tipoo

d_2 said:

On what basis are the numbers unrealistic?  this product is one of the first made using a 3nm chip … EVER … and a 5nm to 3nm “jump” is a far bigger % size decrease than previous (ie 10 to 8)

Tell me you're not confidently just dividing the fabrication plant marketing name as an actual number... 
 
 
 
By the way, TSMC is noting it has a 10-15% performance improvement over N5, and .58 the area density, so it's a decent node jump but not the biggest size decrease ever and the fab names are mostly just marketing, 7nm had 30-50nm fin gate lengths and pitch lengths lol 
 
https://www.anandtech.com/show/18727/tsmcs-3nm-journey-slow-ramp-huge-investments-big-future

At least it's more believable than the 4000 single core score Max Tech ran with lol.  
 
At the higher end of an improvement this couuuld be believable with 3nm offering a big improvement over a few generations of 5nm, but this is also that timeframe of a lot of fake benchmarks, even if one shows up in the GB database sometimes it's later taken down as a fake and I don't see this one on there even. 

>149 minutes 
>70 minutes

2.2x slower doesn't sound "nearly as fast" as the 16. The low GPU utilization is because ML tasks are automatically dispatched to the neural engine instead. 

Their single core performance gains have been very impressive thus far, but it had to slow down in gains somewhere, looks like we're approaching it if this is true. 

They'll probably pitch other aspects of the SoC, the GPU and the 'matrix' engine. 

avon b7 said:

Tweaked A12 for XR successor and A13 for the 'Pro' models?

That might enable them to cut costs on the XR a little although it could push up the cost of the pro models as the A13 would have a smaller market to ship to initially.

iirc, this was also said around A12 leaked benchmarks (that ended up being true), that it was a tweaked A11, but it was not so. It just looks like they've finally approached the exponential difficulty curve part of single threaded performance. 

We deal with PHI data that can't go on AWS or any outside servers. Some of our machines are 768GB RAM, the previous workstation limit, as virtualized instances as mentioned will take a heck of a lot of memory relative to their need for CPU. That bit seems to be throwing off a lot of people online who can't imagine needing 1.5TB in a single workstation. We were already maxing out older platforms.