Used / secondary market: NVIDIA Tesla M40 24GB (mining-era / server pull)
various server liquidators and refurbished parts resellers on Newegg · US / global (Newegg refurbished/marketplace, USD listings)
Fields
id
supply/used-nvidia-tesla-m40-marketKind
used-market
Vendor
various server liquidators and refurbished parts resellers on Newegg
Region
US / global (Newegg refurbished/marketplace, USD listings)
Channels
marketplace
Price (USD)
189
Price basis
USED / Server pull / Refurbished bare passive PCIe accelerator card. Observed 2026-10-11 on Newegg marketplace. $189.00 = HP OEM Tesla M40 24GB (P8Y46A / 855178-001 / 900-2G600-0310-030); Dell OEM 0MY6DR listings ask $399.00. PERISHABLE-FIGURE STAMP: OBSERVED 2026-10-11. RE-VERIFY BY: 2027-01-11.
Availability
in-stock
Lead time source
structural-na
Lead time basis
lead_time_weeks is null because inventory is in-stock at surplus liquidators.
Export controlled
no
Export control regime
EAR 3A001 or EAR99 (legacy Maxwell architecture below 3A090.a thresholds)
Export control note
Legacy Maxwell enterprise accelerator below 3A090.a thresholds.
notes
OBSERVATION DATE 2026-10-11. Source sup-used-newegg-tesla-m40. The Tesla M40 24GB (Maxwell GM200, 24GB GDDR5) is the absolute cheapest 24GB accelerator on the secondary market ($189.00, or ~$7.88 per GB of VRAM). Severe architectural gotchas: Maxwell (compute capability 5.2) lacks both tensor cores and INT8 DP4A instructions; inference is limited to FP32 or emulated formats in llama.cpp, delivering slow decode tokens/sec; passive cooling requires blower mods; EPS-12V 8-pin power wiring differs from standard PCIe 8-pin; driver support capped on modern CUDA toolchains. The corpus contains no standalone accelerator record for tesla-m40, so accelerator_ids is empty []. Direct eBay queries blocked via HTTP 403 on 2026-10-11.
Notes
OBSERVATION DATE 2026-10-11. Source sup-used-newegg-tesla-m40. The Tesla M40 24GB (Maxwell GM200, 24GB GDDR5) is the absolute cheapest 24GB accelerator on the secondary market ($189.00, or ~$7.88 per GB of VRAM). Severe architectural gotchas: Maxwell (compute capability 5.2) lacks both tensor cores and INT8 DP4A instructions; inference is limited to FP32 or emulated formats in llama.cpp, delivering slow decode tokens/sec; passive cooling requires blower mods; EPS-12V 8-pin power wiring differs from standard PCIe 8-pin; driver support capped on modern CUDA toolchains. The corpus contains no standalone accelerator record for tesla-m40, so accelerator_ids is empty []. Direct eBay queries blocked via HTTP 403 on 2026-10-11.