SERIAL ALICEVERIFIABLE AI ENERGY
TDX-CERTIFIED · POLYGON MAINNET · JUNE 2026

GPU Energy Benchmarkmeasured from silicon.

Real power consumption from H100, H200 and B200 — every data point a signed certificate, anchored on Polygon. Not datasheets. Not estimates.

3
GPU architectures
7+
H200 TDX certs
0.8
hardware trust score
0
trust in us needed
Hardware measured

Three generations.
One verifiable truth.

H100 and B200 from vLLM benchmark runs. H200 from a Phala CVM — hardware-attested at the chip, every sample bound to an Intel TDX quote.

Benchmark

NVIDIA H100 SXM 80 GB

267–680 W
8 models · BF16, GPTQ, AWQ
5 batch sizes (b1 → b128)
TDX-certified

NVIDIA H200 SXM 141 GB

80–593 W
TDX TEE · Ed25519 · Polygon
Ghost power · vLLM Phi-3-mini ladder
Benchmark

NVIDIA B200 SXM 180 GB

452–943 W
6 models incl. Mixtral MoE
5 batch sizes · Blackwell
Power by batch size

Phi-3-mini (3.8B) — actual draw vs. TDP

phi-3-mini · power vs batch size H200 TDX certified
H100 SXM 80 GB H200 SXM 141 GB TDX certified B200 SXM 180 GB H200 TDP (700 W)
H200 actual at batch=128: 516.9 W — not the 700 W TDP. Using TDP in “per-MW” claims understates H200 efficiency by ~35%. B200 draws +17% more power than H200 at the same batch (606 W vs 517 W). Verify H200 b=128 cert →
Key findings

What the data says.

13.6%
H200 ghost power
80.3 W idle vs 592.8 W at load. The floor every workload pays, even at zero throughput.
+17%
B200 vs H200 draw
B200 pulls +17% more power than H200 at batch=128, same model. More compute, more watts.
30–80×
Batch efficiency gain
Energy per token from batch=1 to batch=128 across all hardware. Batch size is the biggest lever.
≈0%
FP8 vs BF16 power
661.7 W (FP8) vs 662.8 W (BF16) on H200. FP8 gives +68% throughput at the same wattage.
Energy efficiency at batch=128

μWh per token (lower = better)

μWh / token · b=128
H100 B200
Mixtral 8×7B MoE: 36.4 μWh/tok vs dense 7B at ~10 μWh/tok. Routing overhead costs ~3.5× more energy per token.
Quantization impact

Llama-3 8B H100 · batch=64

quantization · power + efficiency
Power (W) μWh/tok
AWQ vs BF16: −28% power, −47% energy per token.
Ghost power
H200 idle vs load · TDX certifiedcertified
80.3 W idle is always-on — at b=1 it accounts for ~26% of energy per request; at b=128 it amortises to ~15.5%.
Verified certificates

H200 — Intel TDX · Ed25519 · Polygon Mainnet

sa — h200 sxm 141gb · phala tdx cvm VERIFIED ON-CHAIN
WorkloadPowerEnergyCertificateVerify
H200 idle (0% util) 80.3 W sa-c020e48c… verify →
H200 BF16 matmul load 592.8 W sa-40cb8908… verify →
vLLM Phi-3-mini BF16 · b=1 309.8 W 5.30 Wh sa-ee4cd7b3… verify →
vLLM Phi-3-mini BF16 · b=16 354.8 W 6.06 Wh sa-1a051cf4… verify →
vLLM Phi-3-mini BF16 · b=64 451.5 W 7.83 Wh sa-1751b088… verify →
vLLM Phi-3-mini BF16 · b=128 516.9 W 8.81 Wh sa-d64b2e20… verify →
FP8 matmul (same draw as BF16) 661.7 W 11.52 Wh sa-0244226d… verify →
All H200 certificates: trust_score=0.8 · posture: hardware_attested · TEE: Intel TDX · MRTD verified · TCB: UpToDate · Merkle-anchored on Polygon Mainnet. Active signing key →