No. increase the block size to 384KB does not help scrypt algorithm. Litecoin with scrypt parameters N=1024, R=P=1, that only requires 128KB memory space for single one compuation. And they are already fully aligned, not sparse array. Modeling small data in larger memory space means poor locality.
The key to speedup scrypt is parallelism, because the mining process is to find a valid nonce among the very large solution space. So if the memory is the bottleneck, CPU needs to wait for the completion of memory read/write operations before continue current hashing computation, but in fact, CPU's ALU (or compute units) can keep going and executing other instructions of other hashing computation threads which do not require memory access or whose data has ready now. That's the key to introduce SIMD. So CPU AVX performance is better than SSE/MMX, and newer AVX-2 can achieve better performance.
In GPU, there are many computation units, it does not like CPU only has 2/4/8/... cores. GPU has many streaming processors and scheduling hardware. When a computation is waiting for the completion of GDDR VRAM access, it can schedule the ALU hardware to do other computations.
In fact, current GPU mining has the best memory block size configuration per one scrypt hashing, the lookup-gap is 2. It means only 64KB per one scrypt operation. Because in CPU or GPU, accessing the global memory is very expansive, and scrypt does not have fully linear memory access pattern. (it accesses vectorized data among index 1 to 1024 randomly). So fewer memory access can also speedup the total computation time, but the overhead is that GPU needs to compute data that did not save to memory before.
For example, GPU computes vectorized data from 1 to 1024, but only saves 1,3,5,7,... So when scrypt algorithm needs data of index=4, GPU needs to retrieve data of index=3 and re-compute to gain the data for index=4. But the computation for index=4 in private memory (GPRs, General Purpose Registers, GPU has plenty of them) is very fast.
But the performance does not always get improved by increasing the look-up gap. Although the memory access are significantly reduced, but GPU needs to do redundant job to re-compute the data frequently. For example, when look-up gap = 3, GPU only saves 1,4,7,... But when scrypt needs data of index=6, it needs to retrieve data of index=4 and do the computation twice.