February 2025 Summaries

Perplexity has launched Deep Research, a web-based mode that autonomously conducts extensive searches, reviews hundreds of sources, iteratively reasons through findings, and produces comprehensive reports in roughly two to four minutes. Designed for complex work in areas including finance, marketing, technology, health, product research, travel, and current affairs, the feature can export reports as PDFs or documents or convert them into shareable Perplexity Pages. Deep Research is free with limited daily access for non-subscribers and higher usage limits for Pro subscribers, with iOS, Android, and Mac availability planned. Perplexity reports that the tool scored 21.1% on Humanity’s Last Exam and 93.9% on SimpleQA, while completing most tasks in under three minutes.

Perplexity has made its latest in-house Sonar model available to Pro users, positioning it as a fast, factual, and readable default option for search-based questions. Built on Llama 3.3 70B and further trained for answer quality, Sonar reportedly outperformed GPT-4o mini and Claude 3.5 Haiku in online user-satisfaction testing, exceeded Claude 3.5 Sonnet, and approached GPT-4o performance at lower cost. Running on Cerebras infrastructure, it generates up to 1,200 tokens per second, which Perplexity says is nearly ten times faster than comparable models such as Gemini 2.0 Flash. The company also reports improvements over the base Llama model in factual grounding, conflict resolution, readability, instruction following, and broad world knowledge, measured through evaluations including IFEval and MMLU. Pro subscribers can select Sonar as their default model, while an API version is available but has not yet been migrated to Cerebras infrastructure.

Perplexity developed a custom RDMA-based networking system for AWS p5 instances to transfer non-contiguous GPU memory regions between machines while supporting dynamic Kubernetes membership and peer-to-peer communication, requirements that did not fit NCCL’s static, collective-oriented model. Using AWS Elastic Fabric Adapter, the Scalable Reliable Datagram protocol, libfabric, GPUDirect RDMA, two-sided SEND/RECV control messages, and one-sided RDMA WRITE data transfers, the system bypasses CPU and kernel data copies for direct GPU-to-network-card communication. Performance depended on managing application-owned registered buffers, pre-posted receives, polling completion queues, and awareness of the instance’s NUMA and PCIe topology, which links H100 GPUs and four 100 Gbps network cards beneath each PCIe switch. After adding queuing, connection warmup, dedicated pinned threads, state sharding, batching, lazy posting, and NUMA-aware resource allocation, the implementation reached 3,108 Gbps across 32 network cards, or 97.1% of the platform’s theoretical 3,200 Gbps bandwidth.