Cerebras WSE-3: what does NOT work well — capacity, flexibility, and ecosystem
memory bound
Fields
id
flops/alt-cerebras-limitationsClass
custom_kernel
Bound by
memory
Arithmetic intensity
Cerebras WSE-3 limitations: (1) capacity — 44 GB SRAM per wafer is small vs GPU HBM (80-192 GB), so models >44 GB require multi-wafer pipelining; (2) no dynamic batching — the wafer-scale architecture is optimized for single-stream low-latency decode, not high-throughput batched inference; (3) limited software ecosystem — no vLLM, no PyTorch native support; uses Cerebras SDK only; (4) no fine-tuning on CS-3/CS-4 (training is separate); (5) high cost per system; (6) limited to Cerebras Cloud or on-prem deployment.
Scales with
model_size (44 GB limit per wafer)batch_size (optimized for single-stream, not high-throughput batching)software_ecosystem (limited vs GPU)deployment_flexibility (Cerebras Cloud or on-prem only)
Workarounds
Use multi-wafer pipelining for models >44 GBUse disaggregated inference: prefill on GPU/Trainium, decode on CerebrasUse Cerebras Cloud for on-demand access without hardware purchaseAccept the software ecosystem limitations — use Cerebras SDK
notes
Honest negatives: Cerebras publishes no per-chip throughput figure for standard benchmarks (no MLPerf, no third-party throughput comparisons). The 750 tok/s figure is for a specific model (GPT-5.6 Sol Ultrafast) and is a vendor claim. The architecture is optimized for low-latency single-stream decode, not high-throughput batched inference. The software ecosystem is limited — no vLLM, no PyTorch native support. Training is not supported on CS-3/CS-4 (separate systems). These are significant limitations for general-purpose inference workloads.
Notes
Honest negatives: Cerebras publishes no per-chip throughput figure for standard benchmarks (no MLPerf, no third-party throughput comparisons). The 750 tok/s figure is for a specific model (GPT-5.6 Sol Ultrafast) and is a vendor claim. The architecture is optimized for low-latency single-stream decode, not high-throughput batched inference. The software ecosystem is limited — no vLLM, no PyTorch native support. Training is not supported on CS-3/CS-4 (separate systems). These are significant limitations for general-purpose inference workloads.