MI250X 128GB

The MI250X 128GB is a CDNA 2-generation AMD GPU built on the Aldebaran architecture, manufactured on a TSMC 6nm process node with 58 billion transistors. Released in 2021, it delivers 383 TFLOPS of FP16 throughput and 383 TFLOPS BF16 — the two precision formats most commonly used for transformer model training and inference. The Aldebaran architecture represents AMD's approach to balancing compute throughput, memory bandwidth, and power efficiency for data center AI workloads. At 560W TDP, the MI250X 128GB sits in the high-power data center tier (560W), compatible with standard high-density GPU server configurations.

Memory capacity is 128GB of HBM2e with 3277 GB/s bandwidth. This determines which models can run without quantization: approximately 64B parameters at FP16 (2 bytes/param), 128B at INT8 (1 byte/param), or up to 256B parameters at INT4/GGUF quantization (0.5 bytes/param). These figures are theoretical maximums — actual capacity is reduced by KV cache, framework overhead, and activation memory, typically by 10–20% for inference and 30–40% for training. The arithmetic intensity ceiling is approximately 117 FLOP/byte (383 TFLOPS ÷ 3277 GB/s). Most autoregressive LLM inference falls well below this threshold, making the 3277 GB/s memory bandwidth the binding constraint on tokens-per-second throughput rather than raw TFLOPS.

The MI250X 128GB uses Infinity Fabric / PCIe 4.0 for host connectivity. Without NVLink, VRAM cannot be pooled across multiple cards — the single-card 128GB capacity is the hard ceiling for model size without model sharding over slower PCIe. For workloads that exceed 128GB, the alternative is pipeline parallelism (splitting model layers across cards) rather than tensor parallelism, which introduces inter-card communication overhead at each layer boundary. This makes the MI250X 128GB best suited for workloads that fit within a single card's VRAM budget.

The primary workloads for the MI250X 128GB are Large model training, Memory-bound HPC, AMD ROCm workloads. Dual-die AMD CDNA 2 GPU. 128GB HBM2e across two dies. Predecessor to MI300X. Available on AWS (p4de) and select cloud providers at competitive pricing. Key limitations to factor into your evaluation: Dual-die design — software must handle NUMA topology; ROCm ecosystem lags CUDA maturity; Lower FP16 throughput than NVIDIA H100 at similar price points. When comparing this GPU against alternatives at similar price points, the most important metrics are memory bandwidth (for inference throughput), VRAM capacity (for model size), and FP16/BF16 TFLOPS (for training speed). Raw TFLOPS figures can be misleading for inference — a GPU with lower TFLOPS but higher memory bandwidth will often outperform a higher-TFLOPS card on tokens-per-second for memory-bound autoregressive generation.

In the broader GPU market, the MI250X 128GB is a 5-year-old design that is increasingly being replaced by newer architectures in cloud deployments, though it remains available at competitive rental rates. Software compatibility is a key consideration: AMD's ROCm stack supports PyTorch, TensorFlow, and JAX, but the ecosystem is less mature than CUDA. Some libraries (FlashAttention, certain quantization kernels) have limited or no ROCm support, which can affect inference performance and compatibility with popular serving frameworks like vLLM. For cloud rental, availability varies significantly by provider — some specialize in this GPU tier while others may have limited stock. Compare on-demand and spot pricing across providers using the rental comparison table on this page, and factor in region availability if latency is a concern for your inference workload.