NVIDIA Tesla M40 24GB
nvidia maxwell 2015
Fields
id
accelerators/nvidia-tesla-m40Vendor
nvidia
Architecture
maxwell
Release year
2015
Process (nm)
28
Form factors
pcie
VRAM (GB)
24
Memory type
GDDR5
Memory bus (bit)
384
Memory bandwidth (GB/s)
288
Bandwidth basis
24 GB GDDR5 on a 384-bit bus at ~288 GB/s (Maxwell GM200, 6 Gbps GDDR5 x 384 / 8 = 288). Vendor-published figure.
Peak FLOPS
precisionfp32tflops7denseyesvendor_claimyessource_id—basissecondary-aggregatorbasis_detailTesla M40 (GM200) single-precision ~7 TFLOPS, the widely-republished figure. Maxwell CC5.2 has NO tensor cores and NO INT8 DP4A path.unit—
TDP (W)
250
Unified memory
no
Consumer part
no
notes
Maxwell-generation passive datacenter card - the cheapest 24 GB of anything (~$189 Newegg refurb, HP P8Y46A server pull, captured 2026-10-11) but a TRAP for LLM. Maxwell compute-capability 5.2 has NO INT8 DP4A and NO tensor cores, so llama.cpp falls back to the FP32/CUDA-core path and is SLOWER than the P40 it undercuts by only ~$90, despite identical VRAM. No clean llama-class measurement was found - it exists because it is the cheapest way to hold 24 GB, not because it is fast. Passive cooling + EPS-12V 8-pin same as P40. Modern toolchains are dropping Maxwell. Buy only if ~$189 is truly the ceiling and you accept FP32-only decode.
Notes
Maxwell-generation passive datacenter card - the cheapest 24 GB of anything (~$189 Newegg refurb, HP P8Y46A server pull, captured 2026-10-11) but a TRAP for LLM. Maxwell compute-capability 5.2 has NO INT8 DP4A and NO tensor cores, so llama.cpp falls back to the FP32/CUDA-core path and is SLOWER than the P40 it undercuts by only ~$90, despite identical VRAM. No clean llama-class measurement was found - it exists because it is the cheapest way to hold 24 GB, not because it is fast. Passive cooling + EPS-12V 8-pin same as P40. Modern toolchains are dropping Maxwell. Buy only if ~$189 is truly the ceiling and you accept FP32-only decode.