AMD MI350P PCIe GPU cards unveiled

AMD unveiled its Instinct MI350P PCIe cards on May 7, positioning the product as an AI inference accelerator for operators that want to add GPUs to existing data-center servers rather than deploy new liquid-cooled platforms. The company’s product page says the card is designed to “deploy and scale generative and agentic AI within your existing infrastructure,” and AMD’s launch blog says it is built for standard air-cooled servers. (amd.com 1) (amd.com 2) The new card extends AMD’s MI350 lineup with a PCIe form factor aimed at enterprise and on-premises deployments. AMD’s materials frame the product around inference, retrieval-augmented generation pipelines and installations that fit current rack power and cooling limits. (amd.com) AMD did not publish pricing or a shipment date on the product page, brochure or launch blog reviewed by Reuters. The company instead emphasized the card’s fit with existing infrastructure and its ROCm software stack. ### What exactly did AMD introduce? (amd.com) AMD’s product page identifies the new offering as the “AMD Instinct MI350P PCIe card,” a member of the Instinct MI350 series. The page says the card is intended to provide AI performance “within your existing infrastructure,” distinguishing it from larger accelerator platforms in the same family. (amd.com) The May 7 AMD blog says the MI350P is a “dual-slot drop-in” card for standard air-cooled servers. AMD said the PCIe form factor is meant for enterprises that need more AI computing power than CPUs can provide but are “not ready to invest in dedicated GPU accelerator platforms.” (amd.com) ### Why is AMD stressing air-cooled, standard-server deployment? AMD’s brochure says the MI350P is a full-height, full-length, dual-slot PCIe card designed for mainstream air-cooled servers. The brochure adds that the card is meant to fit without “specialized cooling,” rack redesigns or building new systems from scratch. (amd.com) The launch blog says the card is built to deploy inference on premises within a customer’s current power, cooling and rack infrastructure. AMD also says systems can be configured with up to eight MI350P cards, giving server makers and operators a way to scale within familiar PCIe server designs. (amd.com) ### What performance and memory numbers did AMD publish? AMD’s brochure lists estimated peak performance of 4,600 TFLOPS at MXFP4 and MXFP6, alongside 2,300 TFLOPS-class figures for several 8-bit and 16-bit matrix formats. The same document says the card includes 144GB of HBM3E memory. (amd.com) AMD’s May 7 blog repeats the headline figures, saying the card delivers estimated 2,299 TFLOPS and up to 4,600 peak TFLOPS at MXFP4, plus 144GB of HBM3E memory running at up to 4TB/s. The company described that as the highest performance currently available in an enterprise PCIe card, though that characterization is AMD’s own. (amd.com) ### Which workloads is AMD targeting with the MI350P? AMD’s launch blog says the MI350P is “ideal for small, medium and large AI models for inference and RAG pipelines.” The broader MI350 series page says PCIe cards are for “mainstream enterprise,” while MI350X and MI355X platforms are aimed at larger-scale training and inference workloads. (amd.com) The distinction matters because AMD is not presenting the MI350P as a replacement for its highest-density accelerator platforms. Instead, the company’s own materials place it alongside those systems as a lower-friction option for enterprises that want AI capacity inside current server footprints. (amd.com) ### What did AMD leave unsaid? AMD’s public materials reviewed on May 14 do not include a list price, customer availability date or named launch customers for the MI350P. The product page, PDF brochure and launch blog focus on specifications, deployment format and software support. (amd.com) AMD’s next public step is likely to come through updates to its Instinct product pages, partner-system listings or additional launch materials. As of May 14, 2026, the company’s MI350P product page and brochure remained live with specifications but without pricing or shipment timing. (amd.com) (amd.com)