3 ms·
> Does the performance scales linearly? On GPUs, pretty much. It’s even called embarrassingly parallel for this reason.
by tambourine_man 5y ago
> Does the performance scales linearly?
On GPUs, pretty much. It’s even called embarrassingly parallel for this reason.
- JohnBooty 5y agoAt some point you hit bandwidth limitations, right? Unless each core has local, dedicated memory they will need to access the GPU's shared memory for various things right?
- smoldesu 5y agoPCIe 4.0+ is capable of ~31504MB/s on an x16 interface. Besides that, you might run into memory issues, but I honestly haven't seen a memory-bottlenecked workload on GPUs outside of the crypto mining scene.
- ksec 5y ago>At some point you hit bandwidth limitations Yes, with HBM3 you could go up to 5TB/s and in the future 6.4TB/s. To give a comparison a 3980 RTX Ti has ~900 GB/s with GDDR6x. But just with everything we hit TDP or cost ceiling well before any technical barriers.