Four years ago, you could get an Intel chip and overclock it to 5.0GHz, well before Intel made it the official "turbo" speed. Intel actually had the potential to just clock chips much higher than it did, giving a lot of room to just sit around and bump clock speeds over the past four years without any significant changes.
We don't know how high these new AMD chips can clock on all cores at this point, and it may require better motherboards than many people have. An Asus ROG Crosshair VI Hero from the first generation may be able to do better than many second generation boards, just because of the VRMs.
7nm with EUV might help (esp. for defect rates), but generally, these ultra-dense nodes haven't been kind to clock speeds or power delivery systems (Intel is still struggling with 10nm that is equivalent to TSMC 7nm). Copper resistance increases as it gets smaller, so putting high current through wires with higher resistance will generate a lot of localized heat energy. Without more exotic materials for transistors and wire connects (and studying different types of transistors like GAA FETs for 5nm and below), I think clock speeds will stagnate and energy usage for higher clocks will ramp significantly as we move to smaller lithographies. I expect cache sizes will be the next battle to retain as much data on-chip to both reduce power and increase performance per clock by not having to make as many trips out to slow RAM for instructions and/or data.
My guess: standing by for a last 2019 or 2020 refresh. AMD simply doesn't have to play their full hand right now to be competitive. Going to 16 core on AM4 looks to be trivial on paper since they're already doing 12 core: just a matter of clock speeds and core voltage to make it happen inside of AM4 parameters.
My guess is the 16 core is the base level threadripper . No way for dual channel memory to provide the bandwidth 16 cores need . Even the 12 core will have memory limitations in some applications
You can probably use higher speed memory with Zen 2 and it also looks like it may get better memory bandwidth utilization. Throw in some prefetch improvements and AVX 256 and it may outperform 16-core ThreadRipper for most applications. ThreadRipper may still start at 16-core though. I don’t know if they would sell a 64-core ThreadRipper. To run 64 cores at high clock would take ridiculous power. Epyc 2 with 64-cores will probably be under 2.5 GHz. Perhaps 48 core ThreadRipper.
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
307 Comments
View All Comments
Targon - Monday, May 27, 2019 - link
Four years ago, you could get an Intel chip and overclock it to 5.0GHz, well before Intel made it the official "turbo" speed. Intel actually had the potential to just clock chips much higher than it did, giving a lot of room to just sit around and bump clock speeds over the past four years without any significant changes.We don't know how high these new AMD chips can clock on all cores at this point, and it may require better motherboards than many people have. An Asus ROG Crosshair VI Hero from the first generation may be able to do better than many second generation boards, just because of the VRMs.
flofixer - Monday, May 27, 2019 - link
because of the lack of cobalt?JasonMZW20 - Tuesday, May 28, 2019 - link
7nm with EUV might help (esp. for defect rates), but generally, these ultra-dense nodes haven't been kind to clock speeds or power delivery systems (Intel is still struggling with 10nm that is equivalent to TSMC 7nm). Copper resistance increases as it gets smaller, so putting high current through wires with higher resistance will generate a lot of localized heat energy. Without more exotic materials for transistors and wire connects (and studying different types of transistors like GAA FETs for 5nm and below), I think clock speeds will stagnate and energy usage for higher clocks will ramp significantly as we move to smaller lithographies. I expect cache sizes will be the next battle to retain as much data on-chip to both reduce power and increase performance per clock by not having to make as many trips out to slow RAM for instructions and/or data.Kevin G - Monday, May 27, 2019 - link
My guess: standing by for a last 2019 or 2020 refresh. AMD simply doesn't have to play their full hand right now to be competitive. Going to 16 core on AM4 looks to be trivial on paper since they're already doing 12 core: just a matter of clock speeds and core voltage to make it happen inside of AM4 parameters.12345 - Monday, May 27, 2019 - link
Sound more like sandbagging to me.brunis.dk - Monday, May 27, 2019 - link
If you can win by only giving 80% .. :)sorten - Monday, May 27, 2019 - link
Exactly. This is like Usain Bolt skipping the last 10m of a 100m dash because he's so far ahead.Outlander_04 - Monday, May 27, 2019 - link
My guess is the 16 core is the base level threadripper . No way for dual channel memory to provide the bandwidth 16 cores need .Even the 12 core will have memory limitations in some applications
shing3232 - Monday, May 27, 2019 - link
No, 2990WX use quad channel and 32core, and it work out okish. Zen2 has larger L3, it would work even better. maybe there will be a 16C wxjamescox - Tuesday, May 28, 2019 - link
You can probably use higher speed memory with Zen 2 and it also looks like it may get better memory bandwidth utilization. Throw in some prefetch improvements and AVX 256 and it may outperform 16-core ThreadRipper for most applications. ThreadRipper may still start at 16-core though. I don’t know if they would sell a 64-core ThreadRipper. To run 64 cores at high clock would take ridiculous power. Epyc 2 with 64-cores will probably be under 2.5 GHz. Perhaps 48 core ThreadRipper.