XPU.news

Meta Has a Lot Riding on the MTIA To read the full post, click through. Once at http://xpu.news, you can also comment and submit your own posts to appear in this feed.

For some reason, recent Meta news reminds me of two songs:

The Llama Song (If you had two small children as I did in 2007, you know it.) Once upon a time, Meta had a popular LLM called Llama. Remember those old days?

(Riding on the) Metro by Berlin (80s teen here). MTIA sounds like the name of a …

The important things are he's getting people to stop wearing brown shoes with navy pants and those horrors with the white soles.

Third-Gen Ceva PentaG Targets Satcom and the IoT To read the full post, click through. Once at http://xpu.news, you can also comment and submit your own posts to appear in this feed.

Ceva has begun licensing its third-generation PentaG intellectual property. Available in two configurations, the updated PentaG enables customers without cellular expertise to add 5G connectivity to their designs. The PentaG-NTN works with non-terrestrial networks, and the PentaG-Edge targets custo…

HyperAccel Promises Power Efficiency Gains To read the full post, click through. Once at http://xpu.news, you can also comment and submit your own posts to appear in this feed.

HyperAccel seeks to displace Nvidia by delivering a lower-cost AI accelerator. Its Bertha 500 promises to double performance while raising power efficiency 12× compared with a Hopper-generation Nvidia GPU. Based on the company’s latency-processing unit (LPU) architecture, the startup’s NPU achieves…

New Funding and Chip Whisk SambaNova to its Next Stage To read the full post, click through. Once at http://xpu.news, you can also comment and submit your own posts to appear in this feed.

Hoary startup SambaNova has closed a $350 million Series E funding round, introduced the SN50 AI accelerator, and signed up SoftBank to deploy the new NPU. Following reports that Intel was considering acquiring the company, which Intel CEO Lip-Bu Tan chairs, the beleaguered giant merely joined the …

Disgraced Arm China CEO Wu to Lower Custom NPU Development with RISC-V To read the full post, click through. Once at http://xpu.news, you can also comment and submit your own posts to appear in this feed.

Allen Wu, founder of CoreLab Technology and the disgraced former Arm China head, aims to break the $100 million cost barrier for custom AI chips. He envisions AI processors as heterogeneous systems combining CPUs, GPUs, and domain-specific accelerators, tailored to workloads. CoreLab’s approach red…

US govt is considering capping the number of (AMD MI325 + Nvidia H200) GPUs to 75k per customer. I'm sure there's no way to get around this brilliant scheme.

Super Cores and Chiplets Define new Apple M5 Processors To read the full post, click through. Once at http://xpu.news, you can also comment and submit your own posts to appear in this feed.

Apple has adopted chiplets for its new M5 and M5 Max PC processors, an approach that it now shares with AMD and Intel. The company has also changed its CPU-core configuration, introducing new super cores and dropping efficiency cores. The highest-spec CPU configuration now features 6 super and 12 p…

Wimpy Cores Fight Back To read the full post, click through. Once at http://xpu.news, you can also comment and submit your own posts to appear in this feed.

As MWC starts, Intel has stated that mobile operators are deploying Clearwater Forest systems. Officially called Xeon 6+ with (only) E-Cores, Clearwater is a 288-core server processor. Fabbed in the company's 18A process and using the tightly packed efficiency cores, it should deliver leading compu…

US Faces Unintended Consequences of China Trade Policy To read the full post, click through. Once at http://xpu.news, you can also comment and submit your own posts to appear in this feed.

China's ambition to become technologically independent became an essential national project when the West cut it off from key technologies.

The TOI is reporting that this has extended to CPU design, discussing how Intel CEO LBT found Huawei has hired a hundred top-notch architects. Intel CEO Lip-B…

Nvidia Discloses Rubin CPX, an Unexpected Data-Center GPU To read the full post, click through. Once at http://xpu.news, you can also comment and submit your own posts to appear in this feed.

Nvidia has unveiled the Rubin CPX, an AI chip for data centers that was not previously on their public roadmaps. It can offload LLMs’ decode stages.

Nvidia to Resolve Dilemma To read the full post, click through to http://xpu.news, where you can also comment and submit your own posts to appear in this feed.

Reuters is reporting that The Wall Street Journal is reporting that Nvidia is developing a new chip to accelerate AI inference, presumably meaning to slash chatbot latency. Set to be announced at GTC San Jose next month, the chip will incorporte technology from Groq, which Nvidia recently "licensed…

Amazon Doubles Down on OpenAI Click through to comment and submit posts to appear in this feed.

OpenAI and AWS are expanding their existing $38 billion multi-year agreement by $100 billion over 8 years. The expansion includes OpenAI committing to consume approximately 2 gigawatts of Trainium capacity. Amazon will also invest $50 billion in OpenAI, starting with an initial $15 billion investmen

But I don't want to get autism when I'm 80.

If a developer says:
I’ll fix it in an hour.

Believe him.

No need to check in every 3 hours.

The estate would be better off suing Marimekko for getting its name associated with Mekko charts when clearly they take after Mondrian's work.

A wise friend of mine said a startup should have no more than one prima donna.

File this under "what could possibly go wrong."

Intel Funds SambaNova Following reports that Intel would acquire AI-chip company SambaNova, chaired by Intel CEO Lip-Bu Tan, who had invested in the startup, the beleagured chipmaker is joining SambaNova's next round and the two companies will develop a cloud service to be run by SambaNova. A stated reason for the tie up is to participate in heterogenous data centers. Presumably these would use Intel GPUs (should it make a competitive one) for some functions (e.g., training) and SambaNova NPUs for others (e.g., inference, KV cache prefill). A further reason is that, despite SambaNova's limited market traction and other startups flaming out, investor interest remains high for an Nvidia alternative. It could've simply been too expensive for Intel to outright acquire SambaNova. The potential acquisition raised questions about LBT self dealing. He recused himself from recent Intel-SambaNova negotiations, but the ultimate deal does little to ameliorate these concerns.

Click through to comment and submit posts to appear in this feed.

Click through to comment and submit posts to appear in this feed.

AI-Chip Startup MatX Nabs $500 Million In the third big AI-chip announcement today, MatX has announced its Series B. Coming a year after its Series A, MatX has raised $500 million. Yeah, that's going to leave a mark if the company doesn't pan out. Founded by ex-Googlers, MatX disclosed it's first chip, the MatX One, will employ both SRAM and HBM (assuming they can find some to buy). They promise the One will have leading LLM throughput and latency. There's no shortage of companies pursuing this market. Regardless of their pedigree, MatX founders must deliver their promised performance and cost breakthrough to escape the also-ran peloton.

Click through to comment and submit posts to appear in this feed.

Adopting Grace, Meta is Nvidia’s First Major Server-Processor Customer Meta will deploy Grace-based servers in Nvidia’s first CPU-only deployment, setting a course for a Vera-only deployment and extending Arm’s data-center success.

Click through to comment and submit posts to appear in this feed.

Renesas Adds Dual-Band Wi-Fi and Bluetooth Low Energy to MCUs Renesas’s first dual-band Wi-Fi 6 MCUs feature an Arm Cortex-M33, ultra-low power modes, and pre-integrated modules for battery-powered IoT devices.

Click through to comment and submit posts to appear in this feed.

On the other hand, we have AI slop filing bug reports.

Will Tsavorite’s Composable AI Chiplets Be a GEMM Gem? Tsavorite debuts its OPU (NPU) with $100M in orders, featuring Cuda support, Arm cores, and modular chiplet scalability.

Click through to comment and submit posts to appear in this feed.

Cadence ChipStack AI Super Agent Cuts Design Time Cadence has released the ChipStack AI Super Agent, a productivity-boosting automated workflow for front-end chip design and verification.

Click through to comment and submit posts to appear in this feed.

Neurophos Taps Metamaterials for AI Revolution Neurophos has raised $110 million to develop exaflop-scale photonic AI chips that use tunable metamaterials to overtop the "power wall" of traditional GPUs.

Click through to comment and submit posts to appear in this feed.

Byrne-Wheeler Report, Episode 9: Neurophos Scales Photonic Computing Legendary semiconductor analysts Bob and Joe discuss Neurophos, a photonic computing startup that recently received funding. #photonics, photonic computing, #ai accelerator (NPU), #semiconductor, #semiconductors, #datacenter

Subscribe to the Byrne-Wheeler Report on YouTube for more content like this.

Byrne-Wheeler Report, Episode 9: Intel 4Q25 Earnings Reveal Capacity and Yield Issues Legendary semiconductor analysts Bob and Joe discuss Intel's fourth quarter earnings, focusing on capacity constraints and yield improvements. They then discuss the Panther Lake launch and 18A and 14A process technology. #intel #semiconductors #semiconductor #pc #microprocessor

Subscribe to the Byrne-Wheeler Report on YouTube for more content like this.

Byrne-Wheeler Report, Episode 9: Microsoft Maia 200 Promises Power Efficiency Legendary semiconductor analysts Bob and Joe discuss the Microsoft Maia NPU. Targeting AI inference, it promises performance to rival the Nvidia Blackwell at much less power. #Microsoft Maia 200, #ai accelerator (NPU), #semiconductor, #semiconductors, #datacenter

Subscribe to the Byrne-Wheeler Report on YouTube for more content like this.