Howdy, Stranger!
It looks like you're new here. If you want to get involved, click one of these buttons!
Quick Links
AMD promises "Navi 2X" this year with 50% improvement in performance per watt over Navi
https://www.anandtech.com/show/15591/amds-rdna-2-gets-a-codename-navi-2x-comes-this-year-with-50-improved-perfperwatt
AMD actually announced a ton of stuff today. This is the one that I think is most interesting. Nvidia will presumably get a large improvement in performance per watt with the move to 7 nm as well, so it's not clear how the architectures will compare.
AMD also said today that they'll use some 7 nm process node for their Zen 3 CPUs and Navi 2X GPUs, but that it may or may not be an EUV process node. They also said that they'll move to 5 nm for Zen 4, which would have been the obvious thing to guess.
AMD actually announced a ton of stuff today. This is the one that I think is most interesting. Nvidia will presumably get a large improvement in performance per watt with the move to 7 nm as well, so it's not clear how the architectures will compare.
AMD also said today that they'll use some 7 nm process node for their Zen 3 CPUs and Navi 2X GPUs, but that it may or may not be an EUV process node. They also said that they'll move to 5 nm for Zen 4, which would have been the obvious thing to guess.
![[Deleted User]](https://images.v-cdn.net/banned_large.png)
![[Deleted User]](http://w7.vanillicon.com/v2/7bfcf0fe052d03fae6e745ee91ec8604.svg)
Comments
instance, and not make it to higher tier products.
50% is a big leap, but it needs to translate to a product I want to buy, and making another 5600 that uses less power won’t exactly excite me.
A 12TFlop part that only costs <$300 though ... that would move the needle. And if that’s what they are packing in a console, a dGPU at that price point doesn’t seem like it should be out of reach.
Remember that even if the CPU core dies are the same across the Ryzen, Threadripper, and EPYC lines, the I/O dies are necessarily different. There's no reason why they couldn't go DDR4 for Ryzen and DDR5 for EPYC with exactly the same CPU cores. That would actually make some sense to do because Ryzen is widely available much sooner than EPYC, so even if DDR5 is going to be ready on time for EPYC, it might not be for Ryzen.
Additionally, the chiplet approach with memory controllers on an I/O die that is separate from the CPU cores makes it easier to offer both DDR4 and DDR5 versions of a chip if so inclined. They just need two different I/O dies. It wouldn't be surprising to see both AM4 and AM5 versions of some generation of Ryzen. That would work as a hedge against DDR5 being delayed or overpriced. It would also give people with older Ryzen CPUs a way to upgrade just the CPU (nearly pure profit for AMD) without having to also replace the motherboard (not much profit for AMD) and risk the user moving to an Intel motherboard (no profit for AMD). Recall that the first generation of Phenom II dies had both AM2+ and AM3 support.
Although there is some benefit in supporting a socket, sometimes it's better to drop the socket completely so you can rethink the platform.
Connecting the CPU and GPU via infinity fabric is going to be for HPC stuff. There are some problems that need a ton of bandwidth between the CPU and GPU to perform well. There are also problems where you don't need a ton of bandwidth to connect them, but the programmer is an idiot and has structured code in such a way that it won't perform well unless you have a ton of bandwidth. But that's just enterprise stuff that will have dodgy driver support that you don't want to have to deal with. Whatever you think of GeForce and Radeon driver support, they're both many miles ahead of Tesla and Radeon Instinct driver support.