InfraMeta will build Petal: the first cable across an ocean with a petabit per second of capacityThe cable will link France and the United States over about 7,000 kilometres and carry 1 petabit per second, per Meta double the most advanced cables at that distance. The trick is fibre with two cores instead of one. Service is expected in 2029.InfraVera Rubin entered MLPerf at up to 3.7 times the previous generation, for now as a preview resultIn MLPerf Inference v6.1 NVIDIA submitted the first Vera Rubin NVL72 results against GB300 NVL72: up to 3.7 times higher throughput on Qwen3-VL and up to 2.5 times on DeepSeek-R1. It is a shared test with rules, but the system sits in the category for upcoming products.InfraGoogle, NVIDIA and Emerald AI form an alliance for data centres that can cut their own power when the grid is strainedThe AI Energy Management Alliance wants flexible data centres to connect to the grid faster in exchange for a verifiable promise to give power back. A day earlier NVIDIA described what that looks like live: from 4 to 3 megawatts, automatically.InfraSaudi HUMAIN is now selling compute on AMD chips, and the next step is 250 megawattsOn 31 August AMD, Cisco and HUMAIN announced that AMD Instinct MI355X systems are live in Saudi Arabia and serving customers. The next phase is a plan: up to 250 megawatts on the new MI400 generation from 2027. Their joint venture's target remains up to 1 gigawatt by 2030.InfraCloudflare rearranged one record in memory and got 100 terabytes back, and the cache got fasterThe platform behind 1.1.1.1 holds over 250 billion cache entries at any moment. Five successive changes to how they are laid out cut each entry by more than half. The memory freed is roughly the entire RAM of 130 of their servers.InfraNVIDIA moved the memory controller off the chip and put it inside the memory itselfOn 26 August NVIDIA added NVHBM to NVLink Fusion - memory whose base die now carries their own controller. The claim is up to 30 per cent more bandwidth, 15 per cent less power and up to a quarter of the compute die freed up. Amazon's Annapurna Labs goes first.InfraMeta unveils MTIA 300, its first chip built to train modelsMTIA 300 is Meta's first accelerator built to train models, not just serve answers from finished ones. It targets recommendation and ranking systems, with the network built straight into the silicon. By Meta's own numbers, total communication time drops 3.9x versus an equivalent GPU cluster.InfraMeta wrote its own AI network protocol and opened it through the Open Compute ProjectMetaRoCE is a transport protocol written from scratch for the conversations between thousands of AI accelerators over plain Ethernet. By the company's own numbers, it keeps about 86 percent of throughput even at one percent packet loss. The spec, the reference code and the compliance test are going public.
42 stories