Physical Layer (PHY)

Posted on Edited on In

image-20260331225337870

Sublayer Layer Type Primary Responsibility Key Processes
PCS Digital Data Preparation Encoding (64b/66b), Scrambling, Alignment
PMA Mixed-Signal Serialization & Timing SerDes, Clock Recovery, Data Framing
PMD Analog/Physical Medium Interface Signal Conversion (Optics/Electrical), MDI Interface

RX Elastic Buffer

Joe Winkles, Elastic Buffer Implementations in PCI Express Devices [pdf]

wikibooks, Clock and Data Recovery/Buffer Memory (Elastic Buffer)/Cascades of Buffers and CDRs, delays and tolerance [link]

image-20260402000348140

bridge between Recovered Clock Domain and Local Clock Domain

image-20260331221404321

image-20260331231925843

The periodic slips (under or overflows) depend on buffer depth \(N\) and is approximately given by \[ T_{per} = \frac{N/2 - 1}{\Delta f} \]

image-20260402232306002

1
2
3
f0 = 1;
f1 = f0*(1+300e-6); f2 = f0*(1-300e-6);
N_t0 = 1/(f1 - f2);  % 1.6667e+03

image-20260402232626604

Lane-to-Lane Skew

TODO ๐Ÿ“…

First In First Out (FIFO)

Clifford E. Cummings, Simulation and Synthesis Techniques for Asynchronous FIFO Design with Asynchronous Pointer Comparisons [pdf]

TODO ๐Ÿ“…

Scrambler

TODO ๐Ÿ“…

TODO ๐Ÿ“…

Viterbi-based MLSD

M. Emami Meybodi, H. Gomez, Y. -C. Lu, H. Shakiba and A. Sheikholeslami, "Design and Implementation of an On-Demand Maximum-Likelihood Sequence Estimation (MLSE)," in IEEE Open Journal of Circuits and Systems, vol. 3, pp. 97-108, 2022 [https://sci-hub.jp/10.1109/OJCAS.2022.3173686]

Zaman, Arshad Kamruz (2019). A Maximum Likelihood Sequence Equalizing Architecture Using Viterbi Algorithm for ADC-Based Serial Link. Undergraduate Research Scholars Program. Available electronically from [https://hdl.handle.net/1969.1/166485]

S. Song, K. D. Choo, T. Chen, S. Jang, M. P. Flynn and Z. Zhang, "A Maximum-Likelihood Sequence Detection Powered ADC-Based Serial Link," in IEEE Transactions on Circuits and Systems I: Regular Papers, vol. 65, no. 7, pp. 2269-2278, July 2018 [https://sci-hub.jp/10.1109/TCSI.2017.2775619]

Vineel Kumar Veludandi. Maximum likelihood sequence estimation (MLSE) using the Viterbi algorithm [https://github.com/vineel49/mlse]

David Banas, Keysight, DesignCon 2026, Tutorial โ€“ Understanding the Viterbi Decoder

A. C. Singer, N. R. Shanbhag and H. -m. Bae, "Electronic dispersion compensation," in IEEE Signal Processing Magazine, vol. 25, no. 6, pp. 110-130, November 2008 [https://shanbhag.ece.illinois.edu/publications/singer-spm-2008.pdf]

โ€”, ISSCC2007 T10: Fundamentals of Electronic Dispersion Compensation (EDC)

image-20260401215526089

Maximum-Likelihood Sequence Detection performed by Viterbi on the ISI trellis induced by the channel

image-20260401214642821

Simple soft-decision voting

image-20260401223623584

image-20260401230936331

Simple soft-decision voting can use more information than a slicer, but without a trellis/path-consistency rule, it can create impossible or unstable decision histories.

Viterbi Decoder

image-20260531002612870

image-20260531002943187

Pre-cursor \(h_{-1}=0.3\), Current (main) cursor \(h_0=0.8\), Post-cursor \(h_1=0.2\)

image-20260531084027728

image-20260531002828300

trellis depth = N = 3 (consistent with 8 states); the 4 is the number of observation samples in that particular walk-through, i.e. the time axis โ€” a different thing that unfortunately shares the label "N" on the slide

David Banas, PyBERT [https://github.com/capn-freako/PyBERT/blob/master/src/pybert/models/viterbi.py]

As currently built, ViterbiDecoder_ISI models post-cursor ISI only. _states[ix][-1] selecting the last element as the current symbol is consistent with that

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
# https://github.com/capn-freako/PyBERT/blob/master/src/pybert/models/bert.py

def my_run_simulation(self, initial_run: bool = False, update_plots: bool = True,
                      aborted_sim: Optional[Callable[[], bool]] = None):
        if self.rx_viterbi_fec:

        else:
            self.status = "Running Viterbi..."
            L = self.L
            sigma = self.rn
            decoder = ViterbiDecoder_ISI(L, N, sigma, pulse_resp_samps)
            path = decoder.decode(list(sig_samps), dbg_dict=self.dbg_dict_viterbi)
            _states = decoder.states
            symbols_viterbi = list(map(lambda ix: _states[ix][-1], path))
            bits_out_viterbi = sum(list(map(lambda ss: dfe.decide(ss)[1], symbols_viterbi)), [])

        n_errs_viterbi, bit_errs_viterbi = calc_ber(bits_out_viterbi)
  • DFE (kept light): recover timing, slice levelsโ†’bits, and optionally trim the far ISI tail beyond the trellis window โ€” but not aggressively cancel the near post-cursors.
  • Viterbi: handle the near post-cursor ISI (the cursor + Nโˆ’1 taps) that was intentionally left in, via MLSE โ€” which tolerates a closed eye where the DFE's hard decisions would error-propagate
Group Members When set Role
Static model _states, _expecteds, _trans, _sigma, _v_prob once, in __init__ the time-invariant channel/decoder description
Dynamic state _trellis mutated every step_trellis the evolving survivor metrics/back-pointers

All three derive purely from the constructor args (L, N, pulse_resp_samps):

  • _states โ€” the Lแดบ symbol-window combinations (all_combs). Fixed because the alphabet and memory depth don't change.
  • _expecteds โ€” the noiseless expected sample per state, ฮฃ h[n]ยทs[-(n+1)]. Fixed because it bakes in pulse_resp_samps (the ISI taps).
  • _trans โ€” the shift-register adjacency (which state can follow which), as a row-normalized PMF. Fixed because the trellis topology is structural, independent of data.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
# https://github.com/capn-freako/PyBERT/blob/master/src/pybert/models/viterbi.py

class ViterbiDecoder_ISI(ViterbiDecoder[State_ISI, float]):
    """
    Viterbi decoder using ISI to define observation probabilities.
    """

    # pylint: disable=too-many-locals
    def __init__(self, L: int, N: int, sigma: float, pulse_resp_samps: Rvec):
        """
        Args:
            L: Number of symbol voltage levels.
            N: Number of symbols per state.
            sigma: Standard deviation of Gaussian voltage noise (V).
            pulse_resp_samps: Upstream channel pulse response samples,
                one per UI, beginning with cursor (V).
                (Must have length >= ``N``!)

        """

vs = np.linspace(-2, 2, 4_000)        # voltage-error grid: โˆ’2..+2 V, 4000 pts โ†’ ~1 mV spacing
v_prob = sp.interpolate.interp1d(
    vs,
    [1e-3 * np.exp(-(v**2)/(2*sigma**2)) / np.sqrt(TWOPI*sigma**2) for v in vs],
    bounds_error=False, fill_value=0)

# bounds_error=False, fill_value=0 โ†’ any error magnitude beyond ยฑ2 V returns probability 0 (treated as impossible โ€” way outside the noise distribution)

    @property
    def v_prob(self):  # pylint: disable=missing-function-docstring
        return self._v_prob

    def prob(self, s: int, x: float) -> float:
        """
        Probability of state at index ``s`` given observation ``x``.
        """
        return self.v_prob(x - self.expectation(s))

\[ \mathcal{N}(v;0,\sigma^2)=\dfrac{1}{\sqrt{2\pi\sigma^2}}e^{-v^2/2\sigma^2} \]

v_prob(ฮด) returns the probability (density) of seeing a noise voltage error ฮด under a zero-mean Gaussian with standard deviation sigma. It's the emission probability of the Viterbi algorithm.

In short: v_prob is the decoder's noise model โ€” a fast Gaussian-likelihood lookup that converts "how far is this sample from what state s predicts?" into "how probable is that?", which is exactly the data-driven term the Viterbi trellis needs

1
2
3
4
5
6
# https://github.com/capn-freako/PyBERT/blob/master/src/pybert/models/viterbi.py

class ViterbiDecoder_ISI(ViterbiDecoder[State_ISI, float]):
    def __init__(self, L: int, N: int, sigma: float, pulse_resp_samps: Rvec):
        # Build initial trellis.
        trellis = [[(1 / num_states, 0)] * num_states] * N
size set by
depth (len(trellis), # columns) N N UIs (cursor + Nโˆ’1 post-cursors)
width (len(trellis[-1]), # rows) Lแดบ = num_states L and N via all_combs
  1. Length of returned list gives trellis depth
  2. Length of all inner lists should equal len(states).
  3. Each location in the trellis matrix contains the probability and previous state index for the corresponding state ( probability , prev_state_index )
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
# https://github.com/capn-freako/PyBERT/blob/master/src/pybert/models/viterbi.py

class ViterbiDecoder(ABC, Generic[S, X]):
    """
    Abstract definition of a Viterbi decoder.
    """
    
    def step_trellis(self, x: X, priming: bool = False) -> int:
    """
    Shift the trellis one column left, using the given observation sample.

    Args:
        x: The new observation sample.

    Keyword Args:
        priming: Don't perform backtrace when True.
            Default: False

    Returns:
        The decided state index of the exiting (i.e. - leftmost) column.
    """

    trellis = self.trellis
    num_states = len(trellis[-1])

    # Shift trellis contents one column left.
    for col in range(len(trellis) - 1):
        trellis[col] = trellis[col + 1]

    # Calculate maximum state probabilities, along w/ previous state, for new rightmost column.
    probs = np.zeros(num_states)
    prevs = np.arange(num_states)
    for r in range(num_states):
        new_probs = np.array(
            [0 if self.trans[r][s] == 0
             else trellis[-1][r][0] * self.trans[r][s] * self.prob(s, x)
             for s in range(num_states)])
        prevs = np.where(new_probs > probs, [r] * num_states, prevs)
        probs = np.maximum(new_probs, probs)

    probs /= probs.sum()
    trellis[-1] = list(zip(probs, prevs))

    prev = 0
    if not priming:
    	prev = self.path[0]

    return prev

The core Viterbi recursion trellis[-1][r][0] * self.trans[r][s] * self.prob(s, x) โ€” the candidate path metric for arriving at current state s through predecessor r \[ P(r \to s \mid x) = \underbrace{\text{trellis}[-1][r][0]}{\text{(1) survivor prob of } r} \times \underbrace{\text{trans}[r][s]}{\text{(2) transition } r\to s} \times \underbrace{\text{prob}(s, x)}_{\text{(3) emission of } s} \]

  • trellis[-1][r][0] โ€” the survivor probability of the predecessor state r
  • self.trans[r][s] โ€” the transition probability r โ†’ s
  • self.prob(s, x) โ€” the emission probability of current state s given observation x

\[ P(\text{path ending } r{\to}s,\ \text{obs}=x) = P(r)\cdot P(s\mid r)\cdot P(x\mid s) \]

decode() has three phases

Phase Samples Decisions emitted?
Prime samps[0 : N] No โ€” just fill the depth-N window
Run samps[N : ] Yes โ€” one decided symbol per step
Purge (none) Flush the Nโˆ’1 symbols still held in the window

path performs the back-trace (trace-back) step of the Viterbi algorithm: starting from the most-likely state in the newest column, it follows the back-pointers leftward through the trellis to recover the maximum-likelihood sequence of states held in the current window.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
# https://github.com/capn-freako/PyBERT/blob/master/src/pybert/models/viterbi.py

class ViterbiDecoder(ABC, Generic[S, X]):
    """
    Abstract definition of a Viterbi decoder.
    """

    @property
    def path(self) -> list[int]:
        """
        Maximum likelihood forward path through the trellis.

        Notes:
            1. First element in returned list corresponds to the time
            just before the first trellis column.
            2. The decided state of the final trellis column is *not* included.
        """

        trellis = self.trellis
        trellis_depth = len(trellis)

        # Starting with highest probability final state, backtrack through trellis.
        prevs = [trellis[-1][np.argmax(list(map(lambda pr: pr[0], trellis[-1])))][1]]
        for ix in range(2, trellis_depth + 1):
            prevs.append(trellis[-ix][prevs[-1]][1])
        prevs.reverse()
        return prevs

At the end you drain the Nโˆ’1 symbols still sitting in the trellis

1
2
3
4
5
6
7
8
9
# https://github.com/capn-freako/PyBERT/blob/master/src/pybert/models/viterbi.py

class ViterbiDecoder(ABC, Generic[S, X]):
 
    def decode(self, samps: list[X], dbg_dict: Optional[dict[str, Any]] = None) -> list[int]:

        # Purge the trellis.
        states.extend(self.path[1:])
        states.append(int(np.argmax(list(map(fst, trellis[-1])))))

Standalone demo

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
"""
Standalone demo: run the PyBERT Viterbi (ISI) decoder on a synthetic link.

Builds a ``ViterbiDecoder_ISI`` for a 2-level (NRZ) channel with strong
post-cursor ISI, generates a noisy received waveform, decodes it via maximum
likelihood sequence estimation (MLSE), and compares the error count against a
naive symbol-by-symbol slicer.

Run (from the repository root)::

    PYTHONPATH=src python examples/run_viterbi.py

This example depends only on ``numpy``/``scipy`` -- not on the optional
``pyibisami`` dependency.
"""

import numpy as np

from pybert.models.viterbi import ViterbiDecoder_ISI

rng = np.random.default_rng(42)

# --- Link / decoder parameters ---
L = 2                                # NRZ (2 levels: -1, +1)
N = 2                                # symbols of memory per state
sigma = 0.22                         # Gaussian noise std-dev (V)
pulse_resp = np.array([1.0, 0.8])    # cursor + strong post-cursor ISI tap (V/UI)

dec = ViterbiDecoder_ISI(L=L, N=N, sigma=sigma, pulse_resp_samps=pulse_resp)
expecteds = [dec.expectation(i) for i in range(len(dec.states))]
print(f"States ({len(dec.states)}): {dec.states}")
print(f"Expected observations:     {np.round(expecteds, 3)}")
print(f"State Transition Probability matrix:\n{dec.trans}\n")

# --- Generate a random NRZ symbol stream (-1/+1) ---
n_syms = 5000
levels = np.array([-1.0, 1.0])
tx = rng.choice(levels, size=n_syms)

# --- Pass through the ISI channel + add noise to make the received waveform ---
clean = np.convolve(tx, pulse_resp)[:n_syms]
rx = clean + rng.normal(0.0, sigma, size=n_syms)

# --- Decode ---
state_idxs = dec.decode(list(rx))
# Each decoded state is an N-symbol vector; the current symbol is its last element.
decoded = np.array([dec.states[i][-1] for i in state_idxs])

# Align (decoder output corresponds to the symbol stream) and count errors.
m = min(len(tx), len(decoded))
errs = int(np.sum(tx[:m] != decoded[:m]))

# --- Compare against a naive symbol-by-symbol slicer (threshold at 0) ---
slicer = np.where(rx > 0, 1.0, -1.0)
slicer_errs = int(np.sum(tx[:m] != slicer[:m]))

idx_erros = np.where(tx[:m] != slicer[:m])[0]

print(f"Symbols decoded     : {m}")
print(f"Viterbi MLSE errors : {errs}  (BER ~ {errs / m:.3g})")
print(f"Naive slicer errors : {slicer_errs}  (BER ~ {slicer_errs / m:.3g})")

print("\nFirst 10 errors (index, original symbol, decoded symbol, slicer symbol):")
print("index: ", idx_erros[:10])
print("original symbols: ", tx[idx_erros[:10]])
print("decoded symbols: ", decoded[idx_erros[:10]])
print("slicer symbols:  ", slicer[idx_erros[:10]])


#  [0.  0.  0.5 0.5]
#  [0.5 0.5 0.  0. ]
#  [0.  0.  0.5 0.5]]

# Decoding 5000 samples with trellis depth 2 and 4 states...
# Symbols decoded     : 5000
# Viterbi MLSE errors : 0  (BER ~ 0)
# Naive slicer errors : 452  (BER ~ 0.0904)

# First 10 errors (index, original symbol, decoded symbol, slicer symbol):
# index:  [ 21  71  75  91 101 102 104 113 116 126]
# original symbols:  [-1. -1. -1.  1. -1.  1.  1. -1. -1. -1.]
# decoded symbols:  [-1. -1. -1.  1. -1.  1.  1. -1. -1. -1.]
# slicer symbols:   [ 1.  1.  1. -1.  1. -1. -1.  1.  1.  1.]

image-20260601071327138

Proakis, John G., and Masoud Salehi. Digital Communications. 5th ed. McGraw-Hill, 2008. [pdf]

image-20260601071629821

whitening filter

image-20260601072615963

Partial Response

David Johns, Partial Response and Viterbi Detection [https://www.eecg.utoronto.ca/~johns/ece1392/slides/partial_response.pdf]

Dariush Dabiri , Enabling Improved DSP Based Receivers for 100G Backplane [https://www.ieee802.org/3/bj/public/sep11/dabiri_01_0911.pdf]

Partial Response Signaling (PRS) and Maximum Likelihood Sequence Detection (MLSD) are paired together to maximize data rates in bandwidth-limited, noisy channels like fiber optics, magnetic hard drives, and high-speed backplanes

image-20260528202037244

Flit (flow control unit)

TODO ๐Ÿ“…

FBER (First Bit Error Rate)

TODO ๐Ÿ“…

1
2
FBER is like BER except that, for example, a 10-bit burst error counts as one FBER but would count as 10 BER.
The FEC in Flit Mode can correct up to a single 16-bit burst error in any given Flit.

FEC (Feed-Forward Error Correction)

Cathy Liu, Broadcom. DesignCon 2024: 200+ Gbps Ethernet Forward Error Correction (FEC) Analysis

โ€”, Broadcom, DesignCon 2026 What is FEC and how do I use it in 200G/400G/800G/1.6T Ethernet?

Leslie A. Rusch , GEL7114 Digital Communications, Module 5 - Error Correcting Codes [https://wcours.gel.ulaval.ca/GEL7114/assets/pdfs/Module5_en_1by1_1.pdf]

1
2
3
4
5
6
7
Steve Glaser, How to understand FBER?

Note that the PCIe FEC is weak as compared to Ethernet.
This is intentional. Stronger FEC codes incur higher latency penalty (because they cover more bits and require more computation).
PCIe Links are short and replays are quick. The tradeoff was made that additional delay due to an occasional replay was preferable to a constant delay due to a more powerful FEC. 
The PCIe FEC's purpose is to reduce the replay rate to an acceptable level, not eliminate replays. 
Ethernet has longer Links with longer replay times. That motivates their different tradeoff.

TODO ๐Ÿ“…

Feature Intersymbol Interference (ISI) Forward Error Correction (FEC)
Definition A signal distortion where symbols overlap. A coding technique to detect/fix bit errors.
Origin Unintentional (caused by channel physics). Intentional (added by the system designer).
Data Impact Smears pulses together, making them unreadable. Adds redundant bits to protect original data.
Primary Cause Multipath fading and limited bandwidth. Noise, interference, and signal attenuation.
Relationship Negative dependency (interference). Positive dependency (structured redundancy).
Typical Solution Equalization or Pulse Shaping (e.g., Root-Raised Cosine). Block Codes (Reed-Solomon) or Convolutional Codes (LDPC).
Goal To clean the signal before decoding. To recover the data after decoding errors.

Reed-Solomon (RS) codes

RS is a forward error-correction (FEC) algorithm used to correct data corruption after it has been received

ECC (Error Control Coding)

Takayuki Kawahara, ISSCC2007 T5: Error-Correcting Codes for Memories

TODO ๐Ÿ“…

PCIeยฎ Retimer Latency

Jay Li, Product Marketing Director, kandou, T3S02_PCIe Retimer Latency

image-20260604204510926

image-20260604072254491

image-20260604072304996

PCI Express Standards Information

CLB: Compliance Load Board (System Board)

CBB: Compliance Base Board (Add-in Card)

NCB (Non-ChangeBar): A clean version of the official specification document that contains the full text of the updated standard

CB (ChangeBar): A version that highlights only the changes and additions made from the previous version of the specification. The added or modified lines typically have a vertical bar (a changebar) in the margin for easy scanning

Base Specification: The PCI Express Base describes the architecture, interconnect attributes, fabric management, and the programming interface required to design and build systems and peripherals that are compliant with the PCI Express Specification.

CEM (Card Electromechanical) Specification: The PCI Express CEM specifications define an implementation for small form factor PCI Express cards

reference

John Swindle, PCIe 1.1 Phy Design Considerations

Hidehiro Toyoda, Shinji Nishimura, and Masato Shishikura, PMD architecture with skew compensation mechanism for parallel link [https://www.ieee802.org/3/hssg/public/nov06/nishimura_01_1106.pdf]

Kamesh Velmail, Samsung, Challenges, Complexities and Advanced Verification Techniques in Stress Testing of Elastic Buffer in High Speed SERDES IPs [pdf]

Marianne Nourzad (Intel), DesignCon 2022. Transmitter Jitter Measurement Methodologies for PAM-4 IOs