64 Gbyte memory for 64 cores?

[8l]

1 core per 1 Gbyte memory, is this mean something?

[Dade]

[piotr5]

you mean 32 kbytes.
it's the usual problem with address-space: 32bit can only address 2^32 bytes or twice that many words. that's just 4GB or 8GB or 16GB at most. if you want more, you need 64bit processor. but on the other hand it is not as much of a gain to have 64bit when the 64GB chips take up almost the same space as the whole parallella board, so I guess we need to wait till memory-developers catch up with the new possibilities of 64-bit...

parallella is based on all of the core's memory being directly accessible.1024 cores with 32kb each, that's 32MB used up by them and the rest of the 4GB address-space free for the 2 main processors and whatever other hardware. guess there is some space for improvement, but 64GB wasted by the cores alone would really be overkill -- the way software nowadays is lacking complexity...

btw, 32kb is not really the maximum-program-size for each core, 8kb are reserved for stack, and maybe the program will want another 8kb for storing the data it generates. so better look out for compilers that produce code under 16KB!

[ysapir]

The Epiphany architecture maps segments of 1MB of address space to each core. In a system, this 1MB range is (more or less) exclusive to a core, but can be accessed by any other master from inside or outside the chip.

Out of the 1MB, some space is reserved for the core registers (the upper 64KB of the slice) and some is covered by the on-core SRAM. The amount of SRAM depends on the chip design and in the current products it is 32KB per core.

[fmotta]

[ChrisGranger]

[aolofsson]

The 64-core Parallella will be built, and will be shipped to a few Parallella backers ($750 tier). The $3M was needed to bring the cost of the board down to $199. We hope to ship some of the early 64-core prototypes (based on the Zedboard) to the $10k tier backers next week.

[tnt]

Oh damn, I misread the reward tier, I tought that all 64 core reward tier were dependent of the stretch goal. I should have taken the 64CORE-MAY instead of 64-CORE ...

[ChrisGranger]

[PickPacket]