Exclude opt just for non-ampere gpus

js/web/lib/wasm/jsep/backend-webgpu.ts

@@ -94,17 +94,34 @@ const getProgramInfoUniqueKey =
       return key;
     };
 
+
+/**
+ * this class is designed for info and related helper functions of the GPU adapter in use
+ */
 export class AdapterInfo {
+  private architecture: string;
   private vendor: string;
 
   constructor(adapterInfo: GPUAdapterInfo) {
     if (adapterInfo) {
+      this.architecture = adapterInfo.architecture;
       this.vendor = adapterInfo.vendor;
     }
   }
 
-  // vendor could be intel, nvidia, amd, etc.
+  // architecture could be ampere, gen12-lp, etc.
+  isArchitecture(architecture: string): boolean {
+    if (typeof this.architecture === 'undefined') {
+      return false;
+    }
+    return this.architecture === architecture;
+  }
+
+  // vendor could be amd, intel, nvidia, etc.
   isVendor(vendor: string): boolean {
+    if (typeof this.vendor === 'undefined') {
+      return false;
+    }
     return this.vendor === vendor;
   }
 }

js/web/lib/wasm/jsep/webgpu/ops/conv.ts

@@ -148,11 +148,11 @@ const conv2d = (context: ComputeContext, inputs: readonly TensorView[], attribut
   // const hasPreluActivationWeights = false; /* TODO: add support for prelu activation weights */
   const isChannelsLast = attributes.format === 'NHWC';
   if (attributes.group !== 1) {
-    // One CI bot with NVIDIA GPU fails with below 2 cases, but we couldn't repro them with any other GPUs, including NVIDIA ones.
+    // One CI with NVIDIA GPU (ampere architecutre) fails with below 2 cases, but we couldn't repro them with any other GPUs, including NVIDIA ones.
     // [webgpu]Conv - conv - vectorize group - B
     // [webgpu]Conv - conv - vectorize group - D
-    // Disable vectorize on NVIDIA to make bots happy. BTW, no obvious perf gain with vectorize is seen on NVIDIA GPUs.
-    const enableGroupedConvVectorize = context.adapterInfo.isVendor('nvidia') ? false : true;
+    // Disable vectorize on NVIDIA ampere to make bots happy. BTW, no obvious perf gain with vectorize is seen on NVIDIA GPUs.
+    const enableGroupedConvVectorize = context.adapterInfo.isArchitecture('ampere') ? false : true;
     if (enableGroupedConvVectorize && isChannelsLast && inputs[1].dims[0] === attributes.group &&
         inputs[1].dims[1] === 1 && attributes.dilations[0] === 1 && attributes.dilations[1] === 1) {
       const outputShape = calculateOutputShape(