At this point, until a GPU overhaul is made like Zen was for the CPU space, we are not going to see much changes. AMD will need to introduce chiplet design gpus for things to really change.
How would a chiplet design really help GPUs, though? You don't have nearly the kind of I/O requirements on-chip for a GPU as you do a CPU. You're already massively parallel, and you can just shut off defective "cores" and put those parts in a lower bin; going to a 2 (or more) chip setup would probably just hurt performance.
Chiplets require more expensive packaging and to really scale, you have to design with that concept in mind. Previously the cost-benefit from an engineering standpoint was simply to eat the cost going with larger dies and release harvested products due to the rarity of fully functional chips. The costs to migrate to newer nodes is increasing and the necessity of using multi-patterning have put tighter limits on how big chips can be. Chiplets are the way forward as they solve the current issues in manufacturing and the packaging doesn't carry the same premium as before.
I don't think a chiplet would help a consumer GPU much. Chiplets allow you to put more transistors than you otherwise would be able to. But consumer GPUs don't reach the reticle limit. More transistors would also increase the price of the GPUs. The cost per transistor isn't going down as much from node to node as it used to. So if AMD tried to boost performance by building a big chip with multiple chiplets it would be very expensive not only because of the complexity of chiplet technology but also because of the cost of all those transistors.
What AMD needs to do is to continue to modify their architecture. The RDNA is a good first step. It's still behind NVIDIA in memory bandwidth, energy, and workload efficiency, but it looks like it makes a good jump over GCN in those areas.
Seriously... everything you are saying is only your opinion. Chiplet is bringing better yield, better binning, reduce waste and end up giving more margins due to modularity.
As of now, we have no clue what RDNA can provide, the reviews are not even out yet and it is the first kick at the can. If all the game industry is backing up AMD, guess what, there must be a reason for it.
Chiplets bring better yields, yes, but at the expense of worse power efficiency. If you put the chiplets on silicon interposers it's very expensive, too. So you must put them on less expensive subtrate and that affects the communication speed. Maybe when the technology is more mature it could make sense, but there's no sense to introduce the complexity now. In any case, AMD fans talk about chiplets as if AMD is the only one pursuing them. The whole industry is. Intel, AMD, NVIDIA,TSMC, everybody.
Why do you say the game industry is "backing up AMD"? What does that mean?
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
eva02langley - Tuesday, July 2, 2019 - link
At this point, until a GPU overhaul is made like Zen was for the CPU space, we are not going to see much changes. AMD will need to introduce chiplet design gpus for things to really change.sing_electric - Tuesday, July 2, 2019 - link
How would a chiplet design really help GPUs, though? You don't have nearly the kind of I/O requirements on-chip for a GPU as you do a CPU. You're already massively parallel, and you can just shut off defective "cores" and put those parts in a lower bin; going to a 2 (or more) chip setup would probably just hurt performance.eva02langley - Tuesday, July 2, 2019 - link
By making smaller chips, better yield, better binning, reducing waste... on multiple product.It is like designing a frame for multiple car model assembly. You are cutting cost and end up with better quality products...
Why chiplets? It is obvious as a business strategy.
jordanclock - Tuesday, July 2, 2019 - link
If it's obvious, then I'm sure AMD has considered it and decided it doesn't make sense.Why does everyone in the comments act like they know better than entire teams of the world's best engineers?
Kevin G - Tuesday, July 2, 2019 - link
Chiplets require more expensive packaging and to really scale, you have to design with that concept in mind. Previously the cost-benefit from an engineering standpoint was simply to eat the cost going with larger dies and release harvested products due to the rarity of fully functional chips. The costs to migrate to newer nodes is increasing and the necessity of using multi-patterning have put tighter limits on how big chips can be. Chiplets are the way forward as they solve the current issues in manufacturing and the packaging doesn't carry the same premium as before.Gastec - Wednesday, July 17, 2019 - link
Chiplets and multi-GPU configs FTW! Also convince the electric companies to reduce the electricity bill by 50% :)Yojimbo - Tuesday, July 2, 2019 - link
I don't think a chiplet would help a consumer GPU much. Chiplets allow you to put more transistors than you otherwise would be able to. But consumer GPUs don't reach the reticle limit. More transistors would also increase the price of the GPUs. The cost per transistor isn't going down as much from node to node as it used to. So if AMD tried to boost performance by building a big chip with multiple chiplets it would be very expensive not only because of the complexity of chiplet technology but also because of the cost of all those transistors.What AMD needs to do is to continue to modify their architecture. The RDNA is a good first step. It's still behind NVIDIA in memory bandwidth, energy, and workload efficiency, but it looks like it makes a good jump over GCN in those areas.
eva02langley - Tuesday, July 2, 2019 - link
Seriously... everything you are saying is only your opinion. Chiplet is bringing better yield, better binning, reduce waste and end up giving more margins due to modularity.As of now, we have no clue what RDNA can provide, the reviews are not even out yet and it is the first kick at the can. If all the game industry is backing up AMD, guess what, there must be a reason for it.
Yojimbo - Wednesday, July 3, 2019 - link
Chiplets bring better yields, yes, but at the expense of worse power efficiency. If you put the chiplets on silicon interposers it's very expensive, too. So you must put them on less expensive subtrate and that affects the communication speed. Maybe when the technology is more mature it could make sense, but there's no sense to introduce the complexity now. In any case, AMD fans talk about chiplets as if AMD is the only one pursuing them. The whole industry is. Intel, AMD, NVIDIA,TSMC, everybody.Why do you say the game industry is "backing up AMD"? What does that mean?
Meteor2 - Saturday, July 6, 2019 - link
Everything Yojimbo says is fact; eva02 you're clearly very poorly informed.