Originally Posted by ebduncan
carrizo won;t be 3.5ghz all the time, it will be only single thread probably.
My richland laptop is the same way. has a 3.5ghz turbo, but all the cores only run at 2.3ghz when they are loaded.
Anyway Carrizo mobile won't really be anything special. It will just be a general improvement from kaveri. Which was nothing to write home about.
HSA definitely shows promise if it can ever take flight.
I like the concept and I like the direction but HSA just isn't designed to take flight. Just about everything conspires against it. Even the main concept.
Originally Posted by Nintendo Maniac 64
It'd probably be more practical once Zen comes around since HSA main benefit seems to be in productivity-based things, though games could possibly take advantage of it.
Speaking of games, any future console design wins would definitely
take advantage of it, and it's even thought the PS4 and XB1 have some lesser HSA-like functionality already. I know that Nintendo in particular seems to love
memory bandwidth when it comes to how they design their console hardware, so HSA would be right up their ally.
HSA doesn't exactly help with memory bandwidth though
Originally Posted by epic1337
java partly uses GPGPU, one could then imagine how vast java had already spread.
its not really a joke, nor was it meant to be offensive, it was simply a point-out at how GPGPU had already been preferred over general compute.
Which is already know but, consider this. What high performance software is done in, of all things, Java? Especially GPGPU?
GPGPU in Java is like adding a turbo charger to a golf cart.
webpage (browser) rendering are GPGPU heavy (hw acceleration, of course you could disable some of it
Yah my HD 4000 comes to its knees rendering tabs in Chrome. And by knees I mean it wonders if that was supposed to be some kind of challenge.
likewise video decoding and database operations had recently preferred GPGPU.
The former doesn't need more performance so much as it needs better efficiency. The latter? Databases really aren't my kinda thing, but it'll depend on the database and operations and whether they can be done in parallel.
You see, GPGPU is great but not everything can just encorporate GPGPU.
i hate to quote wiki for context but...
while HUMA functions as a bridge to join two different processors to function on the same ram interface, HSA functions to allow direct access without needing allocated space.
basically put, hUMA without HSA falls back to being NUMA.
and on that site says this, in a sense they can implement HSA on discreet GPUs, just that its not practical in their views
edit: on that note i think i get where you're getting at, i do understand HSA that this is only one part of it's function
I think you're getting closer but my main point is that GPGPU by traditional means (dGPU) far outclasses what an APU is capable of.
That GPGPU is best utilized in professional work as you are showing, where professional GPUs are faster, more accurate, more efficient and better supported.
That AMD is in NO position to dictate the market or see the spread of HSA in desktop applications due to the necessity of specialized hardware that ranks among the minority and that development requires a special compiler system.
HSA is sort of like a stepping stone in research, not important for immediate use, but as a gateway to things that are.
There doesn't seem to be any nice solution to the faults in current GPGPU using systems.
HSA main function is to make GPGPU accessible as a general compute core like CPUs.
AMD, Nvidia, people in enterprise, consumers, etc... WISH that were the case.
HSA isn't even close to doing that. It wishes to use them as fully as GPGPU can but it doesn't make any breakthroughs there.
Originally Posted by SpeedyVT
They are building 2016 cpus with both ARM and x86 on the same chip. CPUs are often taxed by pure math operations and a GPU can accelerate that. However nothing yet. nothing major has been compiled for HSA.
You mean MANY math operations which can be run in parallel. Quick and dirty example, matrix multiplication.