Caporegime
That would bring AMD ST R23 Score to 2100 points and provided they have the same 16c 32t layout would score 37,000 in MT.
-
Competitor rules
Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.
*** AMD "Zen 4" thread (inc AM5/APU discussion) ***
That would bring AMD ST R23 Score to 2100 points and provided they have the same 16c 32t layout would score 37,000 in MT.
Optimistic on both timing and IPC improvements, timing is most likely off but 18% better IPC is not outside the realm of possibility. N7 -> N5 gives AMD a lot of extra transistors on the same die size to play with, allowing them to make bigger cores.
Assuming that's correct, it could also allow them to close down the gap with Apple on Perf/Watt. 18% more IPC would put it roughly 30% below M1 Pro in terms of Perf/Watt in a laptop. Really massive gain and definitely closes down the massive gap. Intel can still put their electric heater inside a laptop.
Is 18% IPC and 7% core clock enough for ZEN4 to eclipse current ADL single core ? I guess it is doing some quick fag packet maths but not by much, considering ZEN4 needs to last out past the next Intel iteration its going to be tight at the top I would have thought.
Caporegime
Should put it about 10% past ADL ST with a core 1/4 the size.
I doubt it will be only 18% IPC just as i doubt it will be released in april. But still we don't know how ipc will translate to gaming performance, for example, Zen3d is named as leading gaming performance, while having about 10% less IPC than ADL. Zen 4 won't have 3d cache at the beginning for what we know so they need to raise IPC much more to be translated to gaming performance,otherwise there will be small difference between Zen3d and Zen 4, and Zen 4 will be no doubt pricey. That leaker is having a fun, but we will know truth in 2 month, if they don't release Zen 4 at that time then this whole leak is a scam.
Caporegime
It looks about right to me, i seem to remember them saying "the same sort of IPC as from Zen 2 to Zen 3" which according to AMD was +19% IPC, overall that turned out to be about right, Zen 3 also clocked 10% MT to 20% ST higher but that has nothing to do with IPC. If Zen 4 is clocking 7% higher ST that would put it about 5.3 to 5.4Ghz, B2 versions of Zen 3 are clocking to 5.15Ghz easily, up from 5.0Ghz (+3%) so i can believe that too.
Its only MLID whose claimed +30% IPC with +40% overall per core higher performance, i don't believe that, AMD are ####### good X86 designers, i would argue better than Intel but they are not that good. Zen 5 may get another 20% but that's another design iteration.
Last edited:
I recall the expectation has been that Zen 3->4 will be a similar gain as Zen 2->3. So another ~20% IPC improvement, with as much extra clock as possible while remaining in the same power envelope as Zen 3. One area AMD could improve is their multithreaded FP scaling, which DDR5 should provide some good gains.
I'm really happy we're back to the times when every generation actually gives us a reason to upgrade.
Caporegime
What? Why?
Because they're not up to par with Intel or Apple due to memory bandwidth limitations.
You can see on SPECint2017 (the average score), 5950X with 16 cores is tied to 12900K in MT while more or less tied in ST and having more cores. And in SPECfp2017 it actually falls behind 12900K (and M1 Max) noticeably. Intel and Apple would likely scaled just as poorly if they were on DDR4 (indeed you can see 12900K DDR4 doesn't scale well either).
This is because each DDR5 channel is really two 32-bit channels instead of one 64-bit channel (as it was in DDR4), and each bank can operate independently of each other, doubling burst length (e.g. two 64-byte ops instead of one during the same time) and getting much better efficiency at utilising the bandwidth. Like this isn't an issue in Threadripper/EPYC as they can get more channels, but a serious issue limiting usecases of 5950X for workstations doing certain workloads. Seriously dual-channel DDR4 was not enough for 16 cores of Zen 3.
This doesn't show itself in gaming or rendering (those never saturate memory bandwidth and are more sensitive to latency), so the limitation isn't visible on gaming/blender/cinebench tests, but it is quite visible in SPEC lbm, cam4, fotonik, roms, etc... These are more representative of scientific/financial compute workloads, code compiling, particle simulations, monte-carlo simulations, etc...
Once AMD goes to DDR5 they will very likely scale as well as Intel and Apple (as AMD does scale well in the Threadripper/EPYC range, so no architectural limit there), and their MT FP performance would blow Intel out of the water given the # of core advantage.
Caporegime
Because they're not up to par with Intel or Apple due to memory bandwidth limitations.
You can see on SPECint2017 (the average score), 5950X with 16 cores is tied to 12900K in MT while more or less tied in ST and having more cores. And in SPECfp2017 it actually falls behind 12900K (and M1 Max) noticeably. Intel and Apple would likely scaled just as poorly if they were on DDR4 (indeed you can see 12900K DDR4 doesn't scale well either).
This is because each DDR5 channel is really two 32-bit channels instead of one 64-bit channel (as it was in DDR4), and each bank can operate independently of each other, doubling burst length (e.g. two 64-byte ops instead of one during the same time) and getting much better efficiency at utilising the bandwidth. Like this isn't an issue in Threadripper/EPYC as they can get more channels, but a serious issue limiting usecases of 5950X for workstations doing certain workloads. Seriously dual-channel DDR4 was not enough for 16 cores of Zen 3.
This doesn't show itself in gaming or rendering (those never saturate memory bandwidth and are more sensitive to latency), so the limitation isn't visible on gaming/blender/cinebench tests, but it is quite visible in SPEC lbm, cam4, fotonik, roms, etc... These are more representative of scientific/financial compute workloads, code compiling, particle simulations, monte-carlo simulations, etc...
Once AMD goes to DDR5 they will very likely scale as well as Intel and Apple (as AMD does scale well in the Threadripper/EPYC range, so no architectural limit there), and their MT FP performance would blow Intel out of the water given the # of core advantage.
Yes
I'm looking at that and i'm thinking the 16t 5800X is ahead of the 20t Coffee Lake and marginally ahead of the 16t Rocket Lake.
The 5900X is ahead of the 12900K with DDR4 and the Apple M1 is using some form of close to die fast memory.
In conclusion i would suggest that Zen 3 MT FP performance is very good, it just lacks DDR5
Soldato
So early??
They are already late. Zen 3 was launched in 2020..
Link to where AMD said they would be releasing earlier?
In every interview, official presentation they said h2 this year, from what i know h2 is july to december, so that guy that said in april is just having fun with us.
Caporegime
I wonder what AMD are doing for Computex in May.
INT/FP performance difference in modern CPUs has more to do with number and structure of execution units
And then limited by package power (I suspect 12900K pulls ahead because its power limit is twice as high)
But sure, DDR5 will give Zen4 more room to do more/fatter execution units.
And then limited by package power (I suspect 12900K pulls ahead because its power limit is twice as high)
But sure, DDR5 will give Zen4 more room to do more/fatter execution units.
Soldato
Digitimes reporting that 5nm AMD Zen4 shipping in September. Falls in line nicely with the early peek due at Computex later this month.
Caporegime
That and RDNA3 GPU's are launching in October, November or December. probably October / November.
You would think AMD would want the new CPU's out before then to give them the best chance for a good none bottlenecked showing on the bar charts.
I'm all in for September, seems just right. Busy latter half on the year, Nvidia will join the party with new GPU's too while nothing but crickets out of Intel's GPU camp i bet. eep eep eep....
Last edited: