amd-experiments / results /bench_kernels.json
Bobliu's picture
Upload folder using huggingface_hub
a30d21b verified
Raw History Blame Contribute Delete
3.33 kB
{
"device": "AMD Instinct MI300X VF",
"n_vectors": 65536,
"head_dim": 128,
"python_wrapper": {
"compress": {
"op": "tq3_compress",
"n_vectors": 65536,
"median_us": 3121.65,
"throughput_gbs": 11.8,
"vectors_per_us": 20.99,
"note": "rotation (GEMM\u2192MFMA) + centroid lookup + bitpack"
},
"decompress": {
"op": "tq3_decompress",
"n_vectors": 65536,
"median_us": 632.55,
"throughput_gbs": 58.4,
"vectors_per_us": 103.61,
"note": "bitunpack + centroid lookup + inverse rotation (GEMM\u2192MFMA)"
},
"fused_dot": {
"op": "tq3_fused_dot",
"n_queries": 64,
"n_kv": 65536,
"median_us": 609.95,
"throughput_gbs": 33.1,
"note": "bitunpack + centroid lookup + dot product (GEMM\u2192MFMA)"
},
"fp16_matmul": {
"op": "fp16_matmul (reference)",
"n_queries": 64,
"n_kv": 65536,
"median_us": 32.8,
"throughput_gbs": 767.7
}
},
"standalone_binary": {
"binary_results": {
"Memory": {
"throughput_gbs": 5300.0,
"source": "standalone_binary"
},
"For": {
"throughput_gbs": 5300.0,
"source": "standalone_binary"
}
},
"raw_output": "=== TurboQuant MI300X Kernel Benchmark ===\nGPU: AMD Instinct MI300X VF | warpSize=64\nn_vectors=65536 n_warmup=50 n_iters=50\n\nInput: 33.6 MB (float32, 65536 vectors \u00d7 128 dim)\nTQ3 out: 3.4 MB (9.8\u00d7 compression)\nTQ4 out: 4.5 MB (7.5\u00d7 compression)\n\nKernel | n | min_ms | avg_ms | BW_in (GB/s) | BW_out (GB/s)\n\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n TQ3 quantize (compress) | n= 65536 | min=3.155 ms | avg=3.193 ms | BW_in=10.6 GB/s | BW_out=1.1 GB/s\n TQ3 dequantize (decompress) | n= 65536 | min=0.169 ms | avg=0.177 ms | BW_in=20.1 GB/s | BW_out=198.1 GB/s\n TQ4 quantize (compress) | n= 65536 | min=3.160 ms | avg=3.206 ms | BW_in=10.6 GB/s | BW_out=1.4 GB/s\n TQ4 dequantize (decompress) | n= 65536 | min=0.172 ms | avg=0.180 ms | BW_in=25.9 GB/s | BW_out=194.7 GB/s\n TQ3 fused dot (1 query) | n= 65536 | min=0.037 ms | avg=0.038 ms | BW_in=93.1 GB/s | BW_out=7.2 GB/s\n\n=== Theoretical MI300X limits ===\n Memory bandwidth: 5300 GB/s\n For TQ3 compress (FP32 in = 33.6 MB): theoretical 5300 GB/s bounded = 0.006 ms min\n For TQ3 decompress (TQ3 in = 3.4 MB): theoretical 5300 GB/s bounded = 0.0006 ms min\n"
},
"memory_analysis": {
"n_vectors": 65536,
"fp16_bytes": 16777216,
"fp8_bytes": 8388608,
"int4_bytes": 4194304,
"tq3_bytes": 3407872,
"tq4_bytes": 4456448,
"fp8_ratio": 2.0,
"int4_ratio": 4.0,
"tq3_ratio": 4.92,
"tq4_ratio": 3.76
}
}