Meta has developed an AI accelerator chip called MTIA, following in the footsteps of Google and Amazon. Employing RISC V CPUs and fixed-function units, the chip offered 102 INT8 TOPS at 25 W TDP.

Lightelligence’s ONoc interposer connects chiplets optically through silicon waveguides on a monolithic multireticle substrate. The company launched a machine-learning application as a demonstrator.

The MLPerf benchmark added tests to assess GPT-training performance. AI processors from Nvidia and Intel (Habana) both exhibit good scaling.

Fifth Generation Mali/Immortalis GPU Architecture Defers Vertex Shading

Updated Arm v9.2 Cluster Interconnect Adds CPUs, Doubles L3 Cache

South Korean AI startup Rebellions has launched Atom, an AI-accelerator card targeting the high-frequency trading market. The company has raised over $100 million and has released ResNet-50 MLPerf benchmark performance data.

Imagination Technologies’ new licensable CXM GPU family brings high-dynamic-range (HDR) support to its mainstream offering. With three new models, the company claims high area efficiency.

The company has streamlined its data-center AI roadmap, creating a new version of Falcon Shores that combines the Xe GPU architecture with some aspects of Habana’s AI design, which is delayed to 2025.

Moffett, a startup headquartered in Shenzen, China, has released AI accelerators that address sparsity to increase performance 32x. The company offers three cards spanning a 70–250 W TDP range.

The Arm-compatible AmpereOne server processor implements 192 single-thread cores of Ampere’s own design in a chiplet-based package.

TechInsights’ floorplan analysis of desktop-PC processors quantifies the size difference among AMD and Intel CPUs.

Nanosheet transistors are superior to FinFETs for logic transistors, offering an adjustable gate width, superior electrostatic characteristics, and better channel-thickness uniformity as well as faster switching speed. They have drawbacks for SRAM and I/O transistors, however.