Originally posted by mrg666
View Post
Announcement
Collapse
No announcement yet.
AMD Ryzen 9 7950X & Intel Core i9 14900K: Ubuntu 22.04 vs. 23.10 vs. 24.04 Linux Performance
Collapse
X
-
Originally posted by mrg666 View PostThat 7950x is such a nice CPU. Though, I am resisting the temptation to upgrade my 5950x systems.
And, to those claims about stability patch, 14900k is already at a "low" 132w power level. Stability patch will not affect the performance, I think. That CPU goes past 250w when the power is not limited; that is where problems start.
with an upgrade from 2 ram channels to 4 ram channels doubling the performance(but not for the 2 DIMM designs)
and then they could go from DDR5 to DDR5 MRDIMM 17 600 MT/s what would also again double the performance.
there is already the threadripper platform with 4 ram channels but thats not designed to also have APUs
so a hypothetical AM6 with up to 32cores and 4 ram channels and also made for APUs whould be pretty nice.
TDP could go from 170watt to 250 watt for the 32cores.
Phantom circuit Sequence Reducer Dyslexia
- Likes 1
Comment
-
Originally posted by qarium View Post
a AM5 socket system is maybe not what you want... just imagine a AM6 socket system made for up to 32cores
with an upgrade from 2 ram channels to 4 ram channels doubling the performance(but not for the 2 DIMM designs)
and then they could go from DDR5 to DDR5 MRDIMM 17 600 MT/s what would also again double the performance.
there is already the threadripper platform with 4 ram channels but thats not designed to also have APUs
so a hypothetical AM6 with up to 32cores and 4 ram channels and also made for APUs whould be pretty nice.
TDP could go from 170watt to 250 watt for the 32cores.
Comment
-
Originally posted by mrg666 View PostFor that, I would definitely skip the AM5 and wait for AM6. I agree that 32 cores might be happening in the consumer market as SMT is going out in the coming Intel generations. AMD could do the same.
that big.little is death you see in raspberry pi 5 vs orange pi 5
raspberry pi pi5 does not have big.little design
the orange pi 5 has big.little design and the benefit is near to zero. means according to becnchmarks raspberry pi 5 and orange pi 5 have the same performance.
just think about this hypothetical AM6 32core cpu 16cores could be Zen5 and 16core could be Zen5c... but because it is not for the notebook/laptop market the benefit is nearly zero and a 32 core zen5 would be better.
right now amd only does a mix of zen4 and zen4c in the notebook oriented APUs even if they also release it for AM5..
in the next 1-2 years amd will clearly release a AM6 platform because their threadripper platform is to expensive for the mass market. and AM5 will not scale well for more than 16cores. .
Phantom circuit Sequence Reducer Dyslexia
Comment
-
Originally posted by qarium View Post
the qualcomm elite X proof that not only SMP/hyperthreating is death in the consumer market but also big.Little design is death to.
that big.little is death you see in raspberry pi 5 vs orange pi 5
raspberry pi pi5 does not have big.little design
the orange pi 5 has big.little design and the benefit is near to zero. means according to becnchmarks raspberry pi 5 and orange pi 5 have the same performance.
just think about this hypothetical AM6 32core cpu 16cores could be Zen5 and 16core could be Zen5c... but because it is not for the notebook/laptop market the benefit is nearly zero and a 32 core zen5 would be better.
right now amd only does a mix of zen4 and zen4c in the notebook oriented APUs even if they also release it for AM5..
in the next 1-2 years amd will clearly release a AM6 platform because their threadripper platform is to expensive for the mass market. and AM5 will not scale well for more than 16cores. .
Comment
-
Originally posted by mrg666 View PostI agree, one of the reasons I like AMD CPUs is that they are not big-little. When they drop SMT, I guess that will be another reason. I suspect Intel went with big-little out of desperation that they could not fit more cores without scaling down their process tech. Can you imagine 16 P cores in an i9 when 8 P cores can go past 250W?
their big-little design is really bullshit. also SMT it was nice tech with singlecore cpus or dualcore cpus or with quadcore or 8core
but with 16-32 cores SMT becomes a burden because the complexity to optimise the tasks for so many cores goes into a impossibility. just see Qualcomm elite X... no SMT and no big-little design and the benchmark results are very good.
a hypothetical AM6 socket will only have 250W TDP not because of the cpu cores its because of the big fat iGPU similar to AMD Strix Halo 40cU 2560shader units.Phantom circuit Sequence Reducer Dyslexia
Comment
-
Originally posted by qarium View Post
the qualcomm elite X proof that not only SMP/hyperthreating is death in the consumer market but also big.Little design is death to.
that big.little is death you see in raspberry pi 5 vs orange pi 5
raspberry pi pi5 does not have big.little design
the orange pi 5 has big.little design and the benefit is near to zero. means according to becnchmarks raspberry pi 5 and orange pi 5 have the same performance.
just think about this hypothetical AM6 32core cpu 16cores could be Zen5 and 16core could be Zen5c... but because it is not for the notebook/laptop market the benefit is nearly zero and a 32 core zen5 would be better.
right now amd only does a mix of zen4 and zen4c in the notebook oriented APUs even if they also release it for AM5..
in the next 1-2 years amd will clearly release a AM6 platform because their threadripper platform is to expensive for the mass market. and AM5 will not scale well for more than 16cores. .
By community request, we present our findings on how the AMD Ryzen 9 3900X performs with SMT disabled. This approach has potential, especially for gaming, because it ensures more physical hardware units are available for each thread, and could also benefit the processor's power management.
Then you should also notice that in multithreaded benchmarks, the rk3588 with 4 A76 + 4 A55 cores design has clearly much more horsepower than the 4 A76 cores of the raspberry pi 5: https://www.phoronix.com/review/rasp...5-benchmarks/3 (see specifically the JPEG XL benchmarks, plus compilation benchmarks on page 4).
Clearly you would have got a lot more performance with 8 A76 rather than 4+4 CPU, but rather than performance alone, big.LITTLE designs are to be compared in power consumption and efficiency against omogenous clusters, because efficiency for light tasks is the main intent of those kind of designs since they are ubiquitous in smartphones and similar devices.
Comment
-
Originally posted by blackshard View PostYou're drawing bad conclusions from bad preconditions. First of all, SMT on x86 is still giving great benefits in some number-crunching oriented tasks (those which, obviously, benefit most from multiple cores):
https://www.techpowerup.com/review/a...l-9900k/3.html
a ryzen-9-3900x has 12 cores this means it is far from the core count you need to eliminate any benefit from the overhead...
there are phoronix.com has benchmarks with real high core count cpus and it shows that with 64 or128cores SMT has near to zero benefit and in most benchmarks the overhead of generating more threats kills of your memory. means if you want 2GB per threat a 128core gpu want 256GB of ram if you activate SMP you need 512GB ram many systems can not even have 512GB ram
any benchmarks with 1-16 cores about SMT are invalide for the future and 32 core cpus will be the last generation with any benefit from SMT... and 64 or 128 or 192 or 256core cpus will not have any benefit from SMP what so ever.
and ARM cpus who are 8-wide decode path do not need SMP at all.
Originally posted by blackshard View PostThen you should also notice that in multithreaded benchmarks, the rk3588 with 4 A76 + 4 A55 cores design has clearly much more horsepower than the 4 A76 cores of the raspberry pi 5: https://www.phoronix.com/review/rasp...5-benchmarks/3 (see specifically the JPEG XL benchmarks, plus compilation benchmarks on page 4).
Clearly you would have got a lot more performance with 8 A76 rather than 4+4 CPU, but rather than performance alone, big.LITTLE designs are to be compared in power consumption and efficiency against omogenous clusters, because efficiency for light tasks is the main intent of those kind of designs since they are ubiquitous in smartphones and similar devices.
you pick out 2 benchmarks "JPEG XL benchmarks, plus compilation benchmarks" but if you watch all benchmarks the geometric overall performance benefit is very small to irrelevant.
also keep in mind that rk3588 in 8nm node is already obsolete try this again against the qualcome elite x produced in 4nm...
Phantom circuit Sequence Reducer Dyslexia
Comment
-
Originally posted by qarium View Post
you base your argument on outdated cpus ryzen 9 3900 and intel 9900K...
a ryzen-9-3900x has 12 cores this means it is far from the core count you need to eliminate any benefit from the overhead...
Originally posted by qarium View Postthere are phoronix.com has benchmarks with real high core count cpus and it shows that with 64 or128cores SMT has near to zero benefit and in most benchmarks the overhead of generating more threats kills of your memory. means if you want 2GB per threat a 128core gpu want 256GB of ram if you activate SMP you need 512GB ram many systems can not even have 512GB ram
If such a benchmark exists, I would deduct that the application is scaling rather poorly with high number of cores rather than blaming SMT of being ineffective.
Originally posted by qarium View Postany benchmarks with 1-16 cores about SMT are invalide for the future and 32 core cpus will be the last generation with any benefit from SMT... and 64 or 128 or 192 or 256core cpus will not have any benefit from SMP what so ever.
and ARM cpus who are 8-wide decode path do not need SMP at all.
Originally posted by qarium View Postthe performance benefit of the rk3588 with 4 A76 + 4 A55 cores do exist but in average its irrelevant.
you pick out 2 benchmarks "JPEG XL benchmarks, plus compilation benchmarks" but if you watch all benchmarks the geometric overall performance benefit is very small to irrelevant.
You can take one hundred of single core tasks, benchmark a single core cpu vs 64 core cpu and the geometric overall will say that both cpus are equal.
The "average" does not exist, the typical workload of the machine does.
Originally posted by qarium View Postalso keep in mind that rk3588 in 8nm node is already obsolete try this again against the qualcome elite x produced in 4nm...
Comment
-
Originally posted by blackshard View PostRyzen 3xxx is not outdated as you may describe. It's Zen2 architecture which is already manufactured and sold in many products, notebooks notably.
its outdated because on socket AM5 you can not even buy Zen2... you can only buy ZEN4...
the only product on the consumer market who is not yet clearly outdated is the zen2 based steam deck but its a 4core system who has such a low core count that it clearly benefit frm SMT..
this does not mean any high core count cpu has any benefit from SMT...
outside of steam deck no one produce a 4core you system anymore on socket AM5 the smalles cpu you can buy is a 6core ...
Originally posted by blackshard View PostUsing a very niche case and random numbers and then generalize is not a good idea.
If such a benchmark exists, I would deduct that the application is scaling rather poorly with high number of cores rather than blaming SMT of being ineffective.
you clearly did drink the kool Aid...
that application scale poorly on high core count cpus is a fact and by this fact alone SMT becomes useless
with that you just waste tranistory without geting any benefit.
Originally posted by blackshard View PostThis statement has to be proved, it's just your hypothesis backed by random numbers.
Originally posted by blackshard View PostThat geometric overall is based upon the selection of the benchmarks, so it has a relative significance.
You can take one hundred of single core tasks, benchmark a single core cpu vs 64 core cpu and the geometric overall will say that both cpus are equal.
The "average" does not exist, the typical workload of the machine does.
and the more tasks you do the more your user case profile is more like the geometric average...
and these people buy a 128core system or 192 core system and wonder why the magic bulled SMT no longer give them benefits instead it slows it down and it increase the complexitity of the software.
really no one cares if your 4 core cpu like steam deck or your 12core example still benefits from SMT
the 192core cpu does not. and thats the point.
Originally posted by blackshard View PostA matter which is totally off topic.
thats why i am sure the 12core qualcome elite X will become a great success.
its just a cpu without bullshit.Phantom circuit Sequence Reducer Dyslexia
Comment
Comment