RESERVED MI355X

MI355X reserved rates are quoted per-engagement and depend on cluster size, term length, region, and provider. As the flagship CDNA 4 part with the highest TDP in the AMD lineup (1,400W versus 1,000W on MI350X), MI355X carries an infrastructure premium — providers price in the rack-power and cooling overhead. Commitment length is the single biggest pricing lever on MI355X: the gap between short-term and long-term reserved rates is wider than on any other AMD GPU on the market. Compute Exchange returns indicative pricing within 24 hours of an RFQ, anchored to your specific configuration.

MI355X is the flagship CDNA 4 part, stepping up FP8 compute to 5.0 PFLOPS dense (versus 4.5 PFLOPS on MI350X) while keeping the same 288GB HBM3e memory footprint. TDP also rises to 1,400W versus 1,000W on MI350X, which affects rack-power density and cooling requirements. MI355X is positioned for training-and-inference balance; MI350X leans more purely inference-optimized. For organizations running mixed-precision training (FP8/BF16) alongside inference, MI355X is the better fit — and reserved capacity is more broadly available than MI350X.

MI355X offers 288GB HBM3e versus 192GB HBM3e on B200, with comparable FP8 compute (5.0 PFLOPS dense on MI355X vs ~4.5 PFLOPS dense on B200). The memory advantage favors MI355X for single-device serving of the largest models and for FP8-quantized inference. B200's NVLink 5 fabric remains the advantage for multi-node training at frontier scale. Software ecosystem maturity is CUDA's strongest argument, though ROCm has closed meaningfully in 2025 — the gap is narrower than it was on prior AMD generations.

MI355X is available across the full reservation spectrum from 1-month to 36-month commits. The economics of long commitments differ sharply from MI325X — the gap between short-term and long-term reserved rates is the widest in the AMD lineup, making term length the single biggest pricing lever on MI355X. 1- and 3-month terms make sense for pilot training runs; 12-month is the typical production threshold; 36-month is the right call for sustained training infrastructure or production inference workloads above 60% utilization. Compute Exchange returns indicative pricing across whichever term scenarios you specify in your RFQ, so you can evaluate the trade-off in a single response.