AMD Primus Projection: Tool for Predicting LLM Training Memory and Speed Before Running on Instinct GPU Clusters

AMD has introduced Primus Projection on its ROCm blog — a tool that answers two practical questions for ML engineers before they spend hours of cluster time: “Will the model fit in memory?” and “How fast will it train?”. The tool specifically targets AMD Instinct GPU accelerators and integrates with the existing ROCm stack.

What Does the Tool Calculate?

Primus Projection combines analytical formulas and real GPU benchmarking to estimate two key components of any training run. The memory side breaks down into three parts: model parameters in BF16 format, optimizer state (FP32 master weights + Adam first/second moments, sharded along the data parallelism dimension), and activations — intermediate results the pipeline must retain for the backward pass, scaled by the number of micro-batches and the MoE routing factor.

For speed prediction, the tool offers two complementary approaches. It can benchmark representative layers on available hardware (even on a single GPU), then analytically extrapolate to the full cluster by reversing the removed parallelism dimensions — Pipeline → Expert → Tensor Parallelism. Alternatively, it can perform pure CPU simulation via GEMM and attention analytical modeling, useful when GPUs are not available.

Particularly notable is support for communication modeling: AllReduce, All-to-All, and P2P collectives with topology awareness, along with pipeline schedules such as 1F1B, interleaved, and zero-bubble, with precise calculation of “bubble” periods during which GPUs are idle.

How Accurate Are the Projections?

According to AMD, projections track actual multi-node measured results within approximately 10% error. Validation was conducted on dense models like Llama and MoE architectures like Mixtral, with test hardware comprising MI325X and MI355X accelerators — AMD’s latest Instinct chips.

The practical value of this accuracy is concrete: if an engineer estimates that training will take 72 hours on 512 GPUs, a 10% error means a range of ~65 to ~79 hours — sufficient for planning, budgeting, and reasonable cluster time reservation.

Who Is the Tool For?

The primary audience is ML engineers and research teams working on AMD infrastructure — whether that is an on-premise Instinct cluster or rented capacity from a cloud partner. The tool removes the practical barrier of “blind” experiment launches that for years has favored teams with unlimited budgets for a “try and see” approach.

The broader message is that AMD is steadily filling the software ecosystem around ROCm — historically its weaker point compared to Nvidia’s CUDA world. Tools like Primus Projection, combined with increasingly frequent Hugging Face and PyTorch support for ROCm, are gradually reducing the switching cost for teams considering AMD as an alternative.