Calculators Live Utility 💻 Local Browser Processing

Transformer Calculator: Turns Ratio, kVA & Wire Sizing

Calculate electrical power transformer parameters including turns ratio, line and phase voltages, kVA sizing, AWG wire gauge, 3-phase Dyn11 and Zigzag connections, core losses, and short-circuit fault currents in browser.

Dual-Discipline Engineering Suite

Transformer Engineering Workbench

Unified engineering suite for physical electrical power transformers and artificial intelligence neural network architectures.

Electrical Magnetics & Winding Design Lab

Design single-phase and three-phase transformers with physical core sizing, turns ratios, wire gauge ampacity, vector groups, and fault currents.

Load Engineering Preset:

Phase & Connection Geometry:

Primary Line Voltage (V):

Secondary Line Voltage (V):

Rated Capacity (kVA):

Supply Frequency (Hz):

Peak Flux Density Bmax (Tesla):

Allowable Current Density J (A/mm2):

Percentage Impedance (%Z):

Load Power Factor (cos phi):

Operating Temperature (deg C):

Turns & Voltage Ratio

4.00 : 1

Step-Down Transformer

Dyn11: +30 degrees (LV leads HV)

Full-Load Current (FLA)

Primary Line FLA:90.2 A

Secondary Line FLA:208.2 A

Primary Winding FLA:52.1 A

Core Area & Turns Per Volt

Core Area (Ac):181.8 cm2

Turns Per Volt (TPV):0.1473 TPV

Primary Turns:71 Turns

Secondary Turns (+5%):19 Turns

Conductor Sizing & AWG

Primary Wire:5 AWG (18.6 mm2)

Secondary Wire:00 AWG (74.4 mm2)

Skin depth at 60 Hz: 8.52 mm

Short-Circuit Fault Current

4.63 kA

Primary Fault Current:2.00 kA

Fault Capacity:1.67 MVA

Losses & Full-Load Efficiency

96.85 %

Core Loss (Steinmetz):934 W

Copper Loss (at 80 deg C):1,482 W

Interactive Physical Transformer Magnetic Circuit Schematic

AI / LLM Neural Architecture & VRAM Engine

Calculate total and active parameter counts, precision VRAM footprints, KV cache scaling, training memory, and token generation speeds.

Load Model Architecture Preset:

Hidden Dimension (d_model):

Transformer Layers (L):

Query Attention Heads (h):

Key-Value Heads (h_kv):

Head Dimension (d_k):

FFN Intermediate Dim (d_ff):

Vocabulary Size (V):

SwiGLU Architecture (3 FFN Matrices)Tie Input & Output Embeddings

Enable Mixture of Experts (MoE)

Routed Experts (E):

Active Experts per Token (k):

Shared Experts:

Initial Dense Layers Count:

Shared Expert FFN Dim:

Dense Layer FFN Dim:

Enable Multi-Head Latent Attention (MLA)

KV Latent Dimension (d_c):

Decoupled RoPE Key Dim (d_r):

Query Latent Dimension (d_c'):

Inference Precision & Context Configuration:

Model Weights Precision:

Context Length (Tokens):

Inference Batch Size:

KV Cache Precision:

Total Model Parameters

8.03B

Active Parameters per Token:8.03B

Attention Ratio: GQA 4:1

Model Weights VRAM

14.96 GB

FP16 (2.0 Bytes / Parameter)

KV Cache VRAM Footprint

1.00 GB

Standard GQA Cache

Total Inference VRAM

17.56 GB

Includes 1.5 GB CUDA & Activation buffers

Inference VRAM Footprint Breakdown

Weights: 14.96 GBKV Cache: 1.00 GBOverhead: 1.60 GB

Hardware Fit & Generation Speed

GPU Hardware	VRAM (GB)	Memory Bandwidth	Fit Status	Max Tokens / Sec

Training State Memory

Full Training (AdamW 16 B/p):128.5 GB

LoRA Fine-Tuning (Rank 16):15.3 GB

Compute FLOPs per Token

16.06 GFLOPs

Prefill Context FLOPs:131.5 TFLOPs

Chinchilla Training (15T tokens): 723 ZettaFLOPs

Free tool byAllMyTools.tech

Technical Reference and Manual

About Transformer Calculator: Turns Ratio, kVA & Wire Sizing

Transformers represent two technological pillars: electromagnetic machines that step electrical power across utility grids, and neural architectures driving artificial intelligence.

Physical transformers transfer energy through electromagnetic induction. Sizing distribution units requires balancing core saturation, current density, winding thermal limits, and fault currents.

In machine learning, transformer models process tokens via self-attention. Estimating VRAM across quantization formats and modeling hardware bandwidth prevents out-of-memory errors.

Combine electrical designs with the electricity cost calculator and inspect networking with the subnet calculator.

Step-by-step practical guide

Select your domain: Electrical Power Transformer Lab or AI / LLM Transformer Engine.
For electrical systems, pick a preset (25 kVA Residential or 75 kVA Dyn11) or enter custom voltages, kVA, frequency, and flux density.
Review computed turns per volt, primary and secondary turns, conductor AWG gauge, and short-circuit current.
For machine learning, choose an architecture preset (LLaMA 3 8B, 70B, Mixtral 8x7B, or DeepSeek-V3 671B) or enter custom layer parameters.
Adjust context length, batch size, and quantization precision (FP16, FP8, NVFP4, INT4) to inspect memory allocation and generation speeds.
Copy engineering summaries or download structured calculation logs.

Engineering specifications and execution details

Turns ratio: Primary Phase Voltage / Secondary Phase Voltage = Primary Turns / Secondary Turns.

Three-phase apparent power: S = sqrt(3) * Line Voltage * Line Current / 1000 (kVA).

Core EMF equation: E = 4.44 * Frequency * Turns * Bmax * Core Area.

Turns per volt (TPV): 10000 / (4.44288 * Frequency * Bmax * Core Area in cm2).

Symmetrical fault current: Full Load Current / (Percent Impedance / 100).

KV cache memory: Layers * Batch Size * Context Length * Elements per Token * Bytes per Scalar.

Decoding speed ceiling: Memory Bandwidth (GB/s) / Weights VRAM (GB).

Data privacy and memory isolation

All calculations run entirely in your web browser using client-side JavaScript.

No proprietary winding dimensions or private neural network weights leave your local computer.

Common production use cases

Electrical engineers sizing commercial distribution transformers, checking vector groups, and calculating short-circuit ratings.
Power designers computing turns per volt, magnetic core areas, and skin depth for high-frequency converters.
Machine learning engineers planning GPU cluster memory footprints, context limits, and local quantization tiers.
System architects evaluating tokens-per-second ceilings across Blackwell B200, H100, and RTX 4090 accelerators.

Edge cases and operational boundaries

High-frequency core saturation: Operating silicon steel above 1 kHz causes rapid overheating. Switch-mode supplies require ferrite cores with Bmax below 0.3 Tesla.
AC skin depth penetration: When conductor diameter exceeds twice skin depth, alternating current concentrates at surfaces. Windings require parallel strands or Litz wire.
Extreme context memory scaling: Standard attention caches grow linearly with sequence length. Multi-Head Latent Attention compresses cache memory up to 56 times.

Authoritative Standards & Scientific References

IEEE C57.12.00 Standard for Distribution and Power Transformers
Industry benchmark specifying ratings, impedance, and thermal limits for power transformers.
Vaswani et al. Attention Is All You Need (NeurIPS 2017)
Seminal research introducing the transformer architecture and multi-head attention mechanisms.

Frequently Asked Questions

What is the difference between voltage ratio and turns ratio in 3-phase transformers?

In single-phase units, terminal voltage ratio equals winding turns ratio. In three-phase units, line voltage ratio differs from turns ratio depending on Delta or Wye configurations. A 480V Delta to 208V Wye unit has a line ratio of 2.31 but a turns ratio of 4.0 because each secondary phase winding sees 120V.

How does Multi-Head Latent Attention (MLA) reduce KV cache VRAM in LLMs?

Standard attention stores separate Key and Value tensors for every layer and token. Multi-Head Latent Attention compresses Keys and Values into a single low-dimensional latent vector. During text generation, keys and values are generated dynamically, cutting cache VRAM up to 56 times.

How do I calculate transformer kVA from line voltage and load amperes?

For single-phase, multiply line voltage by current and divide by 1,000. For three-phase, multiply line voltage by line current, multiply by the square root of three (1.732), and divide by 1,000. A 480V three-phase circuit drawing 90 amperes requires roughly 75 kVA capacity.

What is the difference between total parameters and active parameters in MoE models?

Total parameters represent all weights stored in GPU memory. Active parameters represent the subset executed for an individual token. DeepSeek-V3 holds 671 billion parameters in VRAM but executes only 37 billion per token, matching the throughput of a smaller model.

How does temperature affect transformer copper winding resistance and losses?

Copper resistance increases by approximately 0.393 percent per degree Celsius. Operating at 75 deg C increases resistance by 21.6 percent compared to 20 deg C ambient. At 115 deg C limits, resistance rises 37.3 percent, increasing conductor losses.

How much VRAM does a 70B parameter LLM need across FP16, INT4, and NVFP4?

At 16-bit half precision (FP16), a 70B model requires about 131 GB of VRAM for weights. 4-bit quantization (INT4) reduces this to roughly 36 GB. Native NVIDIA NVFP4 requires about 37 GB, allowing the model to fit on two 24GB GPUs.

What is percentage impedance (%Z) and how does it determine short-circuit fault current?

Percentage impedance is the primary voltage percentage needed to circulate full-load current through a shorted secondary. Dividing full-load current by the impedance decimal yields symmetrical fault current. A 75 kVA unit with 4.5% impedance and 208A FLA yields 4,626A fault current.

Why does high frequency in SMPS transformers require ferrite cores and Litz wire?

Above 20 kHz, silicon steel suffers high hysteresis and eddy current losses. Ferrite offers high electrical resistivity to limit heating. Simultaneously, the AC skin effect crowds current to conductor edges, requiring multi-strand insulated Litz wire.

How do you calculate training compute FLOPs using the Chinchilla scaling law?

Forward passes consume two floating point operations per parameter, and backward passes consume four, totaling six operations per parameter per token. Training an 8B model on 15 trillion tokens requires approximately 723 ZettaFLOPs of total compute.