IN Intel SME Discrete GPU

Intel Gaudi 3

Runs downloadable LLMs up to ~199.5B at Q4, ~479-594.6 tok/s on an 8B model.

Intel Gaudi 3
128GB
VRAM
3700
GB/s bandwidth
199.5B
Max model (Q4)
900W
TDP · PSU 1300W

What you can run

At 8k context, FP16 KV. Full-speed tok/s holds up to "max ctx"; beyond it the context spills to system RAM (CPU/PCIe offload) and drops to the degraded figure.

Model sizeFP16Q8Q4
3B 357.7-444 t/s
≤1682k ctx · then ~3.9-4.8 (slow)
715.3-888 t/s
≤1728k ctx · then ~7.7-9.6 (slow)
1277.4-1585.7 t/s
≤1748k ctx · then ~13.8-17.1 (slow)
7-8B 134.1-166.5 t/s
≤765k ctx · then ~1.5-1.8 (slow)
268.3-333 t/s
≤826k ctx · then ~2.9-3.6 (slow)
479-594.6 t/s
≤853k ctx · then ~5.2-6.4 (slow)
13B 82.5-102.5 t/s
≤551k ctx · then ~0.9-1.1 (slow)
165.1-204.9 t/s
≤630k ctx · then ~1.8-2.2 (slow)
294.8-365.9 t/s
≤665k ctx · then ~3.2-4 (slow)
32B 33.5-41.6 t/s
≤199k ctx · then ~0.4-0.5 (slow)
67.1-83.3 t/s
≤321k ctx · then ~0.7-0.9 (slow)
119.8-148.7 t/s
≤375k ctx · then ~1.3-1.6 (slow)
70B won't run 30.7-38.1 t/s
≤141k ctx · then ~0.3-0.4 (slow)
54.7-68 t/s
≤235k ctx · then ~0.6-0.7 (slow)
120B won't run won't run 31.9-39.6 t/s
≤150k ctx · then ~0.3-0.4 (slow)
400B won't run won't run won't run

Estimate: tok/s ≈ bandwidth ÷ (params × bytes/param) × efficiency (Intel). Cyan = fits ~117.8 GB usable. Beyond "ctx", KV spills to system RAM (CPU/PCIe offload) (rough degraded speed shown).

What else you need for a full build

Minimum supporting components for a desktop built around the Intel Gaudi 3. Guidance — exact needs depend on your case, other parts and how much you offload to system RAM.

Power supply
≥ 1300W · 80+ Gold
Headroom over the 900W card plus the rest of the build.
System RAM
≥ 128 GB · 256 GB recommended
At least the 128 GB of VRAM; 2× lets you CPU-offload models larger than VRAM.
Motherboard
1× free PCIe x16 slot (4.0/5.0)
A full-length x16 electrical slot; PCIe 4.0 is ample, 5.0 future-proofs.
CPU
Any modern multi-core
The GPU runs inference; more cores and memory channels only matter when offloading to RAM.
Storage
≥ 1 TB NVMe SSD
Weights are large — a 70B model at Q4 is ~40 GB; budget for a few.

Run popular models on the Intel Gaudi 3

Videos about the Intel Gaudi 3

Intel's Big AI Bet: Gaudi 3 ft. Supermicro -- System Tour, Tested Pytorch & More

Intel Gaudi 3 AI Accelerator with 24x 200GbE #ai

More hardware like the Intel Gaudi 3

About the Intel Gaudi 3

Intel Gaudi 3 is a discrete GPU from Intel, 2024. For local genAI the numbers that matter are its 128 GB (how big a model + context fits) and 3700 GB/s bandwidth (generation speed). Plan for a ~1300W power supply.

Frequently asked questions

What size LLM can the Intel Gaudi 3 run?

With ~117.8 GB usable it runs up to ~199.5B at Q4 (~56.5B at FP16) at 8k context. Bigger spills into system RAM (CPU/PCIe offload) and slows sharply.

How many tokens/sec does the Intel Gaudi 3 generate?

Generation is bandwidth-bound. At 3700 GB/s an 8B model at Q4 runs ~479-594.6 tok/s (estimate).

Why does long context slow it down?

The KV cache for the context must also fit fast memory; once weights+KV exceed 117.8 GB it spills to system RAM (CPU/PCIe offload) and throughput collapses. Each row shows the max usable context.

What else do I need to run the Intel Gaudi 3?

Besides the card: a ≥1300W 80+ Gold power supply, at least 128GB of system RAM (2× the 128GB VRAM is ideal for CPU offload), a free PCIe x16 slot, any modern multi-core CPU, and a ≥1TB NVMe SSD for model weights.