Amazon launches Trainium 3, Marvell ‘ends up big loser,’ SemiAnalysis says

SemiAnalysis said in a newsletter that, “Hot on the heels of our 10K word deep dive on TPUs, Amazon (AMZN) launched Trainium3 general availability and announced Trainium4 at its annual AWS re:Invent. Amazon has had the longest and broadest history of custom silicon in the datacenter. While they were behind in AI for quite some time, they are rapidly progressing to be competitive… With Trainium3, AWS remains laser-focused on optimizing performance per total cost of ownership. Their hardware North Star is simple: deliver the fastest time to market at the lowest TCO. Rather than committing to any single architectural design, AWS maximizes operational flexibility. This extends from their work with multiple partners on the custom silicon side to the management of their own supply chain to multi-sourcing multiple component vendors… Marvell (MRVL) ends up being the big loser from this. While they designed Trainium2, they lost the design bakeoff with Alchip for this generation. Marvell’s Trainium3 version was a chiplet based design, with the I/O being put onto a separate chiplet, instead of on a monolithic die with the compute as is the case with Trainium2 and what will be Trainium3. Marvell lost this socket due to poor execution for Trainium2. The development timeline took far too long. Marvell had problems designing the RDL interposer for the package as well, and Alchip had to step in to help deliver something workable.”