AMD officially released the new generation of accelerated GPU Instinct MI300X/MI300A at the end of last year. The first batch was shipped to large customers such as supercomputing and data centers, including the U.S. government, Microsoft, and Meta. Now some small customers have also received the goods.LaminiAI, an emerging AI startup that has always had a very good relationship with AMD, revealed its batch arrival of MI300X, which will be used to train its own LLM large language model.
What they use isEight-way parallel configuration,There are eight such accelerators in a single system. You can see eight cards from the screenshot, but they are all idle and not at full power.
As early as July last year, Lawrence Livermore National Laboratory (LLNA), a subsidiary of the U.S. Department of Energy, announced that the new "El Capitan" (El Capitan) supercomputer has begun to install the MI300A accelerator and will be fully online in 2024. The performance is expected to exceed 20 billion floating point calculations per second.
Now, small customers have also received the MI300 accelerator. Obviously, the production capacity has been increased and can be supplied in batches, which is very critical for AMD.
AMD had previously expected thatThe MI300 series of accelerators will be the fastest in the company's history to reach $100 million in revenue.
AMDMI300X integrates eight 5nmXCD accelerated computing modules, with a total of 304 computing units, four 6nm IOD modules, integrated 256MB unlimited cache, and eight HBM3 memories with a total of 192GB, totaling 153 billion transistors.
According to official statements, the performance of MI300X can exceed NVIDIAH10080GB.
MI300A is the world's first APU accelerator for AI and HPC. It is equipped with six XCD modules, 228 computing units, three CCD ohms, 24 Zen4 CPU cores, and 128GB HBM3 memory, with a total of 146 billion transistors.