Federated Learning Communication Efficiency (2D)

Compression Method

Top-K keep ratio 30%

Quantization bits 8 bit

Non-IID client skew 30%

Live Readouts

Round

Loss ‖w − w*‖²

—

Payload / round

—

Bandwidth saved

Loss (last 60 rounds)—

How it works

10 clients each hold a locally-shifted copy of an 8-dimensional model target w*. Each round, client k takes one local gradient step toward its own target, producing an update Δw_k. Instead of sending the full 32-bit update, it is compressed before upload:

Top-K sparsification (with error feedback):
  c = Δw_k + e_k            (add carried-over residual)
  keep the top-K% components by |value|, zero the rest
  e_k ← c − compressed(c)    (unsent part is remembered, not lost)

Quantization (b bits, QSGD-style):
  step = 2·range / 2^b
  q_i = round(v_i / step) · step, clipped to range

Server: w ← w + (1/N) Σ_k compressed(Δw_k)

Top-K — only the largest-magnitude gradient components are transmitted; the rest is deferred via an error-feedback buffer so no information is discarded, just delayed, which is why Top-K still converges even at aggressive ratios.
Quantization — every component is still sent, but rounded to a coarse grid of 2^bits levels, trading per-value precision for fewer bits per value.
Non-IID skew — how far each client's local optimum drifts from the shared target; higher skew makes individual client updates disagree more, which compounds with compression noise and slows convergence — the real challenge production federated systems (e.g. FedAvg, FedProx) tune for.
Payload / round sums transmitted bytes across all 10 clients for the current method, compared against the uncompressed 32-bit baseline shown as "bandwidth saved". The bar chart below the diagram breaks this down per client each round.
Drag inside the radial diagram to rotate the client ring — purely visual, it does not affect the math.

This mirrors real communication-efficient FL research (Deep Gradient Compression, QSGD, signSGD): reducing upload bytes per round is often the dominant cost in cross-device federated learning, more than compute.

Round 0

drag to rotate

Global server model Client node Compressed update packet

Bytes sent this round, per client—

Federated Learning Communication Efficiency Simulator (2D)