Releases
Download stable releases of the Pantheon GPU toolkit. The newest release is listed first.
Pantheon v1.2.5 (Latest)
Release Date: November 11, 2026
Release Notes
Source: https://github.com/pantheongpu/pantheon/tree/v1.2.5
Changes since v1.2.4:
- Report each card's memory clock, edge and hotspot temperature, thermal margin, seconds held back by each limit, the state a run started in, host CPU load, and the clocks, bus width, PCIe link, core count and temperatures the card is rated for (#55)
- Collect the fields that read N/A: memory type and vendor on older NVIDIA drivers, and AMD memory type, driver, PCIe link, voltage and fan (#54)
- Add nccl_all_reduce: one all-reduce across every selected GPU, through NCCL (RCCL on AMD) (#50)
- p2p_matrix: a failed peer copy is an interconnect WATCH, and record the 8x H100 NVSwitch run (#49)
- NVLink follow-up from an 8x V100 run: read the counters per field, and add p2p_matrix (#48)
- Check the NVLink links and the GPU fabric on multi-GPU NVIDIA nodes (#47)
- CI: run every workload on simulated GPUs (pantheonsim) when a change touches the workloads (#51)
- README: badges, quick start and what it is / isn't (#46)
- Integrations: point the image tags at 1.2.4 (#45)
- (#52 relicensed to AGPL and #53 reverted it before any release: no net change. Pantheon stays Apache-2.0.)
Scores are comparable with 1.2.4. No existing workload changed.
Multi-GPU NVIDIA nodes: a card that passes every test alone passes a node that has lost an NVLink link. Pantheon now reads each GPU's NVLink links and fabric state before and after a run and reports a finding as interconnect (a WATCH, never a FAULT). Two new workloads load the links: p2p_matrix (how fast each GPU sends to every other) and nccl_all_reduce (the collective a training job makes).
Reports carry more: memory clock, both temperature sensors, the seconds each limit was active next to the seconds the driver counted as held back, the state the card started in, and what it is rated for. The new columns read N/A where a card or driver does not say, never a zero. Memory type and vendor now read on NVIDIA drivers older than 580, and AMD cards report their memory type, driver version, PCIe link, voltage and fan.
Not covered: the interconnect checks were run on an 8x V100 NVLink mesh and an 8x H100 NVSwitch node, not on A100 or B200. RCCL on AMD compiles in CI but has not run. AMD throttle seconds are not collected. On the Windows driver under WSL the violation counters also advance while a card idles.
Install from PyPI (pipx install pantheon-gpu), the apt or
COPR repositories, the container image
(ghcr.io/pantheongpu/pantheon:1.2.5), or the wheel below:
Kernels compile on first run into a per-user cache (roughly a
minute, once) and need a CUDA or ROCm toolchain; use
--platform mock to exercise the tooling with no GPU.
Downloads
| File | Format | Size |
|---|---|---|
| Pantheon v1.2.5 Debian Package | .deb |
225.5 KB |
| Pantheon v1.2.5 Python Wheel | .whl |
423.5 KB |
| Pantheon v1.2.5 Source Tarball | .tar.gz |
426.6 KB |
| Pantheon v1.2.5 Source Distribution | .tar.gz |
349.0 KB |
| Pantheon v1.2.5 Source ZIP | .zip |
562.7 KB |
| Pantheon v1.2.5 Checksums | SHA256SUMS |
484 B |
Pantheon v1.2.4
Release Date: October 17, 2026
Release Notes
Source: https://github.com/pantheongpu/pantheon/tree/v1.2.4
Changes since v1.2.3:
- Start each NVIDIA workload with only its own card visible (#41)
- Report a driver with an older NVENC API as a skip in media_enc_virus (#42)
- Skip the matrix workloads cleanly on AMD parts with no matrix instructions (#40)
- Choose the default GEMM format from the part: no BF16 default below Ampere (#39)
- Integrations: point the image tags at 1.2.3 (#38)
Turing and Volta (T4, T4G, V100): 1.2.3 ran mma_virus and omni_virus on a BF16 GEMM path these cards handle about 40 times slower, without saying so. 1.2.4 chooses the default from the card: below Ampere (compute capability 8.0) both tests keep the WMMA tensor stream and print why. Do not use 1.2.3 results for these two workloads on those cards. Ampere and later are unchanged.
Multi-GPU NVIDIA hosts: each workload now starts with only its own card visible. Before, a stray context on GPU 0 could cause random out-of-memory failures when every card was tested at once.
AMD parts without matrix cores (RDNA2, gfx9 before MFMA): mma_virus, omni_virus and transformer_virus now skip with a reason instead of failing to build.
Install from PyPI (pipx install pantheon-gpu), the apt or
COPR repositories, the container image
(ghcr.io/pantheongpu/pantheon:1.2.4), or the wheel below:
Kernels compile on first run into a per-user cache (roughly a
minute, once) and need a CUDA or ROCm toolchain; use
--platform mock to exercise the tooling with no GPU.
Downloads
| File | Format | Size |
|---|---|---|
| Pantheon v1.2.4 Debian Package | .deb |
202.8 KB |
| Pantheon v1.2.4 Python Wheel | .whl |
386.2 KB |
| Pantheon v1.2.4 Source Tarball | .tar.gz |
366.2 KB |
| Pantheon v1.2.4 Source Distribution | .tar.gz |
300.2 KB |
| Pantheon v1.2.4 Source ZIP | .zip |
496.3 KB |
| Pantheon v1.2.4 Checksums | SHA256SUMS |
484 B |
Pantheon v1.2.3
Release Date: October 13, 2026
Release Notes
Source: https://github.com/pantheongpu/pantheon/tree/v1.2.3
Changes since v1.2.2:
- Build fused_attention on ROCm, and compile every kernel for HIP in CI (#37)
- Fix PCIe AER collection, combined thermal throttle labels, and the --suite docs (#31)
- Correct the README's Efficiency formula and throttle reason labels (#33)
- Keep the operator's home directory out of published reports (#29)
- Fix missing sensors, kernel verification, and skipped-workload exit status (#35)
- Fix verdicts: thermal throttling, skipped and failed workloads, hung tests, PCIe error counters, dmesg blame (#34)
- Follow-up to #26: working memory stream and no crash in omni_virus, fixed BF16 GEMM, safe chain constants (#28)
- Fix power stress in omni_virus and the arithmetic tests, add a vendor GEMM tensor stream (#26)
- Add a Grafana dashboard for the node check's Prometheus metrics (#27)
- Add Prometheus metrics, a Kubernetes Job and Helm chart, and an Apptainer definition (#24)
- Compile the workloads ahead of time: --build-only, PANTHEON_PLATFORM, PANTHEON_CUDA_ARCH (#25)
- Add integrations: a node check, Slurm, NHC and ReFrame (#23)
- README: say how to install a release (#22)
- Test only the cards a job was given: honour CUDA_VISIBLE_DEVICES (#21)
- Remove the percentiles and the baselines file; keep the grouped verdict (#20)
- End every run with a verdict, percentiles against the public database, and a quick suite (#18)
- Report memory temperature and thermal throttling correctly on NVIDIA (#17)
- Document what --profile actually requires and what its counters mean (#16)
Scores that are not comparable with 1.2.2:
Ten workloads were reworked in #26 and #28, so their scores and power draw changed on purpose: voltage, incinerator, fp64_virus, pulse_virus, tensor_virus, sfu_stress, memory_thermal_asym, memory_retention_bake, mma_virus and omni_virus. Do not set a 1.2.3 result for these beside a 1.2.2 result. galpat also changes: it now checks all four 32-bit lanes of every cell it reads instead of one, so the same card scores about 6% lower. mma_virus now runs the vendor BF16 GEMM by default; --gemm 0 selects the old WMMA path. On an A10G, none of the other 35 workloads moved by more than 0.4%.
AMD: every kernel now compiles with ROCm 6.4 and 7.2, including fused_attention, which did not build on any ROCm release before. We have not yet run this release on AMD hardware.
Install from PyPI (pipx install pantheon-gpu), the apt or
COPR repositories, the container image
(ghcr.io/pantheongpu/pantheon:1.2.3), or the wheel below:
Kernels compile on first run into a per-user cache (roughly a
minute, once) and need a CUDA or ROCm toolchain; use
--platform mock to exercise the tooling with no GPU.
Downloads
| File | Format | Size |
|---|---|---|
| Pantheon v1.2.3 Debian Package | .deb |
201.0 KB |
| Pantheon v1.2.3 Python Wheel | .whl |
383.7 KB |
| Pantheon v1.2.3 Source Tarball | .tar.gz |
361.1 KB |
| Pantheon v1.2.3 Source Distribution | .tar.gz |
297.5 KB |
| Pantheon v1.2.3 Source ZIP | .zip |
490.6 KB |
| Pantheon v1.2.3 Checksums | SHA256SUMS |
484 B |
Pantheon v1.2.2
Release Date: September 9, 2026
Release Notes
Source: https://github.com/pantheongpu/pantheon/tree/v1.2.2
Changes since v1.2.1:
- Keep counters with the same name but different units apart (#14)
Install from PyPI (pipx install pantheon-gpu), the apt or
COPR repositories, the container image
(ghcr.io/pantheongpu/pantheon:1.2.2), or the wheel below:
Kernels compile on first run into a per-user cache (roughly a
minute, once) and need a CUDA or ROCm toolchain; use
--platform mock to exercise the tooling with no GPU.
Downloads
| File | Format | Size |
|---|---|---|
| Pantheon v1.2.2 Debian Package | .deb |
168.8 KB |
| Pantheon v1.2.2 Python Wheel | .whl |
334.8 KB |
| Pantheon v1.2.2 Source Tarball | .tar.gz |
259.7 KB |
| Pantheon v1.2.2 Source Distribution | .tar.gz |
244.1 KB |
| Pantheon v1.2.2 Source ZIP | .zip |
377.3 KB |
| Pantheon v1.2.2 Checksums | SHA256SUMS |
484 B |
Pantheon v1.2.1
Release Date: September 9, 2026
Release Notes
Source: https://github.com/pantheongpu/pantheon/tree/v1.2.1
Changes since v1.2.0:
- Report a missing openpyxl as an optional extra, not a failed write (#13)
- Attribute profile counters to the kernel that owned the run (#12)
- Point the banner and final summary at GitHub Discussions (#9)
- Bump version to 1.2.1 (#8)
- Report the declared memory type and vendor of each GPU (#7)
- List every suite in the --test help text (#6)
- Add funding metadata: .well-known manifest pointer and FUNDING.yml (#5)
Install from PyPI (pipx install pantheon-gpu), the apt or
COPR repositories, the container image
(ghcr.io/pantheongpu/pantheon:1.2.1), or the wheel below:
Kernels compile on first run into a per-user cache (roughly a
minute, once) and need a CUDA or ROCm toolchain; use
--platform mock to exercise the tooling with no GPU.
Downloads
| File | Format | Size |
|---|---|---|
| Pantheon v1.2.1 Debian Package | .deb |
168.2 KB |
| Pantheon v1.2.1 Python Wheel | .whl |
334.2 KB |
| Pantheon v1.2.1 Source Tarball | .tar.gz |
258.2 KB |
| Pantheon v1.2.1 Source Distribution | .tar.gz |
243.5 KB |
| Pantheon v1.2.1 Source ZIP | .zip |
375.8 KB |
| Pantheon v1.2.1 Checksums | SHA256SUMS |
484 B |
Pantheon v1.2.0
Release Date: August 31, 2026
Release Notes
Source: https://github.com/pantheongpu/pantheon/tree/v1.2.0
Changes since v1.1.0:
- Make bring-your-own-OptiX usable: cache key, docs, and a skip message that says what to do (#3)
Install from PyPI (pipx install pantheon-gpu), the apt or
COPR repositories, the container image
(ghcr.io/pantheongpu/pantheon:1.2.0), or the wheel below:
Kernels compile on first run into a per-user cache (roughly a
minute, once) and need a CUDA or ROCm toolchain; use
--platform mock to exercise the tooling with no GPU.
Downloads
| File | Format | Size |
|---|---|---|
| Pantheon v1.2.0 Debian Package | .deb |
163.9 KB |
| Pantheon v1.2.0 Python Wheel | .whl |
328.6 KB |
| Pantheon v1.2.0 Source Tarball | .tar.gz |
249.6 KB |
| Pantheon v1.2.0 Source Distribution | .tar.gz |
237.4 KB |
| Pantheon v1.2.0 Source ZIP | .zip |
366.6 KB |
| Pantheon v1.2.0 Checksums | SHA256SUMS |
484 B |
Pantheon v1.1.0
Release Date: August 30, 2026
Release Notes
Source: https://github.com/pantheongpu/pantheon/tree/v1.1.0
Install the wheel (kernels compile on first run into a per-user cache; expect roughly a minute the first time):
Requires a CUDA or ROCm toolchain at first run. Use --platform mock
to exercise the tooling with no GPU.
Downloads
| File | Format | Size |
|---|---|---|
| Pantheon v1.1.0 Python Wheel | .whl |
328.1 KB |
| Pantheon v1.1.0 Source Tarball | .tar.gz |
249.2 KB |
| Pantheon v1.1.0 Source Distribution | .tar.gz |
236.3 KB |
| Pantheon v1.1.0 Source ZIP | .zip |
366.2 KB |
| Pantheon v1.1.0 Checksums | SHA256SUMS |
389 B |
Pantheon v1.0.19
Release Date: August 27, 2026
Release Notes
What's Changed
- Fault maps and a time-based memory retention workload by @saqibkh
- Fold the memory _agg twins into --init_pattern by @saqibkh
- Audit fixes: unreleased hardware, fictional metrics, broken verification, CI hardening by @saqibkh
- Bound launch parameters that could hang the GPU by @saqibkh
- Make the tree publishable: OptiX optional, NOTICE, export script by @saqibkh
- Stop CI running the whole matrix twice per commit by @saqibkh
- Kill the whole process tree, not just the launched process by @saqibkh
- Release v1.0.19 by @saqibkh
Full Changelog: v1.0.18...v1.0.19
Downloads
| File | Format | Size |
|---|---|---|
| Pantheon v1.0.19 Debian Package | .deb |
33.7 MB |
| Pantheon v1.0.19 Tarball | .tar.gz |
67.4 MB |
| Pantheon v1.0.19 ZIP Bundle | .zip |
67.4 MB |
| Pantheon v1.0.19 Checksums | SHA256SUMS |
303 B |
Pantheon v1.0.18
Release Date: August 26, 2026
Release Notes
What's Changed
- Kernel verification hardening: coverage holes, uninitialized lanes, UB, and accounting fixes by @saqibkh
Full Changelog: v1.0.17...v1.0.18
Downloads
| File | Format | Size |
|---|---|---|
| Pantheon v1.0.18 Debian Package | .deb |
33.6 MB |
| Pantheon v1.0.18 Tarball | .tar.gz |
67.3 MB |
| Pantheon v1.0.18 ZIP Bundle | .zip |
67.3 MB |
| Pantheon v1.0.18 Checksums | SHA256SUMS |
303 B |
Pantheon v1.0.17
Release Date: August 26, 2026
Release Notes
Full Changelog: v1.0.16...v1.0.17
Downloads
| File | Format | Size |
|---|---|---|
| Pantheon v1.0.17 Debian Package | .deb |
33.6 MB |
| Pantheon v1.0.17 Tarball | .tar.gz |
67.3 MB |
| Pantheon v1.0.17 ZIP Bundle | .zip |
67.3 MB |
| Pantheon v1.0.17 Checksums | SHA256SUMS |
303 B |
Pantheon v1.0.16
Release Date: August 22, 2026
Release Notes
Full Changelog: v1.0.15...v1.0.16
Downloads
| File | Format | Size |
|---|---|---|
| Pantheon v1.0.16 Debian Package | .deb |
33.6 MB |
| Pantheon v1.0.16 Tarball | .tar.gz |
67.2 MB |
| Pantheon v1.0.16 ZIP Bundle | .zip |
67.2 MB |
| Pantheon v1.0.16 Checksums | SHA256SUMS |
303 B |
Pantheon v1.0.14
Release Date: August 15, 2026
Release Notes
Full Changelog: v1.0.13...v1.0.14
Downloads
| File | Format | Size |
|---|---|---|
| Pantheon v1.0.14 Debian Package | .deb |
33.5 MB |
| Pantheon v1.0.14 Tarball | .tar.gz |
67.1 MB |
| Pantheon v1.0.14 ZIP Bundle | .zip |
67.1 MB |
| Pantheon v1.0.14 Checksums | SHA256SUMS |
303 B |
Pantheon v1.0.13
Release Date: June 18, 2026
Release Notes
What's Changed
- [codex] Fix source and AMD execution paths by @saqibkh
Full Changelog: v1.0.12...v1.0.13
Downloads
| File | Format | Size |
|---|---|---|
| Pantheon v1.0.13 Debian Package | .deb |
33.4 MB |
| Pantheon v1.0.13 Tarball | .tar.gz |
66.8 MB |
| Pantheon v1.0.13 ZIP Bundle | .zip |
66.8 MB |
| Pantheon v1.0.13 Checksums | SHA256SUMS |
303 B |
Pantheon v1.0.12
Release Date: June 18, 2026
Release Notes
What's Changed
- [codex] Prepare Pantheon 1.0.12 release by @saqibkh
Full Changelog: v1.0.11...v1.0.12
Downloads
| File | Format | Size |
|---|---|---|
| Pantheon v1.0.12 Debian Package | .deb |
33.4 MB |
| Pantheon v1.0.12 Tarball | .tar.gz |
66.8 MB |
| Pantheon v1.0.12 ZIP Bundle | .zip |
66.8 MB |
| Pantheon v1.0.12 Checksums | SHA256SUMS |
303 B |
Pantheon v1.0.10
Release Date: June 16, 2026
Release Notes
What's Changed
- pantheon tuning params by @saqibkh
- Debian binary release by @saqibkh
- Debian binary release by @saqibkh
Full Changelog: v1.0.8...v1.0.10
Downloads
| File | Format | Size |
|---|---|---|
| Pantheon v1.0.10 Debian Package | .deb |
87.4 MB |
| Pantheon v1.0.10 Tarball | .tar.gz |
174.8 MB |
| Pantheon v1.0.10 ZIP Bundle | .zip |
174.8 MB |
| Pantheon v1.0.10 Checksums | SHA256SUMS |
303 B |
Pantheon v1.0.8
Release Date: May 21, 2026
Release Notes
What's Changed
- INitial Commit by @saqibkh
- Package Nuitka release archives by @saqibkh
- Update VERSION by @saqibkh
- Automate by @saqibkh
New Contributors
- @saqibkh made their first contribution
Full Changelog: v1.0.7...v1.0.8
Downloads
| File | Format | Size |
|---|---|---|
| Pantheon v1.0.8 Tarball | .tar.gz |
181.1 KB |
| Pantheon v1.0.8 ZIP Bundle | .zip |
247.6 KB |
| Pantheon v1.0.8 Checksums | SHA256SUMS |
173 B |
v1.0.7
Release Date: April 6, 2026
Release Notes
Pantheon v1.0.7 - SDC Validation & FP64 Fixes
What's New in this Release: This update introduces critical diagnostic enhancements for memory integrity and patches the double-precision compute stressor.
-
Active SDC Catching (--verify): Added the --verify flag to actively hunt for Silent Data Corruption (SDC). Instead of just generating extreme heat and waiting for a hardware crash or driver timeout, Pantheon will now actively validate the data payloads returning from the GPU. If the hardware ECC fails to catch a bit-flip caused by thermal or electrical stress, Pantheon will immediately flag the corrupted block.
-
fp64_virus Patched: Fixed the execution and reporting logic for the Double Precision Chokehold (fp64_virus). The kernel now properly saturates the FP64 datapath, accurately exposing physical and artificial silicon limits (such as the strict 1/64th FP64 throttle implemented on consumer NVIDIA GeForce cards).
Downloads
| File | Format | Size |
|---|---|---|
| Pantheon v1.0.7 Tarball | .tar.gz |
131.2 MB |
| Pantheon v1.0.7 ZIP Bundle | .zip |
131.2 MB |