April 2025 Summaries

Perplexity has announced a global partnership with Motorola that will pre-install its answer engine and assistant on millions of new Motorola smartphones, including Razr and Edge 60 models. The integration is optimized for Motorola hardware and software, offering functionality on the folded Razr external display, access through Moto AI prompts, screen-based related-question exploration via Next Move, and assistant actions such as sending emails, setting reminders, playing media, requesting rides, and booking restaurants. Buyers of eligible new Motorola devices will receive three months of Perplexity Pro, including Deep Research, advanced AI models such as Sonar, Claude 3.5 Sonnet, and GPT-4o, Pro Shopping, unlimited file uploads, and enhanced search features. The companies describe the arrangement as one of Perplexity’s most comprehensive mobile-brand integrations and as an effort to make AI-powered information retrieval and task completion more readily available on smartphones.

Perplexity argues that the U.S. Department of Justice’s remedy phase in its antitrust case against Google should focus on restoring consumer and business choice rather than forcing Google to sell Chrome or license its data. It contends that Google used contracts, app-bundling requirements, default-placement agreements, and revenue penalties to require Android device makers and carriers to preload and prioritize Google services, limiting alternatives despite Android’s open-source foundation. The company says consumers often face hidden or difficult-to-change defaults, while manufacturers and carriers risk financial or access-related consequences for offering competing products. Although it acknowledges Google’s contributions through Android and Chromium and argues that Chrome earned its market position through product quality, Perplexity supports separating Android access from mandatory Google app bundles and eliminating penalties for partners that provide alternatives. It presents its own AI-based search and assistant products as an additional option rather than a replacement for Google, maintaining that competition should be determined by product quality and informed user choice.

DeepSeek-V3/R1, a 671-billion-parameter Mixture of Experts model that activates only 37 billion parameters per token, can challenge the usual trade-off between inference latency and throughput by benefiting from larger multi-node GPU deployments. The analysis compares a single 8-GPU H200 node with configurations spanning up to 16 H100 nodes, using tensor, expert, and data parallelism to distribute attention, experts, and requests. Unlike dense models, increasing expert parallelism assigns fewer experts to each GPU, reducing memory-bandwidth pressure and allowing multi-node EP128 configurations to achieve up to five times the throughput of single-node deployments at comparable output speeds. Communication overhead from cross-node AllToAll token dispatch and collection remains a key constraint, but overlapping communication with computation through shared-expert execution and micro-batching can substantially reduce per-layer latency, with micro-batching improving performance at sufficiently large batch sizes. Roofline analysis shows that many GEMM and grouped-GEMM operations remain memory-bound, while speculative multi-token prediction improves matrix efficiency. The implementation uses FP8 block quantization, CUDA and Triton kernels, FlashInfer-based Multi-Latent Attention, CUDA graphs, and custom communication kernels, with future work focused on prefill/decode separation, faster AllToAll and GEMM kernels, improved speculative decoding, and NVLink-connected Blackwell systems.

Perplexity announces that its Enterprise Pro platform has achieved SOC 2 Type II compliance, supplementing its GDPR and PCI DSS adherence to provide independently validated data protection for enterprise customers. Enterprise Pro includes a Security Hub that lets administrators manage user permissions for files, shared content, data connectors, AI models, and internal knowledge bases, alongside SSO, MFA, and short-lived credentials. The company states that customer data is not used to train its systems or third-party models, while connectors, DeepSeek R1 queries, encryption, automatic file deletion, incognito mode, audit logs, and its Trust Center support privacy and transparency. Its security program uses AWS infrastructure, Wiz risk monitoring, continuously monitored incident response, restricted production-data access, and vulnerability disclosure and bug bounty programs. Perplexity says more than 7,000 Enterprise Pro customers, including NVIDIA, Databricks, Stripe, and USADA, use the service, with USADA citing its security controls as a key factor in adoption.

LM Arena’s Search Arena leaderboard, based on more than 10,000 human preference votes on current and complex search queries collected from March 18 to April 13, 2025, ranked Perplexity’s Sonar-Reasoning-Pro-High statistically tied for first with Google’s Gemini-2.5-Pro-Grounding, scoring 1136 versus 1142, while winning 53% of direct comparisons. Perplexity reports that its Sonar models occupied the top four positions and outperformed Google Gemini 2.0 Flash Grounding and OpenAI web-search models, with reasoning-enabled variants placing especially highly. The evaluation found that longer answers, more citations, and citations from community web sources correlated with user preference, while experiments controlling for citations suggested that search depth was an important factor in rankings. Sonar models reportedly cited two to three times as many sources as comparable Gemini systems, and Perplexity presents the results as evidence of strong accuracy, source attribution, and response quality for its Pro and API users, alongside an announcement of an April 24 API overview by co-founder and CTO Denis Yarats.

A portable, open-source Mixture-of-Experts communication library is presented to reduce the latency of routing tokens among distributed GPUs during large-model inference, where experts are sharded across devices and nodes. Built on a minimal set of NVIDIA NVSHMEM primitives, the approach uses GPU-initiated networking through GPUDirect Async, nonblocking remote-memory transfers, and separate send and receive kernels to overlap communication with computation such as attention and shared-expert processing. Its dispatch kernels package and send routed tokens to expert-hosting devices, while combine kernels return expert outputs and compute router-weighted results, using symmetric memory, signaling, and limited synchronization to manage sparse, uneven token traffic. Benchmarks on 128 GPUs across 16 InfiniBand-connected nodes report combined dispatch and combine latency of 902 microseconds with GPU-initiated communication, roughly 10 times faster than dense standard all-to-all operations and 3.6 times faster than a CPU-proxy configuration, while single-node NVLink tests show 2.5 times lower latency than DeepEP. Although it remains about twice as slow as highly hardware-specialized multi-node implementations, the library emphasizes compatibility across NVLink, ConnectX-7, EFA, NVSHMEM versions, and other network environments, and is available through the pplx-kernels GitHub repository.