[Driver] Torchbench training performance on Rolling is slower than LTS

[mengfei25]

🐛 Describe the bug

Observed with torchbench training performance on Rolling driver and LTS driver, looks like Rolling is slower than LTS.
Overall it is ~20% gap. The following is the < 50% models

Category	Name	Eager	Inductor
torchbench_amp_bf16_training	functorch_dp_cifar10	0.440394435	0.376553014
torchbench_amp_bf16_training	lennard_jones	0.410815577	0.406666522
torchbench_amp_bf16_training	resnet18	0.569801695	0.414078304
torchbench_amp_bf16_training	drq	0.419031071	0.416605194
torchbench_amp_bf16_training	dcgan	0.473041413	0.45275431
torchbench_amp_bf16_training	mobilenet_v3_large	0.601874092	0.458255372
torchbench_amp_bf16_training	phlippe_resnet	0.565532503	0.474469922
torchbench_amp_bf16_training	basic_gnn_gcn	0.837236487	0.474493124
torchbench_amp_bf16_training	basic_gnn_gin	0.67058437	0.475682344
torchbench_amp_bf16_training	speech_transformer	0.524068457	0.497168054
torchbench_amp_bf16_training	soft_actor_critic	0.497953362	0.498692661
torchbench_amp_fp16_training	drq	0.40930045	0.399965199
torchbench_amp_fp16_training	phlippe_resnet	0.546009712	0.432983013
torchbench_amp_fp16_training	nanogpt	0.461748452	0.448994252
torchbench_amp_fp16_training	lennard_jones	0.463453479	0.455919189
torchbench_amp_fp16_training	resnet18	0.58230154	0.46118426
torchbench_amp_fp16_training	dcgan	0.46315022	0.468631482
torchbench_amp_fp16_training	mnasnet1_0	0.696290041	0.489389997
torchbench_amp_fp16_training	LearningToPaint	0.761002479	0.4967825
torchbench_bfloat16_training	lennard_jones	0.3521912	0.363774254
torchbench_bfloat16_training	phlippe_resnet	0.534909223	0.408322118
torchbench_bfloat16_training	drq	0.430409868	0.429119883
torchbench_bfloat16_training	nanogpt	0.440041578	0.430549223
torchbench_bfloat16_training	mobilenet_v3_large	0.598219707	0.450442757
torchbench_bfloat16_training	soft_actor_critic	0.46347127	0.452779549
torchbench_bfloat16_training	phlippe_densenet	0.569334382	0.476015257
torchbench_bfloat16_training	resnet18	0.617031303	0.490115099
torchbench_bfloat16_training	functorch_dp_cifar10	0.609626518	0.498543827
torchbench_float16_training	lennard_jones	0.37388514	0.35987522
torchbench_float16_training	phlippe_resnet	0.468794127	0.400084215
torchbench_float16_training	dcgan	0.382458895	0.416399596
torchbench_float16_training	functorch_dp_cifar10	0.540425171	0.417395998
torchbench_float16_training	mobilenet_v3_large	0.583812604	0.434171476
torchbench_float16_training	resnet18	0.570007512	0.441501998
torchbench_float16_training	drq	0.457905383	0.453155682
torchbench_float16_training	squeezenet1_1	0.699713952	0.455786032
torchbench_float16_training	nanogpt	0.424120811	0.469937885
torchbench_float16_training	soft_actor_critic	0.473865774	0.476388316
torchbench_float16_training	timm_efficientnet	0.711824661	0.490151596
torchbench_float32_training	functorch_dp_cifar10	0.517967004	0.411734615
torchbench_float32_training	nanogpt	0.439565001	0.41192107
torchbench_float32_training	drq	0.414325806	0.43835296
torchbench_float32_training	phlippe_resnet	0.54522133	0.45738061

Versions

Device: PVC 1100
Driver: Rolling 24.39.31294 / LTS 23.43.27642.52
PyTorch: 20241202 nightly wheel