An AC-130 and a B-52 are both formidable aircraft. Both can deliver extraordinary firepower, but they were designed for different missions. Comparing them solely by payload, speed, or endurance would miss the point. The important question is not which aircraft is better. It is which aircraft is better suited to the mission.

I have two local AI computers named after those aircraft. AC130 is a compact AMD Ryzen AI Max+ 395 system with 128 GB of unified memory. B52 is a larger workstation equipped with a Threadripper PRO processor and four NVIDIA RTX 4090 GPUs.

They represent two approaches to running AI locally: one emphasizes accessible memory capacity in a compact package; the other emphasizes discrete GPU throughput and parallel execution. I ran identical model files on both systems, measured prompt processing and generation, tested concurrency, and collected reported GPU power.

The question is not which machine wins. It is which machine is suited to the mission.

Two Very Different Architectures

AC130

128 GB Unified Memory

AC130 compact Ryzen AI Max+ workstation
AC130 — compact system with 128 GB unified memory.

AC130 combines a 16-core AMD Ryzen AI Max+ 395 processor with Radeon 8060S integrated graphics. CPU and GPU share 128 GB of LPDDR5X memory, allowing the GPU to access models larger than the dedicated memory on a typical consumer graphics card, subject to allocation and operating-system limits.

Memory capacity is not compute performance. The integrated GPU shares memory bandwidth with the CPU and has much less GPU compute capacity than four RTX 4090 cards. Its appeal is substantial accessible memory in a compact system with comparatively modest power requirements.

B52

4× NVIDIA RTX 4090 GPUs

B52 workstation with four liquid-cooled NVIDIA RTX 4090 graphics cards
B52 — four discrete GPUs, each with 24 GB dedicated video memory.

B52 combines a 24-core Threadripper PRO 5965WX, 128 GB of system RAM, and four RTX 4090 GPUs. Its 96 GB aggregate VRAM is distributed across four devices, not presented as one unified memory pool. A model that fits on one GPU can run independently; larger models can be split across cards at the cost of inter-GPU coordination.

The machine also requires more space and infrastructure: liquid cooling, two 1,500 W power supplies on separate circuits, and supplemental room air conditioning. Those costs and constraints matter alongside throughput.

How I Tested

ModelArchitectureQuantizationFile size
Qwen2.5 7B Instruct7.62B denseQ4_K_M4.36 GiB
Qwen3.6 35B A3B34.66B MoEQ4_K_XL20.05 GiB
Qwen2.5 72B Instruct72.71B denseQ8_071.95 GiB

I verified identical model files on both machines with SHA-256. Both used llama.cpp, but AC130 ran Vulkan build 10889 and B52 ran CUDA build 11844. The results therefore compare complete hardware-and-software configurations, not hardware in isolation.

Standard tests used 2,048 prompt tokens and 128 generated tokens, with three repetitions. Separate extended runs generated 4,096 tokens and collected power telemetry. The 35B model uses a Mixture of Experts (MoE) architecture, activating only a portion of its total parameters for each token.

Prompt Processing and Generation

Prompt processing matters when a model ingests long documents, retrieved context, or extensive conversation histories. B52 led substantially on every model.

ModelAC130 prompt (t/s)B52 prompt (t/s)B52 advantage
7B1,12411,91910.6×
35B MoE1,0708,5188.0×
72B992,68026.9×

Extended generation tells a different, though still B52-favoring, story:

ModelAC130 generation (t/s)B52 generation (t/s)B52 advantage
7B44.61178.914.0×
35B MoE62.84223.423.6×
72B2.8112.274.4×
Measured prompt processing and sustained generation rates for three identical AI model files on AC130 and B52
Measured inference performance: prompt processing and sustained token generation.

The 35B MoE model generated tokens faster than the 7B dense model on both systems. Total parameter count is not a reliable speed predictor: active parameters, model structure, memory behavior, and backend implementation all matter. This result is specific to these models and configurations.

The 72B Reality Check

Both systems ran the 72B Q8 model. AC130 accommodated it in unified memory; B52 distributed it across four GPUs. But 2.81 tokens per second on AC130 is too slow for many interactive uses, and even B52’s 12.27 tokens per second is far below its smaller-model performance. Enough memory makes a model possible, not necessarily practical.

Four Models at Once

For a concurrency test, I launched four independent 7B inference processes. On AC130 they shared the integrated GPU; on B52 each ran on a separate RTX 4090.

MeasurementAC130B52
Concurrent processes44
Aggregate generation50.43 t/s716.38 t/s
Average per process12.61 t/s179.10 t/s
Measured aggregate throughput for four concurrent 7B inference processes on AC130 and B52
Four concurrent 7B workloads: aggregate sustained generation throughput.

B52 delivered 14.2 times the aggregate throughput, with each GPU maintaining approximately its standalone generation rate. AC130’s per-process rate fell to roughly 12–13 tokens per second as four jobs competed for one integrated GPU. The aggregate figures are sums of independent benchmark rates, not separately measured application-serving throughput.

Power Consumption Adds Another Tradeoff

I recorded AC130’s AMD graphics-package power sensor and B52’s NVIDIA GPU telemetry, summing the active GPUs. These are not wall-power measurements, and the sensors do not necessarily cover identical electrical domains. Power averages also include some initialization and loading, so energy-per-token values below are indicative rather than calibrated generation-only measurements.

WorkloadAC130 reported powerB52 reported power
One 7B60.19 W349.12 W
One 35B MoE60.56 W254.57 W
One 72B45.18 W455.72 W
Four concurrent 7B56.82 W1,391.21 W
Measured generation throughput and reported GPU power for AC130 and B52, with power consumption distinctly marked
Generation performance versus reported GPU power; power is not whole-system wall consumption.

The approximate reported energy per generated token was 1.35 versus 1.95 J/token for 7B, 0.96 versus 1.14 for 35B MoE, and 16.08 versus 37.14 for 72B (AC130 versus B52). Those figures suggest a GPU-level energy advantage for AC130 in these tests, but different power boundaries and sampling periods prevent a definitive whole-system efficiency claim.

Electricity cost depends on duty cycle as much as instantaneous power. At $0.14–$0.15/kWh, a hypothetical system averaging 200 W at the wall for 30 days costs roughly $20–$22; one averaging 1,700 W costs roughly $171–$184. These are illustrations, not measured monthly costs for either computer. Cooling and idle consumption also matter.

Matching the Machine to the Mission

Benchmarks are useful, but the real question is what work needs to get done. The same advantage that matters enormously for one workload may be irrelevant for another.

WorkloadLikely better fitWhy
Single-user personal assistantAC130Useful smaller-model speed without four discrete GPUs
Large-model experimentationAC130Unified memory accommodates large models without multi-GPU distribution
Interactive inference on larger modelsB52Faster prompt processing and generation
Multiple independent agentsB52Four GPUs isolate independent inference jobs
High-volume document processingB52Much higher prompt-processing throughput
Batch generationB52Faster generation and independent parallel jobs
Local development and prototypingAC130Compact system and flexible memory capacity
Several separate models at onceB52Dedicated GPU resources when models fit individual cards
Continuous modest inferenceAC130Lower reported power, pending wall-level validation

These are not rigid assignments. AC130 is not necessarily faster at running large models; its advantage is accommodating them without multiple GPUs. B52 is not automatically the sensible choice for a single assistant whose output speed is already adequate. A short, demanding batch job may favor B52 even when its instantaneous power is much higher.

Final Thoughts

B52’s throughput advantage was substantial, especially under concurrent load. AC130’s ability to run a 72B Q8 model within unified memory was noteworthy, even at limited generation speed. Power measurements revealed another dimension, but not a universal answer.

An AC-130 and a B-52 do not perform the same mission in the same way. Their value comes from what each was designed to do. The same principle applies to local AI hardware.

The question is not which machine wins. It is which machine is suited to the mission.

Methodology note: identical SHA-256-verified GGUF files; llama-bench GPU offload enabled; standard 2,048-token prompt and 128-token generation tests repeated three times; extended 4,096-token generation runs. AC130: Vulkan build 10889. B52: CUDA build 11844. GPU telemetry is not wall power; different sensor boundaries and benchmark sampling intervals limit direct energy-efficiency conclusions.