Inference hardware and open-source policy #126
Today's Letter
NVIDIA, Groq 3 LPX long-context benchmark
- NVIDIA introduced Groq 3 LPX as an interactive inference accelerator for the Vera Rubin platform
- Artificial Analysis measured 3,431 output tokens per second on Gemma 4 31B with a 100K-token context
- Groq 3 LPX uses deterministic compiler scheduling for compute, memory, and chip-to-chip communication
- Each system includes 256 LP30 processing units, 128 GB of shared SRAM, and 96 C2C links per chip at 112 Gbps
- Paired with Vera Rubin NVL72, the accelerator targets high-interactivity serving for models exceeding 2T parameters
- On the SPEED-Bench coding benchmark, Gemma 4 reached 4,767 median output tokens per second and a P80 of 5,520
Source: developer.nvidia.com
Google Play, AnkiDroid Donation Link Dispute
- Google Play rejected AnkiDroid updates over an external Open Collective donation link.
- AnkiDroid has more than 10 million installs and is distributed as a free, open-source Android app.
- Donations are processed by Open Source Collective, a US nonprofit with IRS 501(c)(6) tax-exempt status.
- Google’s policy restricts external payment methods but permits tax-exempt donations under specified conditions.
- Google requested clarification on whether 501(c)(6) status satisfies its tax-exempt donation requirement.
- Updates have been rejected since August 28, 2026, with removal from Google Play scheduled for September 11, 2026, except in India and Russia.
Source: github.com
More: arstechnica.com · antigravity.google · webiterate.dev
Jocoletter curates AI, software, and product trends for developers and builders.
#Google #NVIDIA