← AI Hashrate

Gaudi 3 128GB

Intel · 128 GB HBM2e · 3276 GB/s · 600 W · MSRP $12,500

The Gaudi 3 128GB is a Intel accelerator with 128 GB HBM2e and about 3276 GB/s peak memory bandwidth. AI Hashrate estimates local LLM decode (batch 1) at 8K context: fit counts weights + KV cache + 1 GB runtime overhead — not weights alone. At Q4, about 47 curated models fit fully on this card; at FP16, about 38 fit. Top Q4 speeds: Qwen3-0.6B ≈ 2605.9 tok/s (estimated, fits); gemma-3-1b-it ≈ 2084.7 tok/s (estimated, fits); Llama-3.2-1B-Instruct ≈ 1737.3 tok/s (estimated, fits); Qwen3-1.7B ≈ 1042.4 tok/s (estimated, fits); Qwen3-30B-A3B ≈ 694.9 tok/s (estimated, fits). Relative ranking is more reliable than absolute tok/s. See methodology for the bandwidth formula and measured-anchor policy. Methodology.

Table: decode speed estimates (Q4 / FP16). Measured rows override estimates. MSRP is list price, not live retail. Context default 8K.

ModelParams (B)Quanttok/sFits?
Qwen3-0.6B0.8Q42605.9 est.Yes
gemma-3-1b-it1.0Q42084.7 est.Yes
Llama-3.2-1B-Instruct1.2Q41737.3 est.Yes
Qwen3-1.7B2.0Q41042.4 est.Yes
Qwen3-0.6B0.8FP16716.6 est.Yes
Qwen3-30B-A3B30.5Q4694.9 est.Yes
Qwen3-Coder-30B-A3B-Instruct30.5Q4694.9 est.Yes
Qwen3-30B-A3B-Instruct-250730.5Q4694.9 est.Yes
Qwen3-Next-80B-A3B-Instruct81.3Q4694.9 est.Yes
Qwen3-Coder-Next79.7Q4694.9 est.Yes
Llama-3.2-3B-Instruct3.2Q4651.5 est.Yes
MiniMax-M2.7228.7Q4588.4 est.No
gpt-oss-20b21.5Q4579.1 est.Yes
gemma-3-1b-it1.0FP16573.3 est.Yes
Phi-4-mini-instruct3.8Q4548.6 est.Yes
Phi-4-mini-reasoning3.8Q4548.6 est.Yes
gemma-3-4b-it4.3Q4534.5 est.Yes
Qwen3-4B-Instruct-25074.0Q4521.2 est.Yes
Qwen3-4B-Thinking-25074.0Q4521.2 est.Yes
Llama-3.2-1B-Instruct1.2FP16477.7 est.Yes
Qwen3-Coder-Next79.7FP16345.1 est.No
Qwen3-Next-80B-A3B-Instruct81.3FP16331.8 est.No
Mistral-7B-Instruct-v0.37.2Q4289.5 est.Yes
Qwen3-1.7B2.0FP16286.6 est.Yes
Qwen2.5-7B-Instruct7.6Q4274.3 est.Yes
Qwen2.5-Coder-7B-Instruct7.6Q4274.3 est.Yes
DeepSeek-R1-Distill-Qwen-7B7.6Q4274.3 est.Yes
Llama-3.1-8B-Instruct8.0Q4260.6 est.Yes
gpt-oss-120b120.4Q4260.6 est.Yes
Qwen3-8B8.2Q4254.2 est.Yes
DeepSeek-R1-0528-Qwen3-8B8.2Q4254.2 est.Yes
Qwen3-235B-A22B235.1Q4244.4 est.No
Qwen3-235B-A22B-Instruct-2507235.1Q4244.4 est.No
Qwen3-30B-A3B30.5FP16191.1 est.Yes
Qwen3-Coder-30B-A3B-Instruct30.5FP16191.1 est.Yes
Qwen3-30B-A3B-Instruct-250730.5FP16191.1 est.Yes
Llama-3.2-3B-Instruct3.2FP16179.2 est.Yes
gemma-3-12b-it12.2Q4173.7 est.Yes
GLM-4.7-Flash31.2Q4173.7 est.Yes
GLM-4.5-Air110.5Q4173.7 est.Yes
Mixtral-8x7B-Instruct-v0.146.7Q4161.6 est.Yes
DeepSeek-V4-Flash158.1Q4160.4 est.Yes
gpt-oss-20b21.5FP16159.2 est.Yes
Phi-4-mini-instruct3.8FP16150.9 est.Yes
Phi-4-mini-reasoning3.8FP16150.9 est.Yes
gemma-3-4b-it4.3FP16147.0 est.Yes
Qwen3-4B-Instruct-25074.0FP16143.3 est.Yes
Qwen3-4B-Thinking-25074.0FP16143.3 est.Yes
Phi-414.7Q4141.8 est.Yes
Phi-4-reasoning14.7Q4141.8 est.Yes
Qwen3-14B14.8Q4140.9 est.Yes
Qwen2.5-14B-Instruct14.8Q4140.9 est.Yes
Qwen2.5-Coder-14B-Instruct14.8Q4140.9 est.Yes
DeepSeek-R1-Distill-Qwen-14B14.8Q4140.9 est.Yes
Devstral-Small-250723.6Q488.3 est.Yes
Mistral-Small-3.2-24B-Instruct-250624.0Q486.9 est.Yes
Magistral-Small-250924.0Q486.9 est.Yes
Mistral-7B-Instruct-v0.37.2FP1679.6 est.Yes
gemma-3-27b-it27.4Q476.6 est.Yes
Qwen2.5-7B-Instruct7.6FP1675.4 est.Yes
Qwen2.5-Coder-7B-Instruct7.6FP1675.4 est.Yes
DeepSeek-R1-Distill-Qwen-7B7.6FP1675.4 est.Yes
GLM-4.5358.3Q472.9 est.No
Llama-3.1-8B-Instruct8.0FP1671.7 est.Yes
Qwen3-8B8.2FP1669.9 est.Yes
DeepSeek-R1-0528-Qwen3-8B8.2FP1669.9 est.Yes
Qwen3-32B32.8Q463.6 est.Yes
Qwen2.5-32B-Instruct32.8Q463.6 est.Yes
Qwen2.5-Coder-32B-Instruct32.8Q463.6 est.Yes
DeepSeek-R1-Distill-Qwen-32B32.8Q463.6 est.Yes
gpt-oss-120b120.4FP1656.6 est.No
gemma-3-12b-it12.2FP1647.8 est.Yes
GLM-4.7-Flash31.2FP1647.8 est.Yes
Mixtral-8x7B-Instruct-v0.146.7FP1644.4 est.Yes
GLM-4.5-Air110.5FP1644.4 est.No
Phi-414.7FP1639.0 est.Yes
Phi-4-reasoning14.7FP1639.0 est.Yes
Qwen3-14B14.8FP1638.7 est.Yes
Qwen2.5-14B-Instruct14.8FP1638.7 est.Yes
Qwen2.5-Coder-14B-Instruct14.8FP1638.7 est.Yes
DeepSeek-R1-Distill-Qwen-14B14.8FP1638.7 est.Yes
Qwen3-Coder-480B-A35B-Instruct480.2Q437.7 est.No
MiniMax-M1456.1Q431.2 est.No
Llama-3.1-70B-Instruct70.6Q429.8 est.Yes
Llama-3.3-70B-Instruct70.6Q429.8 est.Yes
DeepSeek-R1-Distill-Llama-70B70.6Q429.5 est.Yes
Qwen2.5-72B-Instruct72.7Q428.7 est.Yes
Devstral-Small-250723.6FP1624.3 est.Yes
Mistral-Small-3.2-24B-Instruct-250624.0FP1623.9 est.Yes
Magistral-Small-250924.0FP1623.9 est.Yes
gemma-3-27b-it27.4FP1621.1 est.Yes
DeepSeek-V4-Flash158.1FP1620.2 est.No
DeepSeek-R1684.5Q417.8 est.No
DeepSeek-V3.2685.4Q417.8 est.No
Qwen3-32B32.8FP1617.5 est.Yes
Qwen2.5-32B-Instruct32.8FP1617.5 est.Yes
Qwen2.5-Coder-32B-Instruct32.8FP1617.5 est.Yes
DeepSeek-R1-Distill-Qwen-32B32.8FP1617.5 est.Yes
Llama-3.1-70B-Instruct70.6FP1615.7 est.No
Llama-3.3-70B-Instruct70.6FP1615.7 est.No
DeepSeek-R1-Distill-Llama-70B70.6FP1615.6 est.No
Qwen2.5-72B-Instruct72.7FP1614.3 est.No
GLM-5.1753.9Q413.6 est.No
GLM-5.2753.3Q413.6 est.No
MiniMax-M2.7228.7FP1612.7 est.No
DeepSeek-V4-Pro861.6Q48.5 est.No
Qwen3-235B-A22B235.1FP165.4 est.No
Qwen3-235B-A22B-Instruct-2507235.1FP165.4 est.No
Llama-3.1-405B-Instruct405.9Q42.6 est.No