Free Google Professional-Machine-Learning-Engineer Actual Exam Questions - Question 7 Discussion
preliminary experiments running on your on-premises CPU-only infrastructure were encouraging,
but have slow convergence. You have been asked to speed up model training to reduce time-to-
market. You want to experiment with virtual machines (VMs) on Google Cloud to leverage more
powerful hardware. Your code does not include any manual device placement and has not been
wrapped in Estimator model-level abstraction. Which environment should you train your model on?
Maybe B could work if you want a big speed boost, but since the code isn’t set up for multi-GPU, it might not use all 8 efficiently. D’s just beefier CPU, so probably slower than GPU options.
It’s C since pre-installed GPU support beats manual setup and unoptimized multi-GPU use.
Good point about the manual setup in B slowing things down. I’d pick C since it has a GPU ready with libraries pre-installed, making it easier to start training fast without messing with device placement. C
Maybe C makes more sense since it has pre-installed libraries and a GPU ready to go, avoiding the hassle of manual setup in B. Plus, without device placement code, using all 8 GPUs in B might not speed things up much.
Option B, 8 GPUs, but manual setup might slow things without device placement code.
If the code doesn't handle device placement, jumping straight to 8 GPUs with manual setup (B) might cause headaches. Options with pre-configured environments like C could save time despite fewer GPUs. Could that balance out?
B; 8 GPUs should speed training way more despite manual setup hassle.
B seems like the best shot at speed here. Even if it means doing manual installs, having access to 8 GPUs should dramatically cut training time compared to just one GPU in C. The overhead of managing devices without Estimator might be tricky, but the raw hardware boost from multiple GPUs likely outweighs that. D’s just too CPU-focused, and A’s use of a single TPU might cause compatibility headaches without manual device placement. So, if the goal is to reduce training time quickly and you can handle setup, B makes the most sense.
Option D seems less ideal because more CPU cores alone won’t match the speed gains from GPU acceleration, especially for CNN training. Since the code lacks manual device placement and Estimator wrapping, managing multiple GPUs like in B could cause headaches or require extra code changes. So D might improve CPU performance but won’t offer the same leap as GPUs. Sticking with a single GPU environment like in C looks safer and faster without the multi-GPU complexity or manual setup needed in A or B.
Not B, single GPU in C avoids multi-GPU code issues and boosts speed nicely.
If the code isn’t set for multi-GPU, option B might just complicate things with no real speedup. Option D’s just beefy CPUs, so probably not much faster than your current setup. Could C be the best balance here?
Probably C since it matches the code setup and avoids multi-GPU hassles.
Probably C here too, since it has GPU support and all dependencies ready. With no manual device setup or Estimator wrapping, jumping to 8 GPUs (B) could cause issues or wasted resources.
C—pre-installed GPU support without multi-GPU complexity; good middle ground.
B imo, 8 GPUs should speed up training much more than just 1 GPU or CPU-only options.
C imo, since it comes with pre-installed libraries and a GPU to speed up training without extra setup. Using GPUs makes more sense than just beefing up CPUs for CNNs.