EAGLE-4 Runs on AMD’s MI355X From Day One in SGLang and ATOM
AMD reports 289 tokens per second for EAGLE-4 in SGLang and 309 in ATOM on one MI355X GPU serving one Qwen3.8-27B request, versus 82 for plain decoding.
AMD says EAGLE-4 runs on Instinct MI355X GPUs from day one in SGLang and ATOM. In its October 20 results, AMD reports 289 tokens per second for EAGLE-4 in SGLang and 309 in ATOM on a single Qwen3.8-27B request, compared with 82 for plain decoding.
What the comparison measures
The figures use one MI355X GPU, Qwen3.8-27B with BF16 weights, and SPEED-Bench with thinking enabled and greedy decoding. The same table lists Qwen3.8’s built-in multi-token prediction (MTP) baseline at 178 tokens per second in SGLang under the same single-request setup. These are AMD’s measurements, not an independent reproduction.
EAGLE-4 uses speculative decoding: a smaller drafter proposes several tokens, then the larger model checks the proposed group in one pass. That mechanism explains what the comparison is testing. It does not predict performance for another model, batch, decoding policy, or multi-GPU service, so operators should match the benchmark conditions to their own workload before using the numbers in a capacity estimate.