Update 9-non-portable-kernel-models.rst

content/9-non-portable-kernel-models.rst

@@ -523,7 +523,7 @@ To demonstrate the fundamental features of CUDA/HIP programming, let's begin wit
 
 In this case, the CUDA and HIP codes are equivalent one to one so we will only refer to the CUDA version. The CUDA and HIP programming model are host centric programming models. The main program is executed on CPU and controls all the operations, memory allocations, data transfers between CPU and GPU, and launches the kernels to be executed on the GPU. The code starts with defining the GPU kernel function called **vector_add** with attribute **___global__**. It takes three input arrays `A`, `B`, and `C` along with the array size `n`. The kernel function contains the actually code which is executed on the GPU by multiple threads in parallel.
 
-Accelerators in general and GPUs in particular have their own dedicated memory separate from the system memory (**this could change soon! see AMD MI300 and NVIDIA Hopper!**). When programming for GPUs, there are two sets of pointers involved and it's necessary to manage data movement between the host memory and the accelerator memory. Data needs to be explicitly copied from the host memory to the accelerator memory before it can be processed by the accelerator. Similarly, results or modified data may need to be copied back from the accelerator memory to the host memory to make them accessible to the CPU. 
+Accelerators in general and GPUs in particular have their own dedicated memory separate from the system memory (**this is changing! see AMD MI300A!**). When programming for GPUs, there are two sets of pointers involved and it's necessary to manage data movement between the host memory and the accelerator memory. Data needs to be explicitly copied from the host memory to the accelerator memory before it can be processed by the accelerator. Similarly, results or modified data may need to be copied back from the accelerator memory to the host memory to make them accessible to the CPU. 
 
 The main function of the code initializes the input arrays `Ah, Bh` on the CPU and computes the reference array `Cref`. It then allocates memory on the GPU for the input and output arrays `Ad, Bd`, and `Cd` using **cudaMalloc** (herein, `h` is for the `host`(CPU) and `d` for the 'device' (GPU)). The data is transferred from the CPU to the GPU using hipMemcpy, and then the GPU kernel is launched using the `<<<.>>>` syntax. All kernels launch are asynchronous. After launch the control returns to the `main()` and the code proceeds to the next instructions.