Editorial illustration for DeepSeek-V4 Flash RL Runs on AMD Instinct MI308X GPUs
DeepSeek-V4 Flash RL Runs on AMD MI308X GPUs
Three AMD Instinct MI308X nodes, 24 GPUs total, and one long-context reinforcement learning run on DeepSeek-V4 Flash. That's the setup behind a new test of AMD's AITER-backed RL stack, built to prove out something more specific than raw throughput: whether training and inference can run as separate, coordinated services on the same hardware family without breaking numerical stability or starving the rollout engine of memory.
The split is deliberate. Two MI308X nodes carry a BF16 actor, chosen because gradient computation and optimizer updates need that precision. A third node runs an FP8 rollout engine, where the smaller footprint per parameter frees up GPU memory for KV cache and longer generated sequences. Ray handles placement and coordination across both GPU groups.
This isn't a benchmark of inference alone, or a single isolated training step. The test tracks a full GRPO run, including routing replay, online RDMA weight updates, evaluation, and checkpoint recovery, over the first 75 optimizer steps. The goal was to show one working three-node configuration on MI308X hardware that handles rollout and training together, end to end.
Training and inference run as disaggregated services: two MI308X nodes host a BF16 actor that preserves the numerical stability required by gradient computation and optimizer updates, while a dedicated third MI308X node hosts an FP8 rollout engine whose smaller per-parameter footprint leaves more GPU memory available for KV cache and longer generated sequences.
Why this matters
This is AMD trying to prove Instinct can handle the messy, full-stack reality of RL training, not just a benchmark slide. Running rollout, actor training, RDMA weight updates, and checkpoint recovery across three MI308X nodes is the kind of workflow that breaks quietly in production long before it breaks in a demo. If AITER genuinely holds up under that, it gives teams building on DeepSeek-V4 Flash a real alternative to Nvidia for RL pipelines, which matters a lot given how hard H100/H200 capacity still is to get.
For researchers, the disaggregated design (separating FP8 rollout from BF16 MoE training) is worth studying on its own, independent of the hardware story. For founders and infra teams, the open question is whether this scales past 24 GPUs and three nodes without new bottlenecks showing up. A controlled demo on three OAM nodes tells you the stack works.
It doesn't yet tell you it's cheaper, faster, or more reliable than the Nvidia path at real scale. Watch for independent throughput numbers before treating this as a settled alternative.
Related Reading
Research & Benchmarks Microsoft AI Chief: "If It Isn't Safe We Shouldn't Build It
Research & Benchmarks Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks Nvidia's DLSS 4.5 beta adds 6x Multi Frame Generation for RTX 50 GPUs
From Archives New NVIDIA RTX Laptops Run AI Models Locally on CUDA
From Archives New Analysis Index Widens Gap Between Open and Closed AI Models
Common Questions Answered
How does DeepSeek-V4 Flash separate training and inference on AMD Instinct MI308X GPUs?
DeepSeek-V4 Flash uses a disaggregated services approach where two MI308X nodes host a BF16 actor for training that preserves numerical stability for gradient computation and optimizer updates, while a dedicated third MI308X node hosts an FP8 rollout engine for inference. This split allows the rollout engine's smaller per-parameter footprint to leave more GPU memory available for KV cache and longer generated sequences.
Why is BF16 precision chosen for the actor in this DeepSeek-V4 Flash RL setup?
BF16 precision is chosen for the actor because it preserves the numerical stability required by gradient computation and optimizer updates during training. This precision level is critical for maintaining the accuracy of backpropagation and weight updates in reinforcement learning workflows.
What advantage does the FP8 rollout engine provide in the three-node MI308X configuration?
The FP8 rollout engine has a smaller per-parameter footprint compared to higher precision formats, which leaves more GPU memory available for KV cache and enables longer generated sequences during inference. This memory efficiency allows the rollout engine to generate longer context windows without requiring additional hardware resources.
What production challenges does this AMD AITER RL stack test address?
The test addresses real-world challenges like maintaining numerical stability across disaggregated services, preventing memory starvation of the rollout engine, and coordinating training and inference as separate services without breaking. The test specifically validates that rollout, actor training, RDMA weight updates, and checkpoint recovery can work reliably across three MI308X nodes, which are the kinds of workflows that often fail silently in production.
What does this DeepSeek-V4 Flash demonstration mean for teams building reinforcement learning pipelines?
This demonstration provides teams building on DeepSeek-V4 Flash with a real alternative to Nvidia for RL pipelines, proving that AMD Instinct can handle the full-stack complexity of production reinforcement learning workflows. If AITER holds up under these demanding conditions, it significantly expands hardware options for organizations scaling RL training and inference.
Further Reading
- DeepSeek-V4 Flash RL on AMD Instinct MI308X GPUs with an AITER-Backed RL Stack - AMD ROCm Blog
- veRL on AMD: Production-Ready RL Post-Training on ROCm - AMD ROCm Blog
- Enabling DeepSeek-V4-Flash Training on AMD Instinct MI355X GPUs with Primus - AMD ROCm Blog
- DeepSeek V4 Flash — NeMo-RL - NVIDIA NeMo-RL Documentation
