# GB202 XBAR 2400 MHz override under a non-binding MCLK maximum

Date: 2026-08-11 (Asia/Singapore)

## Scope

Strict Cyberpunk 2077 2.31 A/B on the same RTX 5090 runtime state. Both runs used:

- VBIOS `98.02.2E.80.50`
- driver `610.57.04`
- 6144x3456, Path Tracing, identical game settings
- the same 127-point LACT V/F curve
- memory offset `{0: 2000, 8: 2000}`, measured MCLK approximately 15001 MHz
- board power limit 600 W
- max-only DRAM client `0x75` configured to 16000000 kHz

The sole intended A/B variable was the XBAR strict-client request:

- A: `0xe4 CLIENT_STRICT_XBAR_MAX = 2400000 kHz` and
  `0xe7 CLIENT_STRICT_XBAR_MIN = 2400000 kHz`
- B: neither XBAR client installed

Both DRAM and XBAR requests returned `NV_OK` and were read back before launch.

## Results

| Metric | XBAR min=max 2400 | No XBAR request | A vs B |
|---|---:|---:|---:|
| Average FPS | 11.933449 | 10.141646 | +17.668% |
| Minimum FPS | 10.419945 | 9.215248 | +13.073% |
| Maximum FPS | 13.874783 | 11.562445 | +19.999% |
| GPU clock average | 2934.0 MHz | 2948.5 MHz | -14.5 MHz |
| GPU voltage average | 1026.3 mV | 1038.0 mV | -11.7 mV |
| 1-second board power average | 577.9 W | 492.3 W | +85.6 W |
| XBAR min/avg/max | 2072.5 / 2341.0 / 2407.4 MHz | 1459.4 / 1474.8 / 1488.9 MHz | +866.2 MHz average |
| SYS min/avg/max | 2058.4 / 2324.2 / 2392.1 MHz | 1469.2 / 1482.1 / 1494.9 MHz | +842.1 MHz average |
| MCLK | approximately 15001 MHz | approximately 15001 MHz | unchanged |

No Xid, GPU-lost, reset-required, thermal slowdown, hardware slowdown, or power-brake event occurred.

## Conclusions

**PROVED:** A runtime XBAR strict request can substantially override the low-XBAR/SYS state caused by a finite, non-binding memory maximum on this GB202. It restored 17.668% average rendered throughput while the reported GPU clock decreased by 14.5 MHz. The improvement therefore cannot be attributed to a higher reported GPC clock.

**PROVED:** The 2400 MHz request is not an absolute hardware lock. Under the combined MCLK-max state, measured XBAR occasionally fell to 2072.5 MHz even though both XBAR clients read back as effective at 2400000 kHz. A higher-priority propagated constraint or downstream arbitration can still override the requested floor transiently.

**INFERENCE:** The remaining gap between the requested 2400 MHz and measured 2341 MHz average is caused downstream of the public/RM client readback, likely in shared-rail/topology or final clock arbitration. Locating that arbitration is the next discriminating target.

## Artifacts

- XBAR A run: `benchmark_logs/cp2077_6k-xbar2400-mclkmax16000-vf-mem2k_20260811_032614/`
- No-XBAR B run: `benchmark_logs/cp2077_6k-control-mclkmax16000-vf-mem2k-no-xbar_20260811_033121/`
- A limit logs: `benchmark_logs/cp2077_xbar2400_mclkmax16000_limits_20260811_032612/`
- B limit logs: `benchmark_logs/cp2077_control_mclkmax16000_limits_20260811_033119/`

All temporary DRAM/XBAR clients were cleared after the comparison. LACT V/F and the +2000 memory offset remain applied.