|
Machiavellian said: I am not doubting that the X1 does not do HUMA as AMD does HUMA, I am stating that MS system does something totally different but still does the same thing. A hardware paging table does eactly what you are talking about. The CPU and GPU do not have to move, flush or access each other memory. Memory addresses are paged to the synchronized hardware page table. The complete X1 system, GPU, CPU and Special co-Processors all have access and share this paging table. This mean any memory lookup accesses the paging table not directly with the processors memory. This is a much more complex design and its a bigger scope than what HUMA is designed to do. I agree that MS and AMD arch did not line up with each other as MS was definitely looking to create this VM type of setup. Anyway hit you up later as I need some zss myself and I teach in the morining (API class go figure). |
Just to point all the hardware can't share the same memory table... each processor have it own cache memory table... it is not like CPU access the cache memory table of GPU or vice-versa.... they have thier own cache memory table.
The proposal of the hUMA is that all memory cache table was synced all the time... it is like when the GPU uses some memory space it update its own cache table and sent the information to the CPU-cache table... so both have the same pointers address updated. Withou hUMA... the cache table is not updated so if the CPU need to access some address reserved to GPU it needs to access the GPU first to get the latest GPU-cache table to update it cache table... after that it have all the address points used by GPU and can access it.
What MS did seems like hUMA but not like hUMA lol... confuse? Different ways to do the same things... the point is hUMA is just a tech/way used by AMD to share memory between processors.... I belive Intel have it own tech/way, MS did it own tech/way with Xbone, etc.
Looking at the diagram you can see that MS says that all page tables are syncronized (in the middle).
I believe the MS way is better for what they design for Xbone than the AMD hUMA.
