Intel Gaudi is Intel’s line of data-center AI accelerators, and Gaudi 3 is the current model: two compute dies on TSMC’s 5 nm process, 128 GB of HBM2e at 3.7 TB/s, and 24 integrated 200 Gb Ethernet ports for scaling out. Intel’s Gaudi 3 technical paper lists 1,678 TFLOPS of FP8 and BF16 matrix compute. Intel’s newest data-center AI chip is a GPU rather than a Gaudi part: the Intel GPU D-series (Crescent Island), which Intel said was in customer sampling on October 12, 2026.
On this page
Gaudi 3 at a glance
Intel announced Gaudi 3 on April 9, 2024 and launched it alongside Xeon 6 processors on September 24, 2024. The chip is built from two identical compute dies joined on an interposer, organized into four DCOREs, with a total of eight Matrix Multiplication Engines (MMEs) for matrix math and 64 Tensor Processor Cores (TPCs) for vector operations.
Specifications: which Intel document
Intel’s own materials do not all express Gaudi 3’s compute the same way. The figures below come from the Intel Gaudi 3 AI Accelerator Technical Paper, July 2025, V1 Rev. 3 (Table 1), which compares Gaudi 3 with Gaudi 2. The same paper also rounds peak compute to “1.8 PFlops of FP8 and BF16” in its text, and Intel’s product page states the gain over Gaudi 2 as “up to 2x” FP8 and “up to 4x” BF16 compute. The table uses the paper’s exact numbers as Intel states them.
| Spec (Intel technical paper, Table 1) | Gaudi 2 | Gaudi 3 |
|---|---|---|
| MME FP8 | 865 TFLOPS | 1,678 TFLOPS |
| MME BF16 | 432 TFLOPS | 1,678 TFLOPS |
| Vector BF16 (TPC) | 11 TFLOPS | 28.7 TFLOPS |
| HBM capacity | 96 GB | 128 GB (HBM2e) |
| HBM bandwidth | 2.46 TB/s | 3.7 TB/s |
| On-die SRAM | 48 MB | 96 MB |
| Networking (bidirectional) | 600 GB/s | 1,200 GB/s |
| Host interface | PCIe Gen4 x16 (64 GB/s) | PCIe Gen5 x16 (128 GB/s) |
The paper adds that Gaudi 3 is “manufactured in TSMC 5nm process,” has 14 media decoding units, and exposes 24 x 200 Gbps RDMA NIC ports using RoCE. Intel’s Gaudi 3 overview (version 2.6) gives the same 1,678 TFLOPS figure for FP8 and BF16 matrix compute. Specs for Gaudi 3 next to other accelerators are in the AI chips catalog.
What changed from Gaudi 2
Read across the table, the biggest change is in BF16. On Gaudi 2, BF16 matrix throughput (432 TFLOPS) was half the FP8 rate (865 TFLOPS); on Gaudi 3 the two are equal at 1,678 TFLOPS. That is why Intel quotes roughly a 2x gain for FP8 but about 4x for BF16. Memory grew less: capacity rose from 96 GB to 128 GB and bandwidth from 2.46 TB/s to 3.7 TB/s, about 1.5 times, matching the “1.5x increase in memory bandwidth” Intel cited at the April 2024 announcement. On-die SRAM and network bandwidth both doubled, and the host link moved to PCIe Gen5.
Form factors
Intel sells Gaudi 3 in three formats, according to its overview document and technical paper:
| Product | Format | Networking (Intel overview) |
|---|---|---|
| HL-325L | OAM mezzanine card | 24 x 200 GbE RDMA NICs |
| HL-338 | PCIe add-in card, up to 4 per system | 18 x 200 GbE RDMA NICs |
| HLB-325 | Universal baseboard | 192 x 200 GbE RDMA NICs; 9.6 TB/s bidirectional |
The technical paper rates the HL-325L OAM card at up to 900 W and the HL-338 PCIe card at up to 600 W; the PCIe card is air-cooled. Its product brief describes the PCIe card as full-height, 10.5 inches long and double-width, and lists support for FP32, BF16, FP16 and FP8 (both E4M3 and E5M2).
How Intel positions it
Intel’s pitch for Gaudi is standard Ethernet built into the accelerator. Its product page highlights “standard Ethernet” connectivity and “33 percent more I/O connectivity per accelerator” than the H100.
Intel’s performance claims are its own and were made against Nvidia’s H100:
- At the April 2024 announcement, Intel projected “50% faster time-to-train on average” across Llama 2 7B and 13B and GPT-3 175B, about 50% better average inference throughput, and 40% better inference power efficiency averaged across Llama 7B and 70B and Falcon 180B.
- At the September 2024 launch, Intel claimed “up to 20 percent more throughput and 2x price/performance vs H100 for inference of LLaMa 2 70B.”
These are vendor figures, not independent benchmarks, and they compare with Nvidia’s previous generation rather than Blackwell.
Where to get it
Intel’s April 2024 announcement named Dell Technologies, Hewlett Packard Enterprise, Lenovo and Supermicro as OEMs. Intel’s Gaudi 3 overview lists:
- Dell PowerEdge XE9680 servers, shipping from the first quarter of 2025;
- Supermicro X14 servers with Xeon 6, shipping from the second quarter of 2025;
- Intel Tiber AI Cloud;
- IBM Cloud, in the Dallas, Frankfurt and Washington, D.C. regions.
Intel’s Gaudi product page names IBM Cloud and Denvr Dataworks for cloud access, and Dell, HPE and Supermicro as server partners.
Where it stands, and what Intel has announced next
Intel’s most recent data-center AI accelerator is a GPU rather than a Gaudi part. At the 2025 OCP Global Summit, on October 14, 2025, Intel introduced a data-center GPU code-named Crescent Island, based on its Xe3P microarchitecture, described as “power and cost-optimized for air-cooled enterprise servers” and aimed at inference, including “tokens-as-a-service providers.” It said then that customer sampling was expected in the second half of 2026.
At the 2026 OCP Global Summit, on October 12, 2026, Intel gave the chip a product name, the Intel GPU D-series, and said customer sampling was underway. Its announcement lists up to 480 GB of LPDDR5X memory, 1.5 TB/s of memory bandwidth and 1.3 petaflops of FP8 compute in a 350-watt, air-cooled PCIe card, and says one card supports models of up to 300 billion parameters.
The two chips take different routes to memory: Gaudi 3 uses 128 GB of HBM2e at 3.7 TB/s, while the D-series uses LPDDR5X, with more capacity but lower bandwidth, which Intel frames around the economics and power of inference servers. Company background on Intel and its competitors is in the AI companies directory.
