1 live offerings · 1 tracked regions
Publicly listed 1
Loading
Loading workspaces…
engineering · GPU0101
Specifications with their source, supported precisions, interconnect, benchmark coverage and which published models fit Intel Gaudi 2 96GB.
| Specification | Value | Source | Kovara registry |
|---|---|---|---|
| Memory capacity | 96GB | Kovara registry · verified 2026-09-25 | used |
| Memory type | HBM2E | Kovara registry · verified 2026-09-25 | used |
| Memory bandwidth | 2.45TB/s | Kovara registry · verified 2026-09-25 | used |
| Board power (TDP) | 600W | Kovara registry · verified 2026-09-25 | used |
| Form factor | Mezzanine accelerator | Kovara registry · verified 2026-09-25 | used |
| NVLink | No | Kovara registry · verified 2026-09-25 | used |
| MIG partitioning | No | Kovara registry · verified 2026-09-25 | used |
No public benchmark runs yet
Nobody has published a hardware run on this accelerator. Measured numbers appear here once a run of a compute repository is published.
AcceleratorsGPU0101
Intel
Core hardware and lifecycle, drawn from Kovara master data.
What the recorded specifications mean in practice, and what this accelerator is used for.
Specs, explained
HBM (High Bandwidth Memory) is stacked in layers and sits right next to the processor on the same package. That short, wide path is what gives accelerators with HBM their memory bandwidth, so they are usually chosen for work limited by how fast data reaches the cores, such as large-model inference and HPC simulation.
VRAM is the working space on the card. Model weights, activations and datasets have to fit in it, or be split across several GPUs. Rule of thumb for the weights alone at 16-bit precision (about 2 GB per billion parameters): up to roughly 48 billion parameters, before activations and cache. This is arithmetic on the recorded VRAM, not a benchmark.
Gaudi is Intel's AI accelerator line, built for deep-learning training and inference with networking built into the chip. (General context about the architecture, not a statement about this specific card.)
Launched in 2022. Kovara's records list it as a previous-generation part.
Accelerator modules are mounted on a baseboard rather than slotted in as cards, the layout used for dense multi-accelerator servers.
TDP is the rated thermal design power: the planning figure for power delivery and cooling in a server.
These are the uses Kovara has recorded for this part. The notes explain the workload in general terms; they are not performance claims.
Figures are Kovara master data, last verified 25 Sept 2026. Explanations describe the technology in general; they are not benchmarks and say nothing about price or availability.
Explore this accelerator on ComputePublic Kovara observations; not a claim of complete or universal market coverage. Incompatible raw price units are excluded.
Observed GPU-hour range
USD / GPU-hour
Spread $0.15 USD / GPU-hour
Providers represented by current Kovara-tracked offerings.
1 live offerings · 1 tracked regions
Publicly listed 1
Regions represented by current offerings; this does not describe complete global availability.
1 live offerings · 1 providers
Current configurations, recorded availability, confirmed capacity and the evidence behind each price.
| Provider | Region | Configuration | Availability | Price | Commercial model | Trust / evidence | Action |
|---|---|---|---|---|---|---|---|
| OFR-0684Denvr Data | Unspecified, Multiple / Unspecified | Intel Gaudi 2 96GB8 GPU configuration · 8 GPUs per node | Publicly listedConfirmed capacity: Not confirmed | $1.10 USD / GPU-hourOriginal: USD 1.10 GPU/hour | On-DemandGPU platform · Per-minute billing | Confidence AChecked todayConfirmed 23 Sept 2026Source: SRC000038 | Inspect |
Recorded hardware attributes. Not tracked is distinct from No.
Public evidence and confidence signals supporting Kovara-tracked coverage.