Hot Chips 2026: AMD MI400 and Fujitsu Monaka Debut, Alongside New LLM and SQL Reliability Tools
At Hot Chips 2026, AMD detailed the Helios platform of its MI400 series, pairing EPYC Venice CPUs with Instinct MI455X GPUs and Pensando Vulcano AI NICs, and stressed co-designing AI infrastructure. Fujitsu introduced Monaka, a 2nm server CPU with 144 cores aimed at energy efficiency and modern AI workloads.
Separately, a new protocol demonstrated byte-identical inference for a 72B-parameter LLM across NVIDIA H100 and AMD MI300X, attaching each output to a portable, offline-verifiable signed receipt. SIMURG, a numpy-only monitor, watches LLM token streams character by character for decoding corruption like repetition loops, cross-lingual drift, and structural garbage, releasing the first 350 characters only after verification.
For code review, Provensql compares two SQL queries and returns EQUIVALENT, DIFFERENT, SCHEMA_CHANGE, or UNKNOWN; it never returns a false EQUIVALENT, produced zero false positives across 511 mutations, and uses an SMT engine for the conjunctive fragment with counterexample search as fallback. In bioinformatics, Bonsai offers a distortion-free tree method for single-cell RNA sequencing, addressing t-SNE and UMAP structure hallucinations by representing cells as leaves of a reconstructed tree.
On the frontend, yapyak is a Vite-based i18n compiler that uses source strings as keys and auto-generates translations on save, with SSR adapters for TanStack Start and React Router. Oj is a Rust native replacement for Vite, compatible with existing config and plugins but with faster cold starts and lower memory use. A browser-native media processing report showed a 92% success rate was misleading because out-of-limit videos were excluded from the denominator; including user attempts and adjusting admission control by device capabilities raised task completion by up to 6%.
IBM's Technology Lifecycle Services built a production multi-agent system that gives employees a single interface for inquiries, emphasizing user identity preservation, data access challenges, and cross-team alignment.
» Statistics
- Posts
- 37
- Reads
- 0
- Avg. score
- 7.5
» Most read
- Cross-vendor byte-identical inference for a 72B LLM (AMD MI300X vs. Nvidia H100)
- Bonsai: A Distortion-Free Tree Method for Single-Cell Data Visualization
- Building Production Agentic AI at IBM: Architecture, Decisions, and Lessons
- 900 Data-Agent Runs: Insights on Accuracy, Reproducibility, and Errors
- Provensql Proves SQL Query Equivalence Instead of Guessing
- Read Your Writes: Introducing WAIT FOR in PostgreSQL 19
- LatticeDB: SQLite-Like Solution for Graph Databases
- Admission Control in Browser-Native Media Processing
- Benchmarking RocksDB’s Three Amplification Factors
- Creating Teensy ELF Executables for Linux
» Top scored
- Bonsai: A Distortion-Free Tree Method for Single-Cell Data Visualization
- Provensql Proves SQL Query Equivalence Instead of Guessing
- SIMURG: Real-Time Guard Stops LLM Decoding Corruption Mid-Stream
- Cross-vendor byte-identical inference for a 72B LLM (AMD MI300X vs. Nvidia H100)
- yapyak – An i18n Compiler for React with Automatic Translation
- AMD Helios MI400 System Architecture Unveiled at Hot Chips 2026
- Admission Control in Browser-Native Media Processing
- Oj: Rust Native Replacement for Vite
- Fujitsu Unveils Monaka Data Center CPU at Hot Chips 2026
- Building Production Agentic AI at IBM: Architecture, Decisions, and Lessons