- Oct 22, 2004
- 809
- 1,412
- 136
Kyle Bennet from HardOCP corroborates the information in AdoredTV's latest video:
"There is a whole lot of reality in that video. A lot. There is a little wrong, but not a lot."
https://hardforum.com/threads/adore...d-ryzen-and-radeon-3000-series-leaks.1973015/
I take that as mounting evidence that mainstream desktop will use chiplet design. If MCM packaging is feasible in the mainstream, from a cost and capacity perspective, it makes a whole lot of sense for AMD to reuse their 7nm 8-core chiplet.
Although I've earlier opined that 8-core is more than enough for mainstream, I now think it makes sense for AMD to extend that to 16 cores, to make sure they surpass Intel's offerings. In particular, I didn't think Intel would go beyond 8 cores on 14nm, seeing how i9-9900K pushed the limits on frequency and power. But we now have firm rumours about 10-core 14nm Comet Lake, probably planned for the next annual refresh, late 2019. Also, AMD needed (and still needs) to be prepared for Intel actually getting 10nm to work, which may give Intel opportunity for further core-count increase.
Another concern of mine was memory bandwidth. But a dual-channel memory controller per 16 cores is what 64-core "Rome" is designed around, so 16-core on AM4 may be just fine as a product in the mainstream client market.
I have argued that AMD is well positioned to fight the core war with the Threadripper platform, and I still think that. However, there are cost considerations. At the same price, AMD would prefer to sell a lower-cost mainstream chip, for higher margin.
So I'm persuaded that 16-core will be the new upper limit for mainstream.
I am also persuaded by the tri-chiplet approach. Previously it concerned me that AMD would use two expensive 7nm chiplets in a mainstream socket, but considering they can use a dummy die, as they do for Threadripper now, it makes sense, especially with the opportunity to combine salvaged dies with some cores disabled. It also makes sense to be able to replace a CPU chiplet by a GPU chiplet to implement an APU.
PS. Regarding clock speeds, I know the widespread opinion is that TSMC's 7nm HPC process is far behind Intel's 14nm on performance, but I think this is pessimistic and makes little sense (both AMD and TSMC have done 5 GHz before). I will be disappointed if we don't see boost speed at least reaching 5 GHz for the top bins.
"There is a whole lot of reality in that video. A lot. There is a little wrong, but not a lot."
https://hardforum.com/threads/adore...d-ryzen-and-radeon-3000-series-leaks.1973015/
I take that as mounting evidence that mainstream desktop will use chiplet design. If MCM packaging is feasible in the mainstream, from a cost and capacity perspective, it makes a whole lot of sense for AMD to reuse their 7nm 8-core chiplet.
Although I've earlier opined that 8-core is more than enough for mainstream, I now think it makes sense for AMD to extend that to 16 cores, to make sure they surpass Intel's offerings. In particular, I didn't think Intel would go beyond 8 cores on 14nm, seeing how i9-9900K pushed the limits on frequency and power. But we now have firm rumours about 10-core 14nm Comet Lake, probably planned for the next annual refresh, late 2019. Also, AMD needed (and still needs) to be prepared for Intel actually getting 10nm to work, which may give Intel opportunity for further core-count increase.
Another concern of mine was memory bandwidth. But a dual-channel memory controller per 16 cores is what 64-core "Rome" is designed around, so 16-core on AM4 may be just fine as a product in the mainstream client market.
I have argued that AMD is well positioned to fight the core war with the Threadripper platform, and I still think that. However, there are cost considerations. At the same price, AMD would prefer to sell a lower-cost mainstream chip, for higher margin.
So I'm persuaded that 16-core will be the new upper limit for mainstream.
I am also persuaded by the tri-chiplet approach. Previously it concerned me that AMD would use two expensive 7nm chiplets in a mainstream socket, but considering they can use a dummy die, as they do for Threadripper now, it makes sense, especially with the opportunity to combine salvaged dies with some cores disabled. It also makes sense to be able to replace a CPU chiplet by a GPU chiplet to implement an APU.
PS. Regarding clock speeds, I know the widespread opinion is that TSMC's 7nm HPC process is far behind Intel's 14nm on performance, but I think this is pessimistic and makes little sense (both AMD and TSMC have done 5 GHz before). I will be disappointed if we don't see boost speed at least reaching 5 GHz for the top bins.
Last edited: