Archive for the 'Custom Silicon Solutions' Category
-
January 15, 2025
Next Up for Custom AI Accelerators: Co-Packaged Optics
By Michael Kanellos, Head of Influencer Relations, Marvell
Computer architects have touted the performance and efficiency gains that can be achieved by replacing copper interconnects with optical technology in servers and processors for decades1.
With AI, it’s finally happening.
Marvell earlier this month announced that it will integrate co-packaged optics (CPO) technology into custom AI accelerators to improve the bandwidth, performance and efficiency of the chips powering AI training clusters and inference servers and opening the door to higher-performing scale-up servers.
The foundation of the offering is the Marvell 6.4Tbps 3D SiPho Engine announced in December 2023 and first demonstrated at OFC in March 2024. The 3D SiPho Engine effectively combines hundreds of components—drivers, transimpedance amplifiers, modulators, etc.—into a chiplet that itself becomes part of the XPU.
With CPO, XPUs will connect directly into an optical scale-up network, transmitting data further, faster, and with less energy per bit. LightCounting estimates that shipments of CPO-enabled ports in servers and other equipment will rise from a nominal number of shipments per year today to over 18 million by 20292.
Additionally, the bandwidth provided by CPO lets system architects think big. Instead of populating data centers with conventional servers containing four or eight XPUs, clouds can shift to systems sporting hundreds or even thousands of CPO-enhanced XPUs spread over multiple racks based around novel architectures—innovative meshes, torus networks—that can slash cost, latency and power. If supercomputers became clusters of standard servers in the 2000s, AI is shifting the pendulum back and turning servers into supercomputers again.
“It enables a huge diversity of parallelism schemes that were not possible with a smaller scale-up network domain,” wrote Dylan Patel of SemiAnalysis in a December article.
-
December 19, 2024
Custom, Copper and Cross-Country Connectivity: Eight Big Trends for Marvell in 2024
By Michael Kanellos, Head of Influencer Relations, Marvell
What happened in semis and accelerated infrastructure in 2024? Here is the recap:
1. Custom Controls the Future
Until relatively recently, computing performance was achieved by increasing transistor density à la Moore’s Law. In the future, it will be achieved through innovative design, and many of those innovative design ideas will come to market first—and mostly— through custom processors tailored to use cases, software environments and performance goals thanks to a convergence of unusual and unstoppable forces1 that quietly began years ago.
- At Marvell AI Day, we announced that custom processors would represent 25% of the AI accelerator market by 2028 and exceed $42 billion. We also announced we had secured four designs with three of the major hyperscalers.
- In July, we unveiled StructeraTM A and Structera X,2 CXL controllers for near memory compute and increasing memory capacity, respectively. Merchant and custom versions of the chips were released at the same time, a reverse of the traditional pattern where custom chips might follow merchant ones by a lengthy lag.
- At OCP, Meta showed off FBNIC, a custom network interface developed in collaboration with Marvell to reduce the time needed to resolve network issues.3 Make that four out of four hyperscalers. See the image below.
FB NIC on display at OFC- In December, we announced a five-year collaboration that includes supplying custom AI products and other devices for AWS.4 The collaboration also encompasses performing semiconductor design tasks in the cloud, which promises to reduce the time and cost of creating new chips.
- Finally, we ended the year with an alliance for custom high bandwidth memory with Micron, Samsung Electronics, and SK hynix.5 Custom HBM can reduce chip real estate by up to 25% and increase memory capacity by up to 33%.6 Custom HBM runs against industry orthodoxy,7 writes analyst Joseph Byrne, but it could pay dividends.
-
December 17, 2024
The Rapid Road Ahead for Custom HBM
By Michael Kanellos, Head of Influencer Relations, Marvell
The idea of customizing high bandwidth memory (HBM) has only recently emerged, but expect to see it in the mainstream in just a few years.
“We strongly believe that custom HBM will be the majority portion of the market towards the ’27-28 time frame,” said In Dong Kim, vice president of product planning at Samsung Semiconductor in a video interview with the Six Five at Marvell Analyst Day earlier this month where Marvell, Micron Technology, Samsung and SK hynix announced a collaboration to accelerate the development of custom HBM solutions.
Sunny Kang, vice president of DRAM technology at SK Hynix had a similar outlook. “Usually in the DRAM industry, when we launch a new product, it takes just one or two years to be mainstream,” he said. “That means along the ‘29 timeframe, it is going to be a mainstream product in the HBM market. I’m pretty sure about that.”
-
June 18, 2024
Custom Compute in the AI Era
This article is the final installment in a series of talks delivered Accelerated Infrastructure for the AI Era, a one-day symposium held by Marvell in April 2024.
AI demands are pushing the limits of semiconductor technology, and hyperscale operators are at the forefront of adoption—they develop and deploy leading-edge technology that increases compute capacity. These large operators seek to optimize performance while simultaneously lowering total cost of ownership (TCO). With billions of dollars on the line, many have turned to custom silicon to meet their TCO and compute performance objectives.
But building a custom compute solution is no small matter. Doing so requires a large IP portfolio, significant R&D scale and decades of experience to create the mix of ingredients that make up custom AI silicon. Today, Marvell is partnering with hyperscale operators to deliver custom compute silicon that’s enabling their AI growth trajectories.Why are hyperscale operators turning to custom compute?
Hyperscale operators have always been focused on maximizing both performance and efficiency, but new demands from AI applications have amplified the pressure. According to Raghib Hussain, president of products and technologies at Marvell, “Every hyperscaler is focused on optimizing every aspect of their platform because the order of magnitude of impact is much, much higher than before. They are not only achieving the highest performance, but also saving billions of dollars.”
With multiple business models in the cloud, including internal apps, infrastructure-as-a-service (IaaS), and software-as-a-service (SaaS)—the latter of which is the fastest-growing market thanks to generative AI—hyperscale operators are constantly seeking ways to improve their total cost of ownership. Custom compute allows them to do just that. Operators are first adopting custom compute platforms for their mass-scale internal applications, such as search and their own SaaS applications. Next up for greater custom adoption will be third-party SaaS and IaaS, where the operator offers their own custom compute as an alternative to merchant options.
Progression of custom silicon adoption in hyperscale data centers.
-
June 12, 2023
AI and the Tectonic Shift Coming to Data Infrastructure
By Michael Kanellos, Head of Influencer Relations, Marvell
AI’s growth is unprecedented from any angle you look at it. The size of large training models is growing 10x per year. ChatGPT’s 173 million plus users are turning to the website an estimated 60 million times a day (compared to zero the year before.). And daily, people are coming up with new applications and use cases.
As a result, cloud service providers and others will have to transform their infrastructures in similarly dramatic ways to keep up, says Chris Koopmans, Chief Operations Officer at Marvell in conversation with Futurum’s Daniel Newman during the Six Five Summit on June 8, 2023.
“We are at the beginning of at least a decade-long trend and a tectonic shift in how data centers are architected and how data centers are built,” he said.
The transformation is already underway. AI training, and a growing percentage of cloud-based inference, has already shifted from running on two-socket servers based around general processors to systems containing eight more GPUs or TPUs optimized to solve a smaller set of problems more quickly and efficiently.
