Navi should be cheaper IMO. Navi has way more smaller die (2.5x more smaller than Vega 64 if I recall) and uses GDDR6 instead of HBM. I don't know why AMD priced them that high tho...
"the performance, partially a consequence of going with 12nm, just isn’t there"
People should have been weaned off this by now: process shrinks stopped inherently boosting performance years ago. Power consumption drops and perf/watt increases, but 'perf/transistor' continues to decrease (due to leakage increasing as packing density grows, coupled with power density increases) as it has done for some time, and cost/transistor has been going up since 28nm. A brief period of making dies bigger and bigger (and more and more expensive) has culminated in reticle-limit dies like GV100 and TU102, but that now makes start the wall process scaling hit some time ago in reality. This is only going to continue as processes shrink further. Cost/transistor will rise, perf/transistor will drop, and increasing performance means dies will continue to grow. Performance gains will continue to come from architectural changes, not process changes. Unless you're hitting the reticle limit AND cannot split your die into multiple dies due to latency reasons only then does it make any sense to move to a smaller process, and you will take a hit to both cost/perf as well as perf/transistor in doing so which may eat any gains from packing more transistors in.
I agree that performance per transistor can drop with die sizes due to leakage. And that new nodes are expensive at first. But once a node is mature, the cost per transistor should be lower than previous node. If that wasn't the case than the semi conductor business would be completely sunk.
Changes and improvements in software need to be done which are going very slowly because it takes a lot of work(coding) and knowledge but mostly hard work which younger generations are not willing to do (distracted by social networking and gaming).
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
281 Comments
View All Comments
Korguz - Tuesday, July 2, 2019 - link
um yea ok.. sure...eva02langley - Tuesday, July 2, 2019 - link
Well, AMD will introduce bundles, games and rebates for Ryzen and Navi launch. I will not be surprised to see Navi cut down by 50$ in August.Fritzkier - Tuesday, July 2, 2019 - link
Navi should be cheaper IMO. Navi has way more smaller die (2.5x more smaller than Vega 64 if I recall) and uses GDDR6 instead of HBM. I don't know why AMD priced them that high tho...eva02langley - Tuesday, July 2, 2019 - link
Because Nvidia was asking for 500-600$ for less performances.Meteor2 - Saturday, July 6, 2019 - link
*was*edzieba - Tuesday, July 2, 2019 - link
"the performance, partially a consequence of going with 12nm, just isn’t there"People should have been weaned off this by now: process shrinks stopped inherently boosting performance years ago. Power consumption drops and perf/watt increases, but 'perf/transistor' continues to decrease (due to leakage increasing as packing density grows, coupled with power density increases) as it has done for some time, and cost/transistor has been going up since 28nm. A brief period of making dies bigger and bigger (and more and more expensive) has culminated in reticle-limit dies like GV100 and TU102, but that now makes start the wall process scaling hit some time ago in reality.
This is only going to continue as processes shrink further. Cost/transistor will rise, perf/transistor will drop, and increasing performance means dies will continue to grow. Performance gains will continue to come from architectural changes, not process changes. Unless you're hitting the reticle limit AND cannot split your die into multiple dies due to latency reasons only then does it make any sense to move to a smaller process, and you will take a hit to both cost/perf as well as perf/transistor in doing so which may eat any gains from packing more transistors in.
Threska - Tuesday, July 2, 2019 - link
Chiplets.Hixbot - Thursday, July 4, 2019 - link
I agree that performance per transistor can drop with die sizes due to leakage. And that new nodes are expensive at first. But once a node is mature, the cost per transistor should be lower than previous node. If that wasn't the case than the semi conductor business would be completely sunk.Hixbot - Thursday, July 4, 2019 - link
Edit: I ageee that performance per transistor can drop with process shrinksGastec - Wednesday, July 17, 2019 - link
Changes and improvements in software need to be done which are going very slowly because it takes a lot of work(coding) and knowledge but mostly hard work which younger generations are not willing to do (distracted by social networking and gaming).