Rebellions Rebel100: 4-chiplet UCIe mesh and predictive DMA for inference
memory bound
Fields
id
flops/alt-rebellions-chiplet-meshClass
custom_kernel
Bound by
memory
Arithmetic intensity
Rebel100 uses 4 chiplets connected via UCIe-Advanced mesh at 1 TB/s per bidirectional channel at 11 ns latency. The four chiplets present as one virtual die with no software-visible boundary. 144 GB HBM3E at 4.8 TB/s. The predictive DMA engine enables 2.7 TB/s effective bandwidth for KV data prefetch on long-context decode. The inter-chiplet link, not the memory, is the architectural story — 11 ns is roughly a register-to-register hop.
Scales with
hbm3e_bandwidth (4.8 TB/s, 144 GB capacity)ucie_mesh_bandwidth (1 TB/s per channel, 11 ns latency)predictive_dma (2.7 TB/s effective for KV prefetch)num_chiplets (4, presenting as one virtual die)
Workarounds
Use RBLN SDK compiler and runtimeUse vLLM-rbln for servingLeverage predictive DMA for KV cache prefetch on long-context decodeUse the 4-chiplet mesh for tensor parallelism without software-visible boundaries
notes
The inter-chiplet mesh is the key architectural innovation. 11 ns latency is roughly a register-to-register hop, which means the 4 chiplets behave like a single die. This eliminates the inter-chip communication overhead that plagues multi-GPU inference. The predictive DMA engine is a software-visible optimization for KV cache prefetch — it can be tuned rather than being a fixed property. The 144 GB HBM3E capacity is sufficient for large models. Production deployment: SK Telecom's A.X K1 sovereign LLM runs on RebelServer.
Notes
The inter-chiplet mesh is the key architectural innovation. 11 ns latency is roughly a register-to-register hop, which means the 4 chiplets behave like a single die. This eliminates the inter-chip communication overhead that plagues multi-GPU inference. The predictive DMA engine is a software-visible optimization for KV cache prefetch — it can be tuned rather than being a fixed property. The 144 GB HBM3E capacity is sufficient for large models. Production deployment: SK Telecom's A.X K1 sovereign LLM runs on RebelServer.