Intel's third-generation AI accelerator with 128GB of HBM2e and on-chip 200GbE networking for Ethernet scale-out.
By provider shows one card per company, placed where its best offer ranked. By configuration lists every offer.
What you can rent comes first: in stock, then waitlist, then not reported, then out of stock. Priced offers rank ahead of quote-only, and on-demand ahead of other billing types.
Within a group, five factors set the order:
We match the provider, its country, GPU model, form factor, billing type, availability and instance name. Matching is partial and case-insensitive. Everyday words work too, so "interruptible" finds spot and "sold out" finds out of stock.
No providers match .
Heads up: A provider's own page may quote a different figure, for example on tax, region or a promotion. A monthly-only plan shows a derived hourly rate. Verify before provisioning. More on how we price.
One Gaudi 3 has 128 GB of VRAM. In practice, that's enough memory for roughly 199B parameters at 4-bit or 53B at 16-bit, assuming a 32K context. Below are some open-weight LLMs, with the estimated memory and GPUs each one needs.
| Model | Memory (INT4 / FP4) | Gaudi 3s needed | Cost /hr | Cost /mo |
|---|---|---|---|---|
Qwen3.8-27B
27B
|
18 GB
|
8
|
–
|
–
|
Gemma 4 31B
30.7B
|
22 GB
|
8
|
–
|
–
|
GPT-OSS-120B
117B · 5.1B active
|
67 GB
|
8
|
–
|
–
|
GLM-5.3-Flash
320B · 18B active
|
178 GB
|
8
|
–
|
–
|
MiniMax-M3
428B · 23B active
|
239 GB
|
8
|
–
|
–
|
DeepSeek V4.1 Flash
552B · 16B active
|
306 GB
|
8
|
–
|
–
|
Kimi K3
2.8T · 104B active
|
1,544 GB
|
14
|
–
|
–
|
Estimates based on the median on-demand rate. Memory is weights plus FP8 KV cache at 32K context per request (FP16/BF16 in the 16-bit column). GPU counts assume 90% of advertised VRAM is usable. No guarantee of runtime support, usable performance, or that a matching quantized build exists. Pricing methodology.
Intel Gaudi 3 · Per GPU
| Compute · dense | |
|---|---|
| FP8 | 1,678 TFLOPS |
| FP16 / BF16 | 1,678 TFLOPS |
| Precision support | FP8FP16BF16TF32FP32 |
| Memory | |
|---|---|
| Capacity | 128 GB HBM2e |
| Bandwidth | 3,700 GB/s |
| Silicon | |
|---|---|
| Architecture | Gaudi (3rd Gen) |
| Process | TSMC 5nm |
| Fabric and host | |
|---|---|
| GPU interconnect | RoCE 1,200 GB/s |
| Host interface | PCIe 5.0 x16 |
| Power | |
|---|---|
| Board power | 900 W |
Source: official Intel Gaudi 3 datasheet.
As of October 13, 2026, we track 1 config from 1 provider, all of them custom contracts, priced per deal and usually not published.
GetDeploying currently tracks Gaudi 3 configs from 1 provider. See the full price comparison above for every provider and config.
One Gaudi 3 runs models up to roughly 199B parameters at 4-bit quantization or 53B at 16-bit, assuming a 32K context. Larger models run across multiple GPUs: the model table above shows the estimated memory and GPU count for popular open-weight LLMs.
Comes with 128GB HBM2e per card and 24x 200GbE RoCEv2 ports on-chip, so an 8-card node reaches 1TB of accelerator memory and scales out over standard Ethernet without separate NICs. BF16 and FP8 run at the same peak rate, and a 70B-parameter model in FP8 fits on a single card.
The Gaudi software stack supports fewer frameworks, kernels and models than CUDA, so workloads relying on custom CUDA kernels or niche libraries need porting. For broad ecosystem compatibility, H100 or H200 are easier to adopt; Gaudi 3 remains a fit for PyTorch and Hugging Face workloads that run on Gaudi as is.
Intel Gaudi 2
96 GB HBM2e · 34% less bandwidth than the Gaudi 3
Last updated
