Intel, AMD in x86 alliance; NoC tiling; optical connectivity funding; LLMs’ reasoning

A quick foundry update. While TSMC posted excellent results for the third quarter of 2024 (year over year, revenue increased 39.0% while net income and diluted EPS both increased 54.2%), Samsung Foundry is reportedly struggling with the lack of major customers for its upcoming Taylor, Texas, fab. According to Reuters, this would be the reason why Samsung has postponed taking deliveries of ASML chipmaking equipment for the new factory.

Intel and AMD form the “x86 Ecosystem Advisory Group”

Archrivals Intel and AMD are collaborating for the sake of x86 architecture’s competitiveness. The two companies have created an “x86 ecosystem advisory group”, whose founding members also include Broadcom, Dell, Google Cloud, Hewlett Packard Enterprise, HP, Lenovo, Microsoft, Oracle, and Red Hat. The intended outcomes from this initiative will include enhancing customer choice and compatibility across hardware and software, while accelerating their ability to benefit from new, cutting-edge features; simplifying architectural guidelines to enhance software consistency and standardize interfaces across x86 product offerings from Intel and AMD; enabling greater and more efficient integration of new capabilities into operating systems, frameworks and applications. According to Reuters, the creation of the x86 advisory group is a reaction to growing competition from Arm, and one of its key goals will be to ensure that software can run on both Intel and AMD x86 processors without any modifications.

Arteris’ NoC tiling enables tile-based SoC architecture

Described as “an innovative evolution of its network-on-chip IP products” based on its “proven, robust network-on-chip IP”, the tiling capabilities announced by Arteris promise to pave the way to tile-based architectures for the SoCs that include AI acceleration blocks. According to Arteris, by allowing the replication of the same tile, tile-based architectures simplify design and verification, make it easy to scale up AI performance, and enable a significant power reduction by turning off individual tiles. The new tiling capability, along with mesh topology, is being offered by Arteris’ FlexNoC and Ncore NoC IP products.

Risc-V updates, Andes, Codasip, SiFive

Andes Technology has announced the AndesCore AX66 out-of-order superscalar multicore processor IP supporting the RVA23 profile. The AX66 is the second member of the high-performance out-of-order AX60 series. It features 13-stage pipeline, 4-wide decode, and 8-wide out-of-order execution as the previously announced AX65. In addition to that, the AX66 introduces new functionalities, including Vector and Vector Crypto support, Hypervisor and AIA, Multi-Cluster support with CHI, and RVA23 profile support. AX66 targets high-performance Linux and Android applications such as edge/data center AI, infotainment, networking, and vision/camera applications.

Codasip has introduced its new L730 core, a high-performance Risc-V processor IP that meets automotive safety and security needs enabling ISO/SAE 21434 and ISO 26262 compliance up to the ASIL D integrity level. The L730 is available in a range of pre-verified, baseline configurations that can be further configured and customized through what Codasip calls “bounded customization”. The core supports the Risc-V Scalar Crypto Extension and will offer the CHERI memory safety technology.

Arteris and SiFive have jointly developed a pre-verified solution combining SiFive P870-D CPU and Arteris Ncore cache coherent interconnect IP. According to the companies, the pre-verified solution enables faster, low-risk development cycles for datacenter SoCs with high performance AI workload and power efficiency requirements.

Optical connectivity funding updates: Lightmatter, Xscape, Infinera

Lightmatter has raised a $400 million Series D financing, valuing the company at $4.4 billion and bringing the total capital raised to date to $850 million. With this financing, Lightmatter will prepare its Passage photonic engine for mass deployment in partner data centers.

Startup Xscape Photonics has announced the close of a $44 million Series A funding round, increasing the total amount raised to $57 million. The funding will be used to accelerate the development of Xscape Photonics’ ChromX platform, a scalable, multi-color, programmable photonics platform for AI data center fabrics.

Infinera – a supplier of optical networking solutions and optical semiconductors – and the U.S. Department of Commerce have signed a non-binding preliminary memorandum of terms for Infinera to receive up to $93 million in direct funding as part of the CHIPS and Science Act. This funding would support the expansion and modernization of both Infinera’s semiconductor capabilities in Silicon Valley, California and its test and packaging capabilities in Lehigh Valley, Pennsylvania.

New Easy-PC release

UK-based Number One Systems has released the latest update of its Easy-PC software for PCB design. The new Version 28 features several innovations on areas such as Application Enhancements, Library Management and Creation, Designing and Editing, Design Checking and Integrity, and Plotting and Manufacturing.

Acquisitions

Schneider Electric has signed an agreement to acquire a controlling interest in Motivair (Buffalo, NY), a company specializing in liquid cooling and advanced thermal management solutions for high performance computing systems. Under the terms of the transaction, Schneider Electric will acquire an initial 75% controlling interest in the equity of Motivair for an all-cash consideration of $850 million. The Group expects to acquire the remaining 25% of noncontrolling interests in 2028.

Further reading: state-of-the-art LLMs falling into trap questions

A team of Apple researchers has conducted a large-scale study on several state-of-the-art Large Language Models, investigating the fragility of their mathematical reasoning. The researchers hypothesize that current LLMs are not capable of genuine logical reasoning; instead, they attempt to replicate the reasoning steps observed in their training data. One of the goals of this work was the introduction of a better benchmark to assess the mathematical reasoning of models, but what is particularly interesting is the evidence of powerful LLMs falling into simple trap questions. See the example below, which tests the capabilities of “o1 preview”, a new LLM which belongs to what OpenAI describes as “a new series of reasoning models for solving hard problems”. As the Apple researchers observe, “the model blindly applies the inflation rate, even though the inflation amount is irrelevant as the question clearly indicates the given prices are for ‘now’ and not last year.”

Source: Apple

This entry was posted on Tuesday, October 22nd, 2024 at 9:47 am. You can follow any responses to this entry through the RSS 2.0 feed. You can leave a response, or trackback from your own site.