), no equivalent foundry disclosure was found for this first generation — treated here as genuinely unreported.
6.Which systems use it
SimpleStart here
Sold to most customers as AWS EC2 DL1 instances — up to 8 Gaudi chips per instance, alongside custom Intel Xeon Scalable processors, 768 GiB of instance memory, 400 Gbps networking and 4 TB of local NVMe storage 4.
Deep diveThe technical detail
DL1 launched in two AWS regions: US East (Northern Virginia) and US West (Oregon) 6. AWS also specifies 100 Gbps of chip-to-chip bandwidth between the 8 accelerators inside one instance 4. We found no on-premises server program for this generation comparable to the Dell and Supermicro servers later built for Gaudi 3 (see that page on this site).
7.Official pricing
SimpleStart here
Rented by the hour, not sold as a standalone chip. AWS’s own published on-demand price for a full dl1.24xlarge instance (8 chips) 4:
| Term | Price |
|---|
| On-demand | $13.11/hr |
| 1-year reserved | $7.87/hr |
| 3-year reserved | $5.24/hr |
Deep diveThe technical detail
By our arithmetic, $13.11/hr for 8 chips works out to about $1.64 per chip-hour on demand — in the same range as the $1.25 per chip-hour Denvr Dataworks later charged for Gaudi 2, reserved 4 9. The two prices come from different providers on different terms and cannot be compared directly.
9.What came before, what came next
SimpleStart here
Came before: nothing — this was Habana’s and then Intel’s first AI accelerator chip.
Came after: Gaudi 2, covered on its own page on this site.
Deep diveThe technical detail
The gap from Intel’s Habana acquisition (December 2019) to AWS general availability (October 2021) was almost two years. Gaudi 2 followed about seven months later, in May 2022 7 — a faster turnaround than the gap between acquisition and this chip’s own AWS launch.
10.Hidden in plain sight
SimpleStart here
Intel didn’t put its own name on this chip family for years. The 2020 preview, the 2021 AWS launch and the 2022 Gaudi 2 launch materials all use the “Habana Gaudi” brand, not “Intel Gaudi” — that switch only came with Gaudi 3 in April 2024 2 3 7 10.
Deep diveThe technical detail
Core count is the clearest generational marker Intel has published for this line: 8 Tensor Processor Cores on this first chip 5, 24 on Gaudi 2 7, 64 on Gaudi 3 8 — an eight-fold increase across three generations.