NVIDIA V100 32GB PCIe
nvidia volta 2017
Fields
id
accelerators/nvidia-v100-pcie-32gbVendor
nvidia
Architecture
volta
Release year
2017
Process (nm)
12
Form factors
pcie
VRAM (GB)
32
Memory type
HBM2
Memory bus (bit)
4,096
Memory bandwidth (GB/s)
900
Bandwidth basis
32 GB HBM2 on a 4096-bit bus. Same silicon as the SXM2 32GB part (nvidia-v100-sxm2-32gb) which NVIDIA publishes at 900 GB/s; the PCIe card carries the same HBM2 stacks at the same ~900 GB/s but a 250 W cap and lower clocks (FP16 tensor 112 TF vs 125 TF SXM2).
Peak FLOPS
precisionfp32tflops14denseyesvendor_claimyessource_idnv-v100-datasheetbasisvendor-spec-sheetbasis_detailV100 PCIe 32GB FP32 ~14 TFLOPS (vs 15.7 SXM2 - the 250 W PCIe power envelope trims clocks). NVIDIA V100 datasheet PCIe column.unit—precisionfp16tflops112denseyesvendor_claimyessource_idnv-v100-datasheetbasisvendor-spec-sheetbasis_detailV100 PCIe 32GB FP16 tensor ~112 TFLOPS (vs 125 SXM2). Volta first-generation tensor cores - a REAL FP16 path, which is the advantage over every Pascal part in this price band (P40/P100 have no FP16 tensor).unit—
TDP (W)
250
Unified memory
no
Consumer part
no
notes
The low-drama V100 route and the sweet spot of the used-datacenter slice. Same Volta silicon as nvidia-v100-sxm2-32gb but as a passive dual-slot PCIe card - drops into any x16 slot with NO SXM2 carrier/baseboard problem (the PCIe premium ~$300-350 over the bare SXM2 module buys the carrier problem solved). Passive cooling needs an airflow mod for desktop use. REAL advantage over Pascal: Volta first-gen tensor cores give a genuine FP16 path (~112 TF) so vLLM FP16 works properly. Measured-class envelope: ~60+ tok/s claimed on an 8B-class model in the HF 'V100 Homelab Field Guide'; ~106 tok/s modelled by vramguide for Q4_K_M 8B; low-teens on 32B. Street ~$695-900 (HPE P05913-001 $694.99, HP PG500-216 $809-899, Dell pulls ~$758).
Notes
The low-drama V100 route and the sweet spot of the used-datacenter slice. Same Volta silicon as nvidia-v100-sxm2-32gb but as a passive dual-slot PCIe card - drops into any x16 slot with NO SXM2 carrier/baseboard problem (the PCIe premium ~$300-350 over the bare SXM2 module buys the carrier problem solved). Passive cooling needs an airflow mod for desktop use. REAL advantage over Pascal: Volta first-gen tensor cores give a genuine FP16 path (~112 TF) so vLLM FP16 works properly. Measured-class envelope: ~60+ tok/s claimed on an 8B-class model in the HF 'V100 Homelab Field Guide'; ~106 tok/s modelled by vramguide for Q4_K_M 8B; low-teens on 32B. Street ~$695-900 (HPE P05913-001 $694.99, HP PG500-216 $809-899, Dell pulls ~$758).