Chip Speculative Decoding Comparisons
62 speculative decoding comparisons across DeepSeek V4 Pro 1.6T, DeepSeek R1, Kimi K3 2.8T, GLM 5/5.1, MiniMax M3 428B, and Qwen 3.5 397B-A17B. Each page compares inference with the speculative decoding method (MTP, EAGLE, etc.) enabled versus disabled on the same model and chip — throughput, cost, and interactivity at matched operating points.
DeepSeek V4 Pro 1.6T
8 speculative decoding comparisons with benchmark data on DeepSeek V4 Pro 1.6T.
B200 FP4 — MTP vs Off
NVIDIA · Blackwell
B300 FP4 — MTP vs Off
NVIDIA · Blackwell
GB200 NVL72 FP4 — MTP vs Off
NVIDIA · Blackwell
GB300 NVL72 FP4 — MTP vs Off
NVIDIA · Blackwell
H200 FP8 — MTP vs Off
NVIDIA · Hopper
MI300X FP8 — MTP vs Off
AMD · CDNA 3
MI325X FP8 — MTP vs Off
AMD · CDNA 3
MI355X FP4 — MTP vs Off
AMD · CDNA 4
DeepSeek R1
13 speculative decoding comparisons with benchmark data on DeepSeek R1.
B200 FP4 — MTP vs Off
NVIDIA · Blackwell
B200 FP8 — MTP vs Off
NVIDIA · Blackwell
B300 FP4 — MTP vs Off
NVIDIA · Blackwell
B300 FP8 — MTP vs Off
NVIDIA · Blackwell
GB200 NVL72 FP4 — MTP vs Off
NVIDIA · Blackwell
GB200 NVL72 FP8 — MTP vs Off
NVIDIA · Blackwell
GB300 NVL72 FP4 — MTP vs Off
NVIDIA · Blackwell
GB300 NVL72 FP8 — MTP vs Off
NVIDIA · Blackwell
H100 FP8 — MTP vs Off
NVIDIA · Hopper
H200 FP8 — MTP vs Off
NVIDIA · Hopper
MI325X FP8 — MTP vs Off
AMD · CDNA 3
MI355X FP4 — MTP vs Off
AMD · CDNA 4
MI355X FP8 — MTP vs Off
AMD · CDNA 4
Kimi K3 2.8T
1 speculative decoding comparison with benchmark data on Kimi K3 2.8T.
GLM 5/5.1
11 speculative decoding comparisons with benchmark data on GLM 5/5.1.
B200 FP4 — MTP vs Off
NVIDIA · Blackwell
B200 FP8 — MTP vs Off
NVIDIA · Blackwell
B300 FP4 — MTP vs Off
NVIDIA · Blackwell
B300 FP8 — MTP vs Off
NVIDIA · Blackwell
GB200 NVL72 FP4 — MTP vs Off
NVIDIA · Blackwell
GB200 NVL72 FP8 — MTP vs Off
NVIDIA · Blackwell
GB300 NVL72 FP4 — MTP vs Off
NVIDIA · Blackwell
GB300 NVL72 FP8 — MTP vs Off
NVIDIA · Blackwell
H200 FP8 — MTP vs Off
NVIDIA · Hopper
MI325X FP8 — MTP vs Off
AMD · CDNA 3
MI355X FP8 — MTP vs Off
AMD · CDNA 4
MiniMax M3 428B
11 speculative decoding comparisons with benchmark data on MiniMax M3 428B.
B200 FP4 — M3 EAGLE vs Off
NVIDIA · Blackwell
B200 FP8 — M3 EAGLE vs Off
NVIDIA · Blackwell
B300 FP4 — M3 EAGLE vs Off
NVIDIA · Blackwell
B300 FP8 — M3 EAGLE vs Off
NVIDIA · Blackwell
GB300 NVL72 FP8 — M3 EAGLE vs Off
NVIDIA · Blackwell
H100 FP8 — M3 EAGLE vs Off
NVIDIA · Hopper
H200 FP8 — M3 EAGLE vs Off
NVIDIA · Hopper
MI300X FP8 — M3 EAGLE vs Off
AMD · CDNA 3
MI325X FP8 — M3 EAGLE vs Off
AMD · CDNA 3
MI355X FP4 — M3 EAGLE vs Off
AMD · CDNA 4
MI355X FP8 — M3 EAGLE vs Off
AMD · CDNA 4
Qwen 3.5 397B-A17B
18 speculative decoding comparisons with benchmark data on Qwen 3.5 397B-A17B.
B200 FP4 — MTP vs Off
NVIDIA · Blackwell
B200 FP8 — MTP vs Off
NVIDIA · Blackwell
B200 BF16 — MTP vs Off
NVIDIA · Blackwell
B300 FP4 — MTP vs Off
NVIDIA · Blackwell
B300 FP8 — MTP vs Off
NVIDIA · Blackwell
B300 BF16 — MTP vs Off
NVIDIA · Blackwell
GB200 NVL72 FP8 — MTP vs Off
NVIDIA · Blackwell
GB300 NVL72 FP4 — MTP vs Off
NVIDIA · Blackwell
GB300 NVL72 FP8 — MTP vs Off
NVIDIA · Blackwell
H100 FP8 — MTP vs Off
NVIDIA · Hopper
H200 FP8 — MTP vs Off
NVIDIA · Hopper
MI300X FP8 — MTP vs Off
AMD · CDNA 3
MI325X FP8 — MTP vs Off
AMD · CDNA 3
MI325X BF16 — MTP vs Off
AMD · CDNA 3
MI355X FP4 — MTP vs Off
AMD · CDNA 4
MI355X FP8 — MTP vs Off
AMD · CDNA 4
MI355X BF16 — MTP vs Off
AMD · CDNA 4
RTX PRO 6000 FP4 — MTP vs Off
NVIDIA · Blackwell