English
AMD has officially released the Instinct MI350P, a new PCIe accelerator tailored for enterprise users seeking on-premises AI inference without overhauls to their existing data center infrastructure. Featuring a dual-slot, full-height and full-length form factor, this graphics card is fully compatible with conventional air-cooled servers. It also marks AMD’s first release of a current-generation Instinct chip designed for standard server slots in nearly four years.
AMD Instinct MI350P
AMD’s PCIe-based Instinct product line remained stagnant after the launch of the MI210 in early 2022. All subsequent generations, including the MI300X, MI325X and OAM-format MI350X, adopted OAM socketed modules mounted on dedicated universal baseboards. These modules require customized enclosures with robust power delivery and airflow to support up to eight 1,000W-class accelerators in a single tray. Such hardware architecture suits hyperscale cloud providers that purchase GPU racks in bulk, yet it fails to accommodate regular enterprises unwilling or unable to deploy bespoke AI racks for on-site inference tasks. The MI350P precisely fills this market gap. Currently, NVIDIA lacks a high-end server-grade PCIe competitor in this segment, leaving AMD with temporary market exclusivity.
Hardware Comparison: MI350P versus MI350X OAM
The MI350P is not a cut-down variant of the MI350X; AMD engineered an independent streamlined chip for this new model. The MI350X integrates two I/O dies paired with eight accelerator complex dies (XCDs), delivering 256 compute units in total. In contrast, the MI350P contains one I/O die and four XCDs, equating to 128 compute units. Despite halving the silicon scale, it maintains an identical 2.2 GHz peak clock frequency as its higher-tier counterpart. The memory configuration follows the same downgraded specification: four HBM3E stacks (versus eight), a 4,096-bit memory bus (down from 8,192-bit), alongside 144GB memory capacity and 4 TB/s bandwidth, compared to the MI350X’s 288GB and 8 TB/s throughput.
AMD Instinct MI350P architecture
The peak computing throughput is also reduced by half. The MI350P reaches 4,600 MXFP4 TFLOPS versus the MI350X’s 9.2 PFLOPS, along with 2,300 FP8 TFLOPS compared to the premium model’s 4.6 PFLOPS. Performance metrics for BF16, FP16 and other precision standards follow the same proportional decline. Notably, AMD has published both peak and real-world sustained performance data for transparency. The card delivers 2,299 TFLOPS under MXFP4, 1,529 TFLOPS under FP8, and 713 TFLOPS under BF16. These practical figures reflect real output within a 600W power envelope, where power constraints and memory bandwidth limitations inevitably lower theoretical peak performance.
The editorial team previously evaluated the MI350X platform via Supermicro’s Jumpstart program and recognized its robust inference workload capabilities. The team is eager to conduct hands-on testing of the MI350P, analyzing how this PCIe-form-factor accelerator performs within standard commodity server chassis.
The MI350P does not feature a 50% power reduction despite its halved silicon scale. It carries a 600W TBP power rating, equivalent to roughly 60% of the MI350X’s 1000W limit. This peak wattage hits the upper boundary of the PCIe CEM specification, running the card at the slot’s maximum thermal threshold. A reduced 450W operating mode is available for servers with insufficient cooling, accompanied by moderate performance cuts. Positioned in the same power bracket, the MI350P directly competes with NVIDIA’s H200 NVL and RTX Pro 6000 Server for enterprise procurement.
Unlike NVIDIA’s H200 with NVL4 high-speed interconnects, the MI350P disables Infinity Fabric exposure. All inter-GPU data transmission is limited to the 128 GB/s bandwidth of PCIe Gen5 x16.
Eight-GPU Air-Cooled Deployment
As a standard dual-slot FHFL PCIe card, the MI350P is compatible with existing enterprise servers. Major OEMs offer dense eight-GPU air-cooled models, including the previously reviewed Dell PowerEdge XE7740 and HPE ProLiant DL380a Gen12. Optimized for 600W accelerators, these platforms require no custom racks, liquid cooling or OAM baseboards.
An eight-card MI350P configuration delivers 1,152GB HBM3E and 32 TB/s aggregate bandwidth, sufficient to host trillion-parameter models in MXFP4 precision within one air-cooled chassis. Nevertheless, it sacrifices dedicated scaling fabrics. While the MI350X utilizes Infinity Fabric for fast inter-module communication, the MI350P relies solely on PCIe Gen5. This architecture suits node-local tensor parallelism and cross-node data parallelism for inference, whereas the OAM-based MI350X remains superior for bandwidth-intensive AI training tasks.
Precision Formats
The MI350P inherits all precision formats from the MI350X without upgrades. OCP block-scaling types including MXFP4, MXFP6 and MXFP8 have become mainstream for AI model development, enabling low-precision training with negligible quality degradation. MXFP4 delivers over double the speed of FP8 and four times the throughput of BF16. Real-world cases such as OpenAI gpt-oss and Kimi K2.6 prove the value of native low-bit quantization. Since MXFP4 and INT4 weights occupy only 25% of BF16 memory space, trillion-parameter models can be deployed inside a single eight-GPU server, eliminating cumbersome multi-node clusters for on-prem enterprises.
Bottom Line
Most on-prem AI enterprises are constrained by power, cooling, density and budget rather than raw computing capability. The drop-in MI350P effectively alleviates these deployment barriers. With NVIDIA lacking a flagship PCIe server GPU in this segment, AMD retains a clear competitive advantage for the time being. Further details are available on AMD’s official Instinct webpage.
Beijing Qianxing Jietong Technology Co., Ltd.
Sandy Yang/Global Strategy Director
WhatsApp / WeChat: +86 13426366826
Email: yangyd@qianxingdata.com
Website: www.qianxingdata.com/www.storagesserver.com
Business Focus:
ICT Product Distribution/System Integration & Services/Infrastructure Solutions
With 20+ years of IT distribution experience, we partner with leading global brands to deliver reliable products and professional services.
“Using Technology to Build an Intelligent World”Your Trusted ICT Product Service Provider!
Sandy Yang/Global Strategy Director
WhatsApp / WeChat: +86 13426366826
Email: yangyd@qianxingdata.com
Website: www.qianxingdata.com/www.storagesserver.com
Business Focus:
ICT Product Distribution/System Integration & Services/Infrastructure Solutions
With 20+ years of IT distribution experience, we partner with leading global brands to deliver reliable products and professional services.
“Using Technology to Build an Intelligent World”Your Trusted ICT Product Service Provider!