The Kyle Experience: Texas BBQ *spawn*

It would be really great news for AMD if GP106 lands in the fall and they get a few months alone in the mainstream segment. Which by extension would be good news for almost everybody.

 

Yeah, they could really use that, I guess. But then, Nvidia isn't that kind of charitable. If they deem GP106 competitive with Polaris P10, they will launch as soon as they can after Polaris (assuming they cleared channel inventory enough for their taste).

 

I think 1280CC with 192b at 8GBps or 256b at 6GBps. Seems the 1080 needs about 12GBps on it's 256b bus for feeding all shader.

I think it will be a little, 5-10% below Polaris 10 in performance if it is around 2000Mhz. But have to wait to see how Polaris performs in real world compared to previous GCN too now.

 

NVidia's Drive PX 2 was shown at GTC with two discrete Pascal chips on it, most likely GP106. Those are mounted on MXM carriers with 8 RAM chips each. That suggests a 128 bit bus (or, unlikely but possible, 256!)

 

A 128-bit bus seems most likely given that the GPU memory bandwidth was listed as "80+ GB/s."



That's only a lower bound for the GP106 chip though.

 

Great catch. 80GB/sec matches up with the low 5GT/sec clocked GDDR5 speeds used on mobile MXM parts like GTX 960M and 965M, both with a 128 bit bus. Assuming GTX 1070's GDDR5 speeds on a desktop part, GP106 would have 128GB/sec, just a bit faster than GM206's 112GB/sec.

 

Margins? How bandwidth dependant is Deep Learning as probably used in the PX2's use case scenario? Honest question, I have no idea.

 

For the size of the chip and price bracket it will target, a 128 bit bus seems appropriate. 192 bit bus might start to become pad limited depending on how small the chip is as well as eating into valuable silicon budget for a price sensitive chip. 128 bit is what I would imagine Polaris 11 will be using as well.

Regards,
SB