Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.
C++
// read-bandwidth probe: sum of a 256 MB buffer with float4 loads, a few launch shapes
#include <cstdio>
#include <cuda_runtime.h>
__global__ void rd(const float4* __restrict__ p, size_t n, float* out) {
float s = 0;
for (size_t i = (size_t) blockIdx.x * blockDim.x + threadIdx.x; i < n; i += (size_t) gridDim.x * blockDim.x) {
float4 v = p[i];
s += v.x + v.y + v.z + v.w;
}
if (s == 1234.5f) out[0] = s;
}
__global__ void rd_unroll(const int4* __restrict__ p, size_t n, int* out) {
int s = 0;
size_t stride = (size_t) gridDim.x * blockDim.x;
size_t i = (size_t) blockIdx.x * blockDim.x + threadIdx.x;
for (; i + 3 * stride < n; i += 4 * stride) {
int4 a = p[i], b = p[i + stride], c = p[i + 2 * stride], d = p[i + 3 * stride];
s += a.x ^ a.y ^ a.z ^ a.w ^ b.x ^ b.y ^ b.z ^ b.w ^ c.x ^ c.y ^ c.z ^ c.w ^ d.x ^ d.y ^ d.z ^ d.w;
}
if (s == 1234) out[0] = s;
}
int main() {
const size_t bytes = 256u << 20, n = bytes / 16;
void* p; float* o;
cudaMalloc(&p, bytes); cudaMalloc(&o, 4);
cudaMemset(p, 1, bytes);
cudaEvent_t a, b; cudaEventCreate(&a); cudaEventCreate(&b);
for (int blocks : {112, 224, 448, 896, 3584, 14336}) {
for (int k = 0; k < 2; ++k) {
cudaEventRecord(a);
for (int it = 0; it < 20; ++it) {
if (k == 0) rd<<<blocks, 256>>>((const float4*) p, n, o);
else rd_unroll<<<blocks, 256>>>((const int4*) p, n, (int*) o);
}
cudaEventRecord(b); cudaEventSynchronize(b);
float ms; cudaEventElapsedTime(&ms, a, b);
printf("%s blocks %5d: %.1f GB/s\n", k == 0 ? "loop " : "unroll", blocks, 20.0 * bytes / (ms / 1000) / 1e9);
}
}
return 0;
}