AMDs chiplet design allows them to reduce cost of the Ryzen 9 3950X by half
Much is said and spoken about the Chiplet design from AMD. It works out well for them, really well. But the main reason of using chiplets was dropping expensive monolithic processor die designs, in favor of more economical cores to fabricate. But nobody really knew how much of a benefit that was in terms of money.
AMD revealed that if a monolithic design was used (as Intel uses), a 16-core processor would cost more than double what it currently costs. We mentioned this many times, but if you fabricate big huge die's then chances that on a wafer the yields are bad, is bigger. If you design lots of smaller processors dies on a wafer, the risk of a damaged die is far smaller, and thus far more economical to fabricate as you end up with more working CPU dies (better yields). Of course chipset design have challenges of their own, but AMD isn't rather bothered by it by designing an ultra fast IO chip.
So for a 16-core model you only have to add an additional core die and IO, which is much smaller, cheaper to achieve and this economic. As the new slides indicate, it turns out that it has reduced cost by half. Thanks to this new design, you have the possibility of only marking the price at US $750 to get the Ryzen 9 3950X with 16 cores and 32 threads, while it would surely have cost more somewhere north of USD 1250 perhaps USD 1500, if a monolithic design had been used.
Senior Member
Posts: 14092
Joined: 2004-05-16
which is why with scalable rdna2 AMD will (eventually with Infinity fabric gen 3) make a chiplet gpu
again the same exact factors - smaller dies = higher yields. excepting this time is to b*tch-slap Nvidia instead of Intel.
Nvidia isn't complacent like Intel, however they rely entirely too much on monolithic dies and have no technological alternative like AMD.
Nvidia published numerous papers about MCM, they have working MCM prototypes and have been gearing NVLink for MCM designs, etc.
I don't understand how you can say they don't have an alternative - they've arguably demonstrated more than AMD has in terms of MCM for GPUs.
Senior Member
Posts: 14643
Joined: 2014-07-21
which is why with scalable rdna2 AMD will (eventually with Infinity fabric gen 3) make a chiplet gpu
again the same exact factors - smaller dies = higher yields. excepting this time is to b*tch-slap Nvidia instead of Intel.
Nvidia isn't complacent like Intel, however they rely entirely too much on monolithic dies and have no technological alternative like AMD.
this isn't to say chiplet gpu's don't have issues to date - which is why they haven't been on market. but this design team has been specifically tasked to fix the problems.
this is the AMD end game on the gpu front. once they achieve a fast enough i/o and an upgraded interposer all the advantages Nvidia has had in gpu design will go by the wayside (relatively speaking) as AMD will be able to brute force any spec at a far lower cost widening the delta between them price wise.
https://www.guru3d.com/news-story/nvidia-next-gen-gpu-hopper-could-be-offered-in-chiplet-design.html
Sure AMD's ahead by quite some, but Nvidia's R&D budget is 10 times as big as AMD's, and in comparison to Intel, they have been able to keep the lead up until now. But you do have a valid point, it will be interesting to see what they do and if they can take the fight to Nvidia. Wouldn't hurt to have such an escalation of competition with GPUs as well.