☕️ AMD unveils Helios, its Nvidia rival

Hi there, this is your daily ☕️ Devshot.

AMD unveils Helios, its Nvidia rival LINK

• AMD launched Helios, an integrated rack system pairing Instinct MI455X GPUs, EPYC CPUs, networking, and ROCm software to compete with Nvidia's full AI stack for training and inference workloads.
• Microsoft will deploy Helios on Azure for frontier-model inference, add two agentic AI VM services running on 6th Gen EPYC processors, and integrate Pensando DPUs plus Azure Boost for cloud networking.
• AMD said it will ship Helios to Microsoft and other customers in the second half, with enterprises able to access the infrastructure through Azure Foundry Managed Compute.

DoorDash built a cache handling 1.5M requests a second LINK

• DoorDash built Entity Cache, a transparent proxy caching platform that serves over 1.5 million requests per second across 100+ endpoints and 50 services, cutting redundant service-to-service calls in its microservices architecture.
• Built on Envoy and Valkey, it intercepts HTTP and gRPC requests inside the service mesh before they hit upstream services, so teams get centralized caching without touching application code, managed entirely through configuration.
• Optimizations including custom buffer pools, a lock-free single-flight mechanism, and XFetch-based early refresh reduced allocation rates by 50-60%, boosted per-pod throughput ~5x, and cut P99 latency spikes by up to 80%.

Google builds a chip made for Gemini LINK

• Google is reportedly building a server chip, codenamed "Frozen V2," that hardwires key parts of its Gemini model directly into silicon to process AI queries more efficiently, according to The Information.
• By embedding Gemini's architecture into the hardware, the chip cuts computation and data movement, with engineers estimating it could deliver six to ten times more AI tokens per unit of power than Google's current TPUs.
• Rather than replacing TPUs, Frozen V2 would act as a specialised accelerator alongside them, with Google targeting deployment as early as 2028 while engineers finalise which Gemini components get baked in.

AI's key connection protocol gets big update LINK

• The Model Context Protocol ships a significant update next week that changes how servers handle session IDs, making it far easier to run MCP integrations across large, load-balanced deployments serving millions of users.
• Today an MCP client like Claude gets a session ID on connection and resends it on every request, but spreading traffic across server farms that don't share state forces each machine to track IDs others handed out.
• The new version adopts a stateless server-side approach to session IDs, similar to how ordinary websites work, which should cut maintenance overhead and lower the cost of operating MCP servers at scale.

Other news you might like

• Don't add a read replica until you've read this LINK

• Build a React Native Super App With Re.Pack 5 LINK

• Migrating from LangChain to LlamaIndex LINK

• Scaling document classification to 100k+ labels LINK

• X rewrites Android app in Kotlin and Jetpack Compose LINK

• Upcoming GPU Pricing Updates LINK

• Java News Roundup: Value Objects, WildFly 41, TornadoVM, LangChain4j, Oracle AI Agent Studio LINK

• Hyprland 0.56 released, bringing new Lua APIs and other improvements LINK

• Linux Patches Introduce "KNOD" For In-Kernel Network Offloading Directly To AMD GPUs LINK

• GitLab 19.2 Puts AI Agents to Work on the Security Backlog LINK

• ArgoCon Japan 2026: Meeting the Maintainers, enterprise insights, and the road to Argo CD 3.5 LINK

Trending tools

Sim: a workspace for building and deploying AI agents visually or with code, supporting 1,000+ integrations and every major LLM. LINK

FetchSandbox: lets developers and AI agents test webhooks, retries, and async workflows against 60+ real APIs without burning quota or waiting on staging. LINK

Replay QA: captures your website's behavior in a specialized browser, letting you inspect sessions with devtools and share them for faster bug reporting. LINK

Tamamon: a macOS desktop pet that grows as you code with Claude Code, letting you collect 20 species, feed them, and decorate their habitat locally with no accounts or tracking. LINK

Edgee Claude Code Compressor V2: cuts LLM token costs up to 50% via compression, routes requests to cheaper or faster models, and auto-fails over during outages. LINK

Social Fetch: a scraper API that pulls profiles, posts, comments, and metrics from TikTok, Instagram, YouTube, and X, skipping the need to build scrapers yourself. LINK

Trending papers & reports

Quantum image recognition matched or beat classical models on handwritten digit tasks, hitting about 0.90 accuracy versus 0.85 for the classical version at 1,000 samples, but ran slower and needed more computing power. LINK

Data selection scoring gets a math shortcut that speeds up picking the most valuable training examples by about 35,000 times, making it practical to test which scoring method, including a popular diversity metric, actually predicts real model performance on large datasets. LINK

AI Contagion in Social Networks: We study how artificial intelligence (AI) interacts with social communication networks to shape the stability of collective knowledge. Agents exchange information through a network while AI systems generate content and retrain on the aggreg LINK

A robot control system called HyVLA-0.5 covers the whole pipeline, from collecting real-world training data to fine-tuning and deploying the model, showing how to build practical robots end to end rather than just a lab demo. LINK

Self-driving perception software fuses camera and LiDAR data more smartly, cutting false detections while improving accuracy (mAP 0.502) and running faster (139.82 ms vs. 145.91 ms) than its predecessor. LINK

See you tomorrow for a new dose of ☕️ Devshot!