Eagle Talon
Technical specification · draft · 2026-11-11
Reprogrammable FPGA inference server

Eagle Talon

Seven AMD Alveo V80 FPGAs and one NVIDIA H200 in one 4U server. The FPGAs run the speculative-decoding drafter, and the GPU verifies. A new FPGA image loads in minutes, so the drafter changes with the model.

889.97tok/s
Llama-3.2-3B, single user, measured
249.62tok/s
Qwen3.8-27B, single user, measured
7 + 1
AMD Alveo V80 FPGAs + NVIDIA H200 NVL
~1.1kW
whole server at the wall while serving, measured

Calculations and hypothesis

System

ChassisASUS ESC8000-E11, 4U rack, 8 dual-slot accelerator bays vendor
Dimensions / weight800 × 440 × 174.5 mm (D × W × H), about 27 kg net, barebone vendor
Host CPU2 × Intel Xeon Gold 6426Y, 16 cores each (64 threads) as installed
Host memory32 GB, as tested as installed
Storage3.5 TB system volume, as tested as installed
Power suppliesUp to 4 × 3,000 W Titanium vendor
Operating systemUbuntu 24.04 LTS as installed
Expansion slots freeTBD

Accelerators

FPGA7 × AMD Alveo V80 (Versal HBM), passive cooling as installed
FPGA memory32 GB HBM2e per card, 819 GB/s peak per card vendor
FPGA networking4 × QSFP56 per card, 800 Gb/s aggregate per card vendor
FPGA host linkPCIe Gen5 ×8 ×8 or Gen4 ×16 per card vendor
FPGA card powerUp to 190 W TDP per card vendor; 57–68 W per card measured at idle
GPU1 × NVIDIA H200 NVL, 141 GB HBM3e, 600 W power limit as installed

Measured performance

ModelSystemtok/stokens / stepDate
Llama-3.2-3B7 FPGAs + H200 (best configuration)889.976.912026-11-11
Llama-3.2-3BH200 alone (its best configuration)835.126.632026-11-09
Llama-3.2-3B4 FPGAs + H200, tree depth 15870.067.122026-11-11
Llama-3.2-3BH200 alone, tree depth 15 (same day)700.476.882026-11-11
Qwen3.8-27B4 FPGAs (drafter) + H200249.627.142026-11-09
Qwen3.8-27BH200, drafter on the GPU251.207.132026-10-23
Qwen3.8-27BH200, target model alone61.0812026-10-23

Single user (batch 1), HumanEval 80 prompts × 256 new tokens, SGLang, the same H200 in every row. Tokens per step = average tokens accepted per verify step. All other models: TBD (not measured on this system).

Power

Whole server, idle752 W at the wall (BMC reading, 7 FPGAs + H200 powered) measured 2026-11-04
Whole server, serving1,011–1,110 W at the wall, Llama-3.2-3B single user, 4-FPGA configuration measured 2026-11-04
Peak at power-on1,457 W (two PDU outlets) measured 2026-11-11
Full-load envelope, 7 FPGAs busyTBD

Software and reprogrammability

Serving stack

SGLang with Eagle's EAGLE4 speculative decoding. The FPGAs draft a token tree, and the H200 verifies it in one pass.

Reprogrammable

A new FPGA image is written in about 15–20 minutes per card. The same cards have run drafters for both Llama and Qwen models.

Model coverage

Measured: Llama-3.2-3B and Qwen3.8-27B. Broader open-weights coverage: TBD.

Not yet measured

Memory-bandwidth efficiencyTBD
Multi-user throughputTBD
Larger models (e.g. 8B, 70B)TBD
Energy per token vs GPU-onlyTBD. On 2026-11-04, at tree depth 7, the GPU-only arm used less energy per token than the FPGA arm.