$random in Verilog — Random Number Generation

Chapter 8 · Section 8.2 · System Tasks & Functions

1. The Engineering Problem

A verification engineer writes a stress test for an 8-bit ALU. The goal: hit every opcode + operand combination across thousands of cycles to catch corner-case bugs. The naive approach:

module alu_stress_tb;
    reg [7:0] a, b;
    reg [2:0] op;
    integer i;
 
    initial begin
        for (i = 0; i < 10000; i = i + 1) begin
            a  = $random;
            b  = $random;
            op = $random % 8;     // restrict to 0..7
            #1;
            // ... check result ...
        end
        $finish;
    end
endmodule

The test runs. The opcode histogram comes back skewed — opcode 0 fires twice as often as the others. The engineer suspects a stuck-bit on the opcode register. Three hours of debugging later, the ALU is fine. The bug is in $random % 8 — $random returns a signed 32-bit value, and the modulo of a negative number in Verilog is implementation-defined. Half the time the modulo result is -N, and the engineer's op register sees a wrong-signed value that wraps to a different bucket.

The right fix uses the unsigned variant:

op = $urandom_range(7, 0);   // uniform unsigned 0..7

This page covers the randomization family in depth — the $random mechanism, the seed contract, the negative-modulo trap, the $urandom / $urandom_range upgrades that fix the trap by construction, the seven distribution functions for non-uniform draws, and the reproducibility discipline every regression suite depends on.

2. Anatomy of $random

The simplest random source.

2.1 Syntax

integer r;
r = $random;          // bare call — uses an internal global seed
r = $random(seed);    // seeded call — uses the explicit seed, updates it

The return type is a 32-bit signed integer (-2³¹ to 2³¹-1).

2.2 Semantics

• Pseudorandom, not truly random. The same seed produces the same stream — this is what makes reproducible regressions possible.
• Signed return. The bare $random call can return negative values; this is the source of the % N distribution-bias trap.
• Seeded form mutates the seed. $random(seed) advances the seed in place — the variable's value changes after each call, walking through the pseudorandom stream.
• Unseeded form uses a global hidden seed. All $random calls without an explicit seed share one global stream; the stream's initial value is implementation-defined (typically 0 or a fixed constant per simulator vendor).

2.3 The signedness contract

integer r;
r = $random;
$display("r = %0d   (could be negative)", r);
$display("r = %h    (always 32 hex digits)", r);
$display("r = %b    (always 32 bits)", r);
 
// Cast to unsigned 32-bit for typical use
reg [31:0] u;
u = $random;
$display("u = %h    (high bit interpretation differs by tool)", u);

When assigned to an integer (signed), the value retains its signedness; when assigned to reg [31:0] (unsigned), the bit pattern is identical but the interpretation drops sign. The distinction matters for %d printing and for arithmetic — r % N does NOT behave like u % N when r is negative.

3. Mental Model

4. The Seed Contract

The seed is the random stream's primary key. Two forms.

4.1 The hidden global seed (bare $random)

initial begin
    integer i;
    for (i = 0; i < 5; i = i + 1)
        $display("%0d", $random);   // uses hidden global seed
end

The hidden global seed is shared across every bare $random call in the simulation. Two independent always blocks both calling bare $random interleave their consumption — the order depends on event scheduling and isn't guaranteed across runs even with no source change.

4.2 The explicit-seed form ($random(seed))

initial begin
    integer i;
    integer seed = 42;            // reproducible — same seed every run
    for (i = 0; i < 5; i = i + 1)
        $display("%0d", $random(seed));
end
 
// Output (vendor-dependent, but identical run-to-run with seed=42):
//   303379748
//  -1064739199
//   -1748305800
//   1543339977
//  -1129508905

$random(seed) walks an explicit stream. The seed variable is mutated by each call. Critically, each initial/always/task that owns its own seed variable gets its own stream — independent from the hidden global stream and from other explicit streams.

4.3 Reproducibility discipline

// In every regression-grade testbench
module my_tb;
    integer seed;
 
    initial begin
        // Read seed from command line, default to 0
        if (!$value$plusargs("SEED=%d", seed))
            seed = 0;
        $display("Using seed = %0d", seed);
 
        // ... use $random(seed) or $urandom(seed) throughout ...
    end
endmodule

Run with vsim +SEED=12345 my_tb to fix a specific seed; rerun with the same seed to reproduce any bug found. Most production regression suites log the seed of every failing run so it can be replayed bit-identically.

5. The Negative-Modulo Trap

The most common $random bug in production code.

integer r;
reg [2:0] op;
 
r = $random;             // signed: could be -2³¹ to +2³¹-1
op = r % 8;              // INTENDED: 0..7 uniform
                         // ACTUAL:    -7..7 — values 1..7 each from positive
                         //                    AND negative source ranges

The Verilog standard leaves the sign of (-7) % 8 implementation-defined. Most simulators return -7; some return +1. Either way, assigning the result to an unsigned register either wraps or sign-extends — the histogram is no longer uniform.

The historic Verilog-2001 fix:

r = $random;
if (r < 0) r = -r;       // make positive
op = r % 8;              // now uniform 0..7

But this still has bias — the value r = -2³¹ cannot be negated (overflows back to itself), so the smallest bucket gets one fewer chance.

The modern fix (Verilog-2005 / SystemVerilog):

op = $urandom_range(7, 0);   // unsigned, uniform, by construction

Always prefer $urandom_range for bounded random values.

6. $urandom and $urandom_range

Verilog-2005 added the unsigned random variants.

6.1 $urandom

reg [31:0] u;
u = $urandom;          // unsigned 32-bit, [0, 2³²-1]
u = $urandom(seed);    // seeded — same contract as $random(seed)

Returns a 32-bit unsigned integer. The signed-return trap of $random is gone — every value is in [0, 2³²-1]. The seed contract is identical.

6.2 $urandom_range

reg [7:0]  byte;
reg [3:0]  nibble;
integer    idx;
 
byte   = $urandom_range(255, 0);      // [0, 255] uniform
nibble = $urandom_range(15);          // [0, 15] uniform — min defaults to 0
idx    = $urandom_range(99, 50);      // [50, 99] uniform — inclusive both ends

Returns a uniform unsigned value in [min, max] (inclusive). Distribution is uniform by construction — no modulo arithmetic, no bias.

The argument order is (max, min) (most-significant first); the min argument defaults to 0 if omitted.

6.3 The upgrade rule

Every legacy $random % pattern has a clean $urandom_range replacement that's both shorter and bias-free.

7. Distribution Functions

IEEE 1364-2005 §17.9.2 defines seven distribution functions for non-uniform draws. All take a seed as the first argument and return an integer.

7.1 Example — Gaussian jitter stimulus

module jitter_stimulus;
    integer seed = 42;
    real    jitter_ps;
    integer i;
 
    initial begin
        for (i = 0; i < 10; i = i + 1) begin
            // Mean 0, std dev 50 ps
            jitter_ps = $dist_normal(seed, 0, 50);
            $display("cycle %0d: jitter = %0d ps", i, jitter_ps);
        end
    end
endmodule

$dist_normal(seed, 0, 50) draws an integer from a Gaussian with mean 0 and standard deviation 50. The vast majority of values land in [-150, +150] (±3σ); rare values exceed.

7.2 Example — Exponential inter-arrival times

module packet_generator;
    integer seed = 7;
    integer delay_cycles;
    integer i;
 
    initial begin
        for (i = 0; i < 5; i = i + 1) begin
            // Average 20 cycles between packets
            delay_cycles = $dist_exponential(seed, 20);
            #(delay_cycles) generate_packet();
        end
    end
endmodule

$dist_exponential(seed, 20) draws an integer from an exponential distribution with mean 20 — useful for memoryless traffic models (most packets close together, occasional long gaps).

8. Visual Explanation

Three figures cover the randomization model.

8.1 Visual A — randomization family tree

The functions at a glance.

8.2 Visual B — the negative-modulo bias

Why $random % 8 is biased and $urandom_range(7, 0) is not.

8.3 Visual C — seed contract

Hidden global vs explicit per-context streams.

9. The Reproducibility Discipline

Every working regression suite enforces three rules.

9.1 Always use an explicit seed in production tests

integer seed;
initial begin
    if (!$value$plusargs("SEED=%d", seed))
        seed = 0;          // explicit default — not the hidden global
    $display("[tb] seed = %0d", seed);
    // ... $random(seed), $urandom(seed), $urandom_range(...) ...
end

$value$plusargs("SEED=%d", seed) reads +SEED=12345 from the simulator's command line. If not provided, defaults to 0.

9.2 Log the seed on every failing run

always @(posedge clk) begin
    if (test_failed) begin
        $display("[FAIL] seed=%0d cycle=%0d ...", seed, $time);
        $finish(1);
    end
end

A failed regression-suite run that doesn't print the seed cannot be replayed. Always log it.

9.3 One seed per process / per context

// Each always block owns its own seed → independent streams.
integer seed_stim, seed_check;
initial seed_stim  = 100;
initial seed_check = 200;
 
always @(posedge clk) begin
    stim_value <= $urandom(seed_stim);
end
 
always @(posedge clk) begin
    if ($urandom_range(99, 0) < 5)   // 5% probability
        inject_glitch();
end

Sharing one seed across multiple contexts couples their streams — adding a call in one context shifts the others. Per-context seeds keep the streams independent.

10. Industry Use Cases

Five patterns where the randomization family shows up in production.

10.1 Constrained-random stimulus generation

The canonical workload. Every verification suite generates millions of random transactions:

reg [31:0] addr;
reg [31:0] data;
reg [1:0]  burst_len;
integer    seed = 42;
 
always @(posedge clk) begin
    addr      <= $urandom(seed) & 32'hFFFF_FFFC;  // word-aligned
    data      <= $urandom(seed);
    burst_len <= $urandom_range(3, 0);
end

10.2 Random delays in handshake testbenches

Inserting random delays between handshake phases exposes timing bugs:

initial begin
    repeat (100) begin
        // Random think time before driving valid
        #($urandom_range(20, 1));
        valid <= 1;
        @(posedge clk iff ready);
        valid <= 0;
    end
end

10.3 Statistical-distribution stimulus

Modelling realistic traffic patterns requires non-uniform distributions:

// Gaussian jitter on a 10 ns clock
real jitter_ps;
integer seed = 7;
 
always @(posedge nominal_clk) begin
    jitter_ps = $dist_normal(seed, 0, 25);  // mean=0, sigma=25ps
    #(jitter_ps / 1000.0) actual_clk = 1;
end

10.4 Random opcode generation for ALU / CPU verification

Cover every opcode + operand combination:

reg [7:0] a, b;
reg [2:0] op;
integer   seed = 99;
 
initial begin
    repeat (10000) begin
        a  = $urandom_range(255, 0);
        b  = $urandom_range(255, 0);
        op = $urandom_range(7, 0);
        @(posedge clk);
    end
end

10.5 Bug-replay via seed logging

When a regression finds a failure, the seed is logged with the failure record. The same test rerun with +SEED=<logged_value> reproduces the exact failure for debugging.

11. Common Mistakes

Six pitfalls that catch every verification engineer at least once.

11.1 $random % N biased distribution

The §5 trap. $random returns signed; % N is biased. Use $urandom_range(N-1, 0) instead.

11.2 Forgotten seed → irreproducible regression

// No seed handling — every run produces different results
always @(posedge clk)
    stim <= $random;

A failure on Tuesday's regression run cannot be replayed Wednesday. Always read a seed from $value$plusargs and log it.

11.3 Seed shared across contexts couples streams

integer seed = 0;
always @(posedge clk_a) x = $random(seed);  // consumes from seed
always @(posedge clk_b) y = $random(seed);  // consumes from seed
// Adding any extra $random(seed) in clk_a's block shifts clk_b's stream.

One seed per context. Independent streams stay independent.

11.4 Vendor-portable seed assumption

// Seed 42 in VCS produces sequence A
// Seed 42 in Xcelium produces sequence B — different algorithm

Pseudorandom algorithms are not standardised. Same seed across vendors produces different streams. Reproducibility is vendor-scoped.

11.5 $random in synthesisable RTL

always @(posedge clk)
    state <= $random;    // wrong tool — $random is simulation-only

$random is non-synthesisable. Synthesis tools strip it or warn. For RTL pseudorandom generators, instantiate an LFSR or a Galois polynomial — not a system task.

11.6 $urandom_range(0, 7) vs $urandom_range(7, 0)

The argument order is (max, min). Reversing produces unexpected behaviour — most simulators silently swap, but some return constants.

x = $urandom_range(0, 7);    // ambiguous — depends on tool's swap behaviour
x = $urandom_range(7, 0);    // canonical — max first, min second

Stick to (max, min). Set min = 0 to omit the second argument.

12. Debugging Lab

13. Coding Guidelines

Six rules teams converge on.

1. $urandom_range(max, min) for bounded uniform draws. Never $random % N. The unsigned variant is uniform by construction.
2. Always read the seed from $value$plusargs. Default to a known value (0 or 1). Log the seed at simulation start and on every failing event.
3. One seed per independent context. Don't share a seed across always blocks that should produce independent streams.
4. Use $random / $urandom only in testbench code. Wrap in `ifdef SIMULATION if the file is shared with RTL. Synthesisable pseudorandom generators use LFSRs, not system tasks.
5. Pick the right distribution function. Uniform → $urandom_range; Gaussian → $dist_normal; Exponential (memoryless intervals) → $dist_exponential. Most testbenches use uniform; reach for $dist_* only when the workload demands a specific shape.
6. Document the seed source. Comment whether a seed is a fixed magic number (reproducible smoke test), a $plusargs argument (production regression), or a global counter (multi-test independence).

14. Interview Q&A

15. Exercises

Three exercises that turn $random into reflex.

Exercise 1 — Spot the bias

For each snippet, identify whether the distribution is uniform or biased, and explain why.

// Snippet A
reg [3:0] r;
r = $random % 16;
 
// Snippet B
reg [3:0] r;
r = $urandom_range(15, 0);
 
// Snippet C
reg [3:0] r;
r = $urandom & 4'hF;
 
// Snippet D
reg [3:0] r;
r = $random; if (r < 0) r = -r;
r = r % 16;
 
// Snippet E
reg [6:0] r;   // 0..127
r = $random % 100;   // intend 0..99

Hints. Signed-modulo bias, power-of-two boundaries, the -2³¹ overflow edge case, modulo of non-power-of-two ranges.

Exercise 2 — Add reproducibility

Convert this irreproducible testbench into a regression-grade one (reads +SEED=, defaults to 0, logs the seed at startup and at every failing event).

module sloppy_tb;
    reg [7:0] a, b;
    reg [2:0] op;
 
    initial begin
        repeat (1000) begin
            a  = $random;
            b  = $random;
            op = $random % 8;
            @(posedge clk);
            if (golden(a, b, op) !== dut_result)
                $display("FAIL at t=%0t", $time);
        end
        $finish;
    end
endmodule

What to produce. The rewritten module that (a) reads the seed from $value$plusargs, (b) defaults to 0, (c) uses $urandom(seed) / $urandom_range, (d) logs the seed at startup, (e) logs the seed AND the failing-cycle's randomised inputs at every FAIL event.

Exercise 3 — Pick the right distribution

For each requirement, recommend the best function call.

What to produce. Function name and exact arguments for each.

16. Summary

$random is Verilog's pseudorandom number generator — a 32-bit signed pseudorandom integer drawn from an internal stream. The signed-return trap is what motivates the entire upgrade chain to $urandom and $urandom_range.

The four functions:

• $random — 32-bit signed; legacy form; uses hidden global seed unless $random(seed) form.
• $urandom — 32-bit unsigned; modern replacement for $random.
• $urandom_range(max, min) — uniform unsigned in [min, max]; the working engineer's default.
• $dist_* family — seven non-uniform distribution functions ($dist_normal, $dist_exponential, $dist_poisson, $dist_chi_square, $dist_t, $dist_erlang, $dist_uniform).

The seed contract:

• Bare $random uses a hidden global seed shared across all call sites. Not reproducible by default.
• $random(seed) uses an explicit seed variable, mutated in place. Reproducible by construction.
• Each independent context should own its own seed variable — sharing couples streams.
• Pseudorandom algorithms are not standardised — same seed across vendors produces different streams. Reproducibility is vendor-scoped.

The reproducibility discipline:

• Read seed from $value$plusargs("SEED=%d", seed) with a known default.
• Log the seed at simulation start and on every failing event.
• One seed per context to keep streams independent of unrelated code changes.

The upgrade rule:

The day-to-day discipline:

• $urandom_range(max, min) for bounded uniform. Never $random % N.
• Read the seed from a plusarg, default to a known value, log it. Replayable regressions only.
• One seed per context. Independent streams.
• $random is not synthesisable. Wrap in \ifdef SIMULATION`. Use LFSRs for RTL pseudorandom.
• Reach for $dist_* when the workload demands a non-uniform shape. Uniform is overwhelmingly the common case.

The next three sub-pages close Chapter 8: 8.3 Time Functions covers $time / $realtime / $stime, 8.4 Conditional Simulation covers the assertion-and-severity layer ($error, $warning, $fatal), and 8.5 Simulation Control closes the chapter with $finish / $stop / $exit.

Related Tutorials

• System Tasks & Functions — Chapter 8; the parent overview of the eight functional families.
• $display vs $monitor vs $strobe vs $write — Chapter 8.1; the display family that prints randomised stimulus.
• Conditional Compilation — Chapter 7.4; the \ifdef SIMULATION` discipline that gates every random call.