@@ -0,0 +1,82 @@
+# HYDRA VRAM/Power-limit unlock research notes
+
+> Status: research only. No code from HYDRA has been copied or merged into this
+> project. HYDRA Pro appears to be closed-source/paid; these notes are for
+> evaluating whether a similar feature can be implemented in xbar5090.
+
+## What HYDRA is
+
+- Developer: **1usmus**
+- Latest update/change feed on GitHub:
+  <https://github.com/1usmus/HYDRA_UPDATE>
+- According to public coverage (2026-08), **HYDRA 2.3B Pro** adds support for
+  NVIDIA RTX 50-series:
+  - unlocks GDDR7 memory-clock limits to **36 Gbps** (from the usual 34 Gbps or
+    lower on some models),
+  - raises the power-limit adjustment range to **125%**,
+  - works without flashing a modified VBIOS,
+  - currently distributed as a paid "Pro" version.
+
+## Why it is not directly mergeable
+
+The GitHub `HYDRA_UPDATE` repository contains only:
+
+- `HYDRA_CHANGELOG.txt`
+- `HYDRA_UPDATE_INFO.txt`
+
+It does **not** contain the HYDRA source code. The tool itself is closed-source.
+Therefore we cannot copy the VRAM-unlock implementation directly into xbar5090.
+
+## What would be needed to implement a similar feature
+
+1. **Find the private NvAPI/RM command** used by HYDRA on Windows for:
+   - memory-clock/PLL limit override,
+   - power-limit override beyond the VBIOS default.
+2. **Reverse-engineer the buffer layouts** for the relevant GET/SET control
+   functions on the validated driver family (R572..R610), in the same way this
+   project already did for XBAR/MSVDD/ratio/VF.
+3. **Add read-only probe first**, then a backed-up/read-back SET with strict
+   safety limits, and finally an L2/game stability validation.
+4. **Handle driver-branch differences**; the current project already has a
+   `probe`/`crack` mechanism that could be extended.
+
+## Risks
+
+- VRAM overclocking beyond the factory lock can cause data corruption, visual
+  artifacts, or hardware damage even if the tool reports success.
+- The power-limit and memory-limit controls are likely private RM/NvAPI
+  interfaces and may be driver/VBIOS-specific.
+- HYDRA Pro is paid; reverse-engineering its behavior for compatibility may
+  raise legal/ToS questions. Independent probing of the NVIDIA driver is safer.
+
+## Download / payment (as of 2026-08)
+
+- Official update feed: <https://github.com/1usmus/HYDRA_UPDATE>
+- Official discussion/payment thread is the Overclock.net "Official HYDRA (and
+  HYDRA PRO) Thread"; the exact price/subscription method is behind the forum
+  and was not scrapeable from this environment (WAF/Cloudflare).
+- Public coverage states the RTX 50 VRAM/power unlock is in the **paid Pro**
+  tier; there is no open-source implementation in the GitHub update repo.
+
+## Feasibility for xbar5090
+
+- XBAR was cracked because the private NvAPI/RM control for clock domains is
+  exposed and reversible. A 36 Gbps memory unlock would need the same class of
+  private control for the **memory clock/PLL limit** and/or **power-limit SET**.
+- Current `perf_limits` module is read-only; the validated Windows driver did
+  not expose the PERF SET command.
+- A quick read-only dump of `clk_domains` on the author's RTX 5090 shows only
+  the XBAR domain entry nonzero; the memory clock limit is not visible through
+  the same simple offset path, so additional reverse engineering is required.
+- Conclusion: technically the same methodology could be attempted, but it is a
+  separate reverse-engineering project with higher risk. Do not expect a quick
+  merge from HYDRA; it is closed-source and paid.
+
+## Suggested next step
+
+- Run HYDRA on the same system and capture which NvAPI/RM functions it calls
+  (e.g. with API Monitor/x64dbg or by observing `nvapi64.dll` exports used).
+- Compare those IDs against known RM command tables in
+  `docs/reverse/` and `scripts/validate_nvapi_drivers.py`.
+- If a stable read-only GET can be found, add it to `candidates.json` and
+  extend `probe.py`/`crack.py` before attempting any SET.