AMD, ARM, Imagination, Samsung alliance publish official shared GPU-CPU blueprints An effort to tightly knit together graphics chips, processors and other hardware to boost things like video search on your desktop has taken a step forward. The HSA Foundation today officially published version 1.0 of its Heterogeneous System Architecture specification, which (if we were being flippant) describes how GPUs, …
> without having to copy chunks of data over buses, for example
Unless they have just invented a chip design which works on quantum entanglement of particles, so data can magically jump from the CPU to the GPU caches or to main memory as required, I'm pretty sure that moving data over busses is still firmly in the picture.
Very smooth. I mean external buses, or separate interconnects between on-die components. There will always be some kind of bus between the RAM and whatever is reading from or writing to the RAM, but the HSA people want to do this with one cache-coherent bus and one portable programming model.
One thing springs to mind is bandwidth: jamming everything on the same interconnect may be a problem, so I'll look into that. The specification allows you to break up the physical memory into pools if you really want to, as long as you stick to the standardized programming model.
C.
That was my point - the bus is still there, you are still moving data over it. If processor A has the data and processor B needs the data the only way to get it is to copy it over a bus (on chip or off chip, it makes no difference - there are always wires linking things together, it is just a question of scale).
All this spec seemingly changes is the need for manually keeping the different memory pools in sync - the "clever bit" is the reliance of system level hardware cache coherency - not the removal of the need to move data around the system (which is impossible).
Sorry if I didn't state the point in the most "smooth" way ...
"Hopefully the security issues were dealt with"
Haha you're thinking of things like rowhammer, which work by exploiting the occasional failures if non-EC2 memory, such as used by RAM in laptops .... and especially also by RAM in in GPUs.
GPU code is performance-oriented, because most putative security holes would be contained to the GPU so far I doubt that there has been much effort in finding and closing them.
Looking forward to the torrent of exploits when this stuff comes to consumerland
HSA developed by AMD is a major step forward in PC based system performance because it allows not only the CPU/GPU to use the same memory, more importantly it allows the best compute unit be it the CPU or GPU to process the data in real time and change modes on the fly. This will make a substantial improvement in application performance when run on HSA compliant hardware such as the Carrizo APU - which will be released in just a few weeks time.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
