[Sweclockers] Radeon 380X coming late spring, almost 50% improvement over 290X

[Paul98]

Interesting story by Fudzilla:



Read the whole thing

Terrible article, that guy clearly has no idea what he is writing about.

 

[stuff_me_good]

1. As I understood correctly 3D memory used on pascal, it should be stacked right on top of the GPU core, but why in the picture where is the pascal module, there are 4 chips just beside the GPU die?

2. Isn't his 3D memory approach HMC and not HMB?

3. About nvlink, how can you get faster lane to the system meomry of 80GB bandwidth is the card is in slower PCI-E slot? Does it use some sorth of compression or what? Besides, does it help at all if you have slow DDR4 on the other end of the line not making even close to the 80GB bandwidth limit?

 

[DiogoDX]

1. As I understood correctly 3D memory used on pascal, it should be stacked right on top of the GPU core, but why in the picture where is the pascal module, there are 4 chips just beside the GPU die?

2. Isn't his 3D memory approach HMC and not HMB?

3. About nvlink, how can you get faster lane to the system meomry of 80GB bandwidth is the card is in slower PCI-E slot? Does it use some sorth of compression or what? Besides, does it help at all if you have slow DDR4 on the other end of the line not making even close to the 80GB bandwidth limit?

3D memory = stacking memory chips to form a "stack" (4Hi, 8Hi indicates how many dram chips they are using to from a stack)

3D stacking (vertical stacking) = the "stack" is placed on top of the chip/SOC


Nvidia mockup is showing a 2.5D stacking using a interposer that makes more sense for multiple stacking chips like gpus.