Summary

  • NVIDIA prints dated end-of-support for driver branches (R580 June 2028, R595 March 2027, R620 September 2029), for vGPU software per GPU model, for AI Enterprise hardware removal and for DGX A100 Station systems (12/15/2026 and 12/30/2027); in none of the sixteen documents reviewed does it print one for a standalone A100 or H100 accelerator.
  • The instrument that retires hardware acts on architecture families, and the terms are printed per tier: Volta, compute capability 7.0, first toolkit CUDA 9.0, last toolkit CUDA 12.x, last driver R580. A card's usable CUDA version is bounded by a driver branch number reported at runtime, not by a calendar date printed at purchase.

An A100 arrives with an invoice, a memory figure, a thermal specification and a declared peak. It does not arrive with a date. Ask NVIDIA's own publications when the card stops being supported by current CUDA toolkit releases and the answer divides into three registers that never share a table.

The register that assigns, and retires nothing

The CUDA GPU compute capability table fixes its own subject in one sentence: "Compute capability (CC) defines the hardware features and supported instructions for each NVIDIA GPU architecture." Beneath it, every data-centre part sits on a number. GB300, B300 and GB300 DGX Station at 10.3; GB200 and B200 at 10.0; GH200, H200 and H100 at 9.0; L4, L40 and L40S at 8.9; A100 and A30 at 8.0; T4 at 7.5. The table has no lifecycle column, no end-of-support date, no last-supported-version field. It assigns tiers. It retires nothing.

The register that retires, and names families

The removal lives in the archived CUDA Toolkit 13.0 release notes, and its subject is a family rather than a part: "Architecture support for Maxwell, Pascal, and Volta is considered feature-complete. Offline compilation and library support for these architectures have been removed in CUDA Toolkit 13.0 major version release. The use of CUDA Toolkits through the 12.x series to build applications for these architectures will continue to be supported, but newer toolkits will be unable to target these architectures."

Three sentences, three different acts. The first assigns a status; the second performs a removal; the third names the surviving build path and the one-way consequence. Note what the second sentence removes: compilation and libraries, not the accelerator.

The register that prints last-supported values

The CUDA Toolkit, Driver and Architecture Matrix is the only document in the set that prints a last-supported toolkit and a last-supported driver for a tier:

The register that prints last-supported values
Architecture Compute capability First CUDA Toolkit Last CUDA Toolkit Last driver
Volta 7.0 CUDA 9.0 CUDA 12.x R580

Other rows place Hopper at 9.0, Ampere at 8.0 and 8.6, and Turing at 7.5. My arithmetic on the Volta row's version labels: 9, 10, 11 and 12 are four major CUDA families, and 13 is excluded by the value in the last-toolkit column. That count is mine, not a printed figure.

Now put an A100 against those three registers. The tier table assigns it 8.0. The retirement instrument names families, not parts. The matrix prints a last-supported row for Volta and no equivalent last-toolkit or last-driver value for the 8.0 row where A100 sits, or the 9.0 row where H100 sits. The tier is named, the terms are printed, and the printed terms belong to the tier that has already been retired.

Where the dates actually are

They exist in quantity, keyed to other objects.

The driver branch table, last updated 4 November 2026, prints three dates against three branches: R580, a Long Term Support Branch, End of Life June 2028; R595, a Production Branch, March 2027; R620, a Long Term Support Branch, September 2029. Each carries a maximum of CUDA 13 without the forward-compatibility package. The page also states the limit of its own instrument: "Only Long Term Support and Production branches are listed. New Feature Branches are short-lived and do not have an end-of-life date."

The driver lifecycle policy supplies the periods and, in the same block, the discretion. A Production Branch gets "up to 1 year" of bug fixes and security updates. LTSB releases "will receive bug updates and critical security updates, on a reasonable effort basis, through minor releases during the 3 years that they are supported." Then: "General guidance only. The actual security update and release cadence can change at NVIDIA's discretion."

Two implied introductions fall out of those printed values, and they are mine rather than NVIDIA's: June 2028 minus three years reads June 2025 for R580, and September 2029 minus three years reads September 2026 for R620. The introductions themselves are not printed anywhere in the set.

Per-GPU dates exist too, in one product line. The vGPU end-of-software-support table prints, for the A40, End of Full Support 2024-12 and End of Extended Full Support 2027-12, and leaves the End of Maintenance Support phase undated in that table; for Tesla V100 variants, End of Maintenance Support 2028-07; and it lists A100 variants and H100 PCIe among GPUs that "have not reached the end of the full support phase and remain fully supported", with no end date in that row. The subject is vGPU software support. My subtraction on the A40's two printed months: 36 months apart.

The vGPU hardware FAQ prints that clock's trigger and its precondition: "Software support continues for at least six years after the last GPU of that model is shipped. An active SUMS (Support Updates and Maintenance Subscription) or software subscription is required to receive software support." The clock starts at the last shipment of a model, an event no invoice records.

The AI Enterprise end-of-life notices key hardware removal to a software train: V100, RTX 4000 SFF Ada, RTX A4000 and Quadro RTX 8000, 6000 and 4000 are deprecated and "Removed starting in 8.0 (still supported on Infra 7.x LTSB; Infra 4.10 LTSB reached EOL July 2026)". No H100 or A100 removal date appears there.

Complete systems get a dated notice of their own. NVIDIA Enterprise Support's DGX A100 Station notice, published 21 October 2026, states that the systems "will reach their End of Service Life (EOSL) on the dates noted below": the 160 GB model 920-23487-2530-000 on 12/15/2026, and the 320 GB model 920-23487-2531-000 on 12/30/2027, with service-contract renewal beyond EOSL not available. The subject is a station, not the accelerator inside it.

The real ceiling is a number reported at runtime

The architecture matrix states the mechanism plainly: "The CUDA driver provides an API that is backwards compatible. Thus, new NVIDIA drivers will always work with (applications compiled with) an older CUDA toolkit." Then the constraint: "Each CUDA Toolkit however, requires a minimum version of the NVIDIA driver. Corollarily, when using tools such as nvidia-smi, the NVIDIA driver reports a maximum version of CUDA supported and thus is able to run applications built with CUDA Toolkits up to that version."

The nvidia-smi documentation names the field that carries it: "The CUDA 'User Mode Driver' version. This is the latest CUDA version supported by the driver. This is usually, but not always, the version of the CUDA toolkit installed on the system." The ceiling on a specific machine is therefore read from that machine.

The minor-version compatibility page prints the ranges. "From CUDA 11 onwards, applications compiled with a CUDA Toolkit release from within a CUDA major release family can run, with limited feature-set, on systems having at least the minimum required driver version as indicated below." The table gives CUDA 13.x a floor of 580; CUDA 12.x a floor of 525 and a printed "< 580 (newer drivers still supported through backward compatibility)"; CUDA 11.x a floor of 450 with the same reversing parenthetical. The middle column is a floor. The right-hand column reads as a ceiling and then cancels itself inside the same cell.

What the product pages print

The H100 page names the architecture and one term: "H100 uses breakthrough innovations based on the NVIDIA Hopper™ architecture" and "NVIDIA H100 NVL comes with a five-year NVIDIA AI Enterprise subscription." A five-year software subscription is a term attached to a bundle; it is not an error. It is also not an accelerator end-of-support date, and the page prints none.

The A100 page returned a short marketing excerpt. "NVIDIA A100 Tensor Core GPU Unprecedented acceleration at every scale Enterprise-Ready Software for AI." No support-end date appears in what was returned, and the excerpt is too short to assess the page's full specification or lifecycle content.

The H100 NVL product brief, published 14 March 2024, prints three floors: Linux and Windows driver support R535 or later, CUDA 12.2 or later on x86, and vGPU 16.1 or later for Virtual Compute Server Edition. Three minimums, no maximum, no date.

The CUDA Toolkit archive dates the retirement instrument rather than the hardware: CUDA Toolkit 13.0.0 in August 2025 and CUDA Toolkit 13.5.0 in November 2026. Those are release dates.

And the licence is not a schedule. The CUDA Toolkit EULA, last updated 26 January 2026, allows NVIDIA to terminate at any time if it "decides to no longer provide the SDK in a country or, in NVIDIA's sole discretion, the continued use of it is no longer commercially viable", whereupon the user must "promptly discontinue use of the SDK and destroy all copies in your possession or control". That is a licence-termination clause. It prints no date.

What an invoice holder can compute

From these sixteen documents, a buyer can establish a good deal. The card's tier: 8.0 for A100, 9.0 for H100. The CUDA major family the installed branch supports, read on the machine itself. The vendor's stated floors for one H100 variant as of March 2024. The branch end-of-life months, and the three-year LTSB period that precedes them. Dated software-support schedules for vGPU parts, for AI Enterprise hardware removal and for two DGX A100 Station model numbers.

What the sixteen documents do not supply is a per-accelerator end-of-support date for the standalone card. That statement is about this corpus, not about everything NVIDIA publishes. Two requested addresses, the CUDA C++ Programming Guide and the driver-download page, returned no page text at all; they are not cited here and nothing is asserted about what they print.

Reading limits belong beside the claims they qualify. Quotations are page text as returned by retrieval of the exact addresses, not native browser renders. Several driver and vGPU pages arrived as extracts rather than full renders, and the branch-number-to-column assignment in the end-of-life table comes from that returned text. The architecture matrix's multi-value cells may collapse repeated values in extraction, so the row above is reproduced as returned.

BTW tracks NVIDIA in its directory.