Can you run HiTZ/EriBERTa-base on H100 80GB?

HiTZ/EriBERTa-base needs ~2.96 GB, which is within the 80 GB on H100 80GB by memory capacity (~77.04 GB headroom). Computed from stated size and quant, not a runtime guarantee.

HiTZ/EriBERTa-base needs ~2.96 GB, which is within the 80 GB on H100 80GB by memory capacity (~77.04 GB headroom). Computed from stated size and quant, not a runtime guarantee.

Memory-capacity estimate across common GPUs

GPUVerdictVRAMNeeds
8GB laptop (no dedicated GPU)Fits (by memory)8 GB~2.96 GB
16GB laptop (no dedicated GPU)Fits (by memory)16 GB~2.96 GB
NVIDIA T4 16GB (free Colab)Fits (by memory)16 GB~2.96 GB
NVIDIA L4 24GBFits (by memory)24 GB~2.96 GB
RTX 3060 12GBFits (by memory)12 GB~2.96 GB
RTX 4080 16GBFits (by memory)16 GB~2.96 GB
RTX 3090 24GBFits (by memory)24 GB~2.96 GB
RTX 4090 24GBFits (by memory)24 GB~2.96 GB
A100 40GBFits (by memory)40 GB~2.96 GB
A100 80GBFits (by memory)80 GB~2.96 GB
H100 80GBFits (by memory)80 GB~2.96 GB
Apple M-series 16GB (unified)Fits (by memory)16 GB~2.96 GB
Apple M-series 32GB (unified)Fits (by memory)32 GB~2.96 GB
Apple M-series 64GB (unified)Fits (by memory)64 GB~2.96 GB
Apple M-series 128GB (unified)Fits (by memory)128 GB~2.96 GB
CPU / 32GB system RAMFits (by memory)32 GB~2.96 GB
CPU / 64GB system RAMFits (by memory)64 GB~2.96 GB

How this was computed

  • Weights + KV cache at the stated context + a runtime/overhead allowance; memory capacity only, not a benchmark.

Next steps

Methodology: weights + KV-cache at 8192 tokens + a runtime overhead allowance, from the recommended download size, parameter count, and quant. Computed from stated size and quant, not a runtime guarantee. Not a benchmark; runtimes differ.