Nvidia RTX Spark encrypted firmware slows inference speed to 35% on Linux
Nvidia’s new RTX Spark laptops are reportedly throttling GPU clocks under Linux to around a third of their Windows performance, apparently because key power-management firmware is encrypted and only fully used by the Windows driver. Commenters are split between seeing this as an early-stage driver bug on a Windows-first platform and as a warning sign of tighter vendor control, DRM-style feature locking, and a push toward cloud-based AI. The incident feeds into broader worries about Nvidia’s dominance, weak Linux support compared to its datacenter products, and the long‑term stakes for “sovereign” local AI versus centralized services.
Linux performance issue & firmware behavior
- RTX Spark laptops show severe GPU throttling under Linux (≈35% of Windows inference speed), with clocks stuck around 780 MHz instead of ~2 GHz.
- Oddly, performance is fine on battery and degrades when plugged in, opposite of Windows behavior.
- Discussion highlights that the firmware power table is encrypted; current Linux drivers cannot read or apply the same power-management logic as Windows.
Bug vs deliberate restriction
- One camp sees this as classic Nvidia “Windows-first” driver neglect: new Windows-only platform, immature Linux support, likely to be fixed later.
- Another camp views it as intentional segmentation or lock‑in: encrypted power firmware and keys only for Windows drivers are framed as a way to:
- Nudge developers toward Windows/WSL.
- Discourage repurposing RTX Spark laptops as cheap Linux inference nodes in datacenters.
- Several point out Nvidia’s long history of closed firmware, driver blobs, and artificial product crippling, so suspicion is high.
Local vs cloud and “sovereign AI”
- Many fear a broader industry push to centralize AI in the cloud, citing:
- Nvidia and others exploring DRM for model weights.
- Patents and documentation around encrypted/attested local models.
- Others argue this case is just about laptop firmware, not a grand cloud conspiracy, and accuse commenters of confirmation bias.
- “Sovereign AI” is discussed as the ability to run and modify models locally, even if using Nvidia hardware, versus dependence on proprietary cloud APIs.
Laptop vs home server for inference
- Some argue high‑end AI laptops are irrational: better to run GPUs in a desktop/DGX‑style box at home and access via VPN/Tailscale from a thin laptop or phone.
- Counterpoints: portability, unstable housing, and personal preference for laptop form factors still drive demand for powerful mobile inference.
Alternatives, workarounds, and expectations
- Suggestions include reverse‑engineering the firmware protocol, reusing DGX Spark drivers, or waiting for Nvidia to ship proper Linux support.
- Frustration is expressed that new hardware routinely ships with poor Linux power management. Some vow to avoid Nvidia and favor AMD/Intel or other accelerators.
- Others note that, historically, Nvidia’s proprietary Linux drivers have been high‑performance; this new firmware-encryption wrinkle feels like a regression.