inference/research
verifiedmedium · 0.80updated 2026-10-14

NVIDIA Tesla P40 24GB

nvidia pascal 2016

Fields

id
accelerators/nvidia-tesla-p40
Vendor
nvidia
Architecture
pascal
Release year
2016
Process (nm)
16
Form factors
pcie
VRAM (GB)
24
Memory type
GDDR5
Memory bus (bit)
384
Memory bandwidth (GB/s)
347
Bandwidth basis
24 GB GDDR5 on a 384-bit bus. Widely published at 347 GB/s (384-bit x ~7.2 Gbps effective / 8 = ~346). Vendor spec sheet prints 346-347 GB/s; recorded 347.
Peak FLOPS
precisionfp32tflops12denseyesvendor_claimyessource_id—basissecondary-aggregatorbasis_detailTesla P40 (GP102) single-precision ~12 TFLOPS, the widely-republished figure. Pascal has no tensor cores; FP16 runs at 1/64 FP32 rate so no useful FP16 row exists.unit—precisionint8tflops47denseyesvendor_claimyessource_id—basissecondary-aggregatorbasis_detailTesla P40 INT8 DP4A ~47 TOPS - the Pascal deep-learning feature NVIDIA marketed for inference. This is the ONLY usable compute path for LLM (GGUF Q4/Q8 via DP4A); there is no FP16 tensor path.unit—
TDP (W)
250
Unified memory
no
Consumer part
no
notes
Pascal-generation passive datacenter card - the default floor buy for cheap local LLM. 24 GB GDDR5 at 347 GB/s is the binding constraint: enough to hold a 24-32B GGUF but decode-modelled ~35 tok/s on an 8B Q4_K_M (vramguide, MODELLED not measured). KEY HARDWARE TRAPS: (1) passive card - needs a 3D-printed shroud + high-static-pressure fan for desktop use; (2) EPS-12V 8-pin CPU power connector, NOT a PCIe 8-pin - wrong cable can damage the card; (3) needs Above-4G-Decoding / ReBAR-capable BIOS. SOFTWARE TRAP: Pascal has NO FP16 tensor cores and no INT8 tensor cores - the only usable path is INT8 DP4A (47 TOPS) via GGUF Q4_0/Q8/Q4_K_M in llama.cpp; FP16 runs at 1/64 FP32 and is useless. Prompt eval ~4.7 tok/s makes long-context ingestion painful; interactive chat is fine. eBay 'multi-option' listings ('P4 8GB / P40 24GB / K80 / M40') put a ~$105 headline on the P4 variant; the actual P40 selection inside those listings runs ~$273-310.

Notes

Pascal-generation passive datacenter card - the default floor buy for cheap local LLM. 24 GB GDDR5 at 347 GB/s is the binding constraint: enough to hold a 24-32B GGUF but decode-modelled ~35 tok/s on an 8B Q4_K_M (vramguide, MODELLED not measured). KEY HARDWARE TRAPS: (1) passive card - needs a 3D-printed shroud + high-static-pressure fan for desktop use; (2) EPS-12V 8-pin CPU power connector, NOT a PCIe 8-pin - wrong cable can damage the card; (3) needs Above-4G-Decoding / ReBAR-capable BIOS. SOFTWARE TRAP: Pascal has NO FP16 tensor cores and no INT8 tensor cores - the only usable path is INT8 DP4A (47 TOPS) via GGUF Q4_0/Q8/Q4_K_M in llama.cpp; FP16 runs at 1/64 FP32 and is useless. Prompt eval ~4.7 tok/s makes long-context ingestion painful; interactive chat is fine. eBay 'multi-option' listings ('P4 8GB / P40 24GB / K80 / M40') put a ~$105 headline on the P4 variant; the actual P40 selection inside those listings runs ~$273-310.

Cross-references

Interconnect (1)