To that finish, AMD has launched the MI325X with higher reminiscence capability and bandwidth than the Intuition MI300X, which launched final December. The MI325X relies on the identical CDNA 3 GPU structure, in contrast with 192GB of HBM3 high-bandwidth reminiscence and 5.3 TB/s in reminiscence bandwidth within the MI300X.
AMD mentioned AI inference efficiency within the MI325X offers 40% quicker throughput with an 8-group, 7-billion-parameter Mixtral mannequin over Nvidia’s top-of-the-line Hopper H200, 30% decrease latency with a 7-billion-parameter Mixtral mannequin, and 20% decrease latency with a 70-billion-parameter Llama 3.1 mannequin.
AMD is planning an eight-node platform for subsequent yr, much like Nvidia’s DGX Pods. With eight MI325X GPUs linked over AMD’s Infinity Material, the platform will supply 2TB of HBM3e reminiscence, 48 TB/s of whole reminiscence bandwidth, 20.8 petaflops of FP8 efficiency, and 10.4 petaflops of FP16 efficiency, AMD mentioned.
The MI325X will start transport in programs from Dell Applied sciences, Lenovo, Supermicro, Hewlett Packard Enterprise, Gigabyte, and a number of other different server distributors beginning within the first quarter of subsequent yr, the corporate mentioned.
Learn extra processor information
- Enfabrica appears to speed up GPU communication: Enfabrica’s Accelerated Compute Material SuperNIC (ACF-S) silicon is designed to ship greater bandwidth, higher resiliency, decrease latency and higher programmatic management to information heart operators working data-intensive AI and HPC.
- Nvidia claims effectivity positive aspects of as much as 100,000X: Nonetheless, the chipmaker’s dramatic declare for the efficiency positive aspects of its GPUs is over a 10-year span, and solely applies to at least one sort of calculation.
- Intel launches Xeon 6 processors and Gaudi 3 AI accelerators: Intel has formally launched its subsequent Xeon 6 server processors in addition to the Gaudi 3 AI accelerators, making some fairly huge boasts within the course of.
- Inflection AI shifts to Intel Gaudi 3, difficult Nvidia’s AI chip lead: The announcement follows IBM’s current partnership with Intel, signaling a rising curiosity in Intel’s AI {hardware}.
- Intel’s Altera spinout launches FPGA merchandise, software program: Altera CEO Sandra Rivera shares ‘huge, audacious, formidable purpose’ to dominate FPGA market.
