6 ms·
What's stopping them from creating 1000 core cpus?
by deevolution 7y ago
What's stopping them from creating 1000 core cpus?
- nrki 7y agoHeat, energy usage.
- Ragib_Zaman 7y agoAlso maybe complications involving latency for communication between cores, and memory coherence protocols.
- Symmetry 7y agoYields are the main barrier to single pieces of silicon that big. You have a certain chance of getting a defect per square mm of you chip and as they get bigger the chance of a bad defect gets higher and yields go down. Often they'll occur in a place where you can just disable a core or bank of cache and still sell the chip but not always. So yield rates tend to go down as chips get bigger. Also larger chips make less efficient use of the wafer. Economically, there aren't so many people looking for 1000 cores that it makes sense to put in the NRE to assemble a giant package to put all of that in versus just selling a system that can have multiple sockets. Cooling limits also make spreading out work across multiple sockets a better choice.
- Namrog84 7y agoHow do GPU cores differ in that they have thousands of cores?
- wmf 7y agoThey don't; GPUs have <72 real cores and thousands of marketing cores. And they can disable defective cores so their massive dies are still usable.
- dingo_bat 7y agoMarketing cores lol...
- Symmetry 7y agoAs wmf said, what NVidia calls a "core" isn't something that can issue it's own instructions so it isn't really something you'd normally consider a core, though it does have a PC and can compare its PC against the PC of broadcast instructions to decide if it should execute or not so it's a bit more sophisticated than a simple SIMD vector lane. Maybe on par with an execution port? What's more equivalent to a CPU core would be what NVidia calls an SM and AMD calls a compute unit. These decide which instructions to issue next and broadcast them to the various lanes. You'll have dozens of them in a typical GPU, about the same as the number of CPU cores in the same silicon area.
- mrguyorama 7y agoIs there a market need for 1000 cores running simultaniously that doesn't already need all the other supporting infra you get from a couple extra "full fat" servers each running 64 cores? Presumably if there is any need, it's currently being done by gpus
- imtringued 7y agoMemory bandwidth. Top end Xeon Phis have 288 threads. They need high bandwidth memory and the Xeon Phis are limited to only a few gigabytes of HMC.