Intel Core i9-14900KS Draws as much as 409W at Stock Speeds with Power Limits Unlocked

[Tek-Check]

360 AIO or don't bother? This seems so pointless. Besides, the 14900K is already well positioned enough.

HU tested previous KS models even with 420mm AIO, and the CPUs throttled after a few seconds.

 

[freeagent]

400w at full load with a monster of a cpu like that is to be expected I would think.. how hard can it be to cool if you are already setup for loads like that

 

[Vya Domus]

Rumour says that the next gen won't have HT anyway. I'm curious if it's true and how it'll do. I'm so used to seeing HT/SMT on every CPU by now that I can't even imagine what life is like without it.

That's a really bad sign, it means they're clutching at straws trying to limit power draw any way they can. Since turning SMT off reduces utilization and makes everything less efficient doing this only ever makes sense if you are trying to limit power consumption at all costs.

 

[phanbuey]

That's a really bad sign, it means they're clutching at straws trying to limit power draw any way they can. Since turning SMT off reduces utilization and makes everything less efficient doing this only ever makes sense if you are trying to limit power consumption at all costs.

That's true if you have a homogenous design. If you have a ton of efficiency cores that are faster than a logical thread, and your primary cores are faster without sandwitching in threads during idle time, then in theory HT isn't necessary.

 

[Vya Domus]

If you have a ton of efficiency cores that are faster than a logical thread, and your primary cores are faster without sandwitching in threads during idle time, then in theory HT isn't necessary.

It doesn't make sense to compare an E core to a logical thread, if you disable SMT you are inevitably making that core less efficient, I don't mean just power wise but also performance wise. It's a serious regression architecturally. The whole purpose of SMT was to increase the utilization of resources that would otherwise be idle.

 

[bug]

It doesn't make sense to compare an E core to a logical thread, if you disable SMT you are inevitably making that core less efficient, I don't mean just power wise but also performance wise. It's a serious regression architecturally. The whole purpose of SMT was to increase the utilization of resources that would otherwise be idle.

It goes both ways. It can help use idle resources, if available. But it can also cause resource starvation under heavy load.
Intel engineers are not dumb. They know the usage patterns much better than you and I.

 

[AnotherReader]

It goes both ways. It can help use idle resources, if available. But it can also cause resource starvation under heavy load.
Intel engineers are not dumb. They know the usage patterns much better than you and I.

It varies from application to application so from a performance perspective, the current situation is the right one where the user has the option to disable it. If the rumour is true, then it's more likely driven by the idea of reducing the attack surface in shared environments like the "cloud".

 

[CrAsHnBuRnXp]

Has Intel just not learned that people dont want high wattage CPUs?

 

[Chrispy_]

Does anyone care any more?

Gamers - one of the few remaining demographics who care about single-theaded or low-threaded performance are buying X3D chips that are just better despite far lower power draw.

Everyone else who needs more performance will just add cores; 7950X, ThreadRipper (Pro), EPYC, Xeon.

Pushing so far beyond the point of diminishing returns of efficiency is stupid. It has always been stupid, and it will remain increasingly stupid as the environmental and monetary cost of energy usage increase exponentially.

 

[Vario]

Well by the sounds of it im going to have to buy a case that supports at least 5 rads lol

How about a 1080 external rad, go big. That should be able to dissipate 400 watts. You could even do a 1260.

 

[Deleted member 237813]

yet an optimized 7800x3d with 30-40 watts in gaming equals it. who the fuck is still buying intel for gaming who has clue?

 

[Vario]

yet an optimized 7800x3d with 30-40 watts in gaming equals it. who the fuck is still buying intel for gaming who has clue?

Product is more for non gamers that are limited to an intel ecosystem. I don't think most people will be buying this.

 

[Makaveli]

409 Watts jesus!

 

[Sabotaged_Enigma]

Well who's using water cooling can try connecting it to room heater and use 14900KS to make through winter lol

 

[Chrispy_]

Product is more for non gamers that are limited to an intel ecosystem. I don't think most people will be buying this.

I'm curious if such a demographic even exists any more.

Just about the only possible answer would be hardcore AVX-512 users and just about every situation I can imagine where heavy AVX-512 is needed, you also need gobs of bandwidth. Consumer i9's are godawful for that which is why they're all using Xeons for AVX-512 workloads.

Am I missing something, is there a productivity workload that doesn't work on Zen4 and needs AVX-512, but not with significant memory quantities or bandwidth? AMD are ahead of intel in just about everything except AVX-512 which is already pretty niche outside of the datacenter and ECC-equipped HEDT pro-tier workstations running Xeons already.

 

[Ferrum Master]

I'm curious if such a demographic even exists any more.

Just about the only possible answer would be hardcore AVX-512 users and just about every situation I can imagine where heavy AVX-512 is needed, you also need gobs of bandwidth. Consumer i9's are godawful for that which is why they're all using Xeons for AVX-512 workloads.

Am I missing something, is there a productivity workload that doesn't work on Zen4 and needs AVX-512, but not with significant memory quantities or bandwidth? AMD are ahead of intel in just about everything except AVX-512 which is already pretty niche outside of the datacenter and ECC-equipped HEDT pro-tier workstations running Xeons already.

AVX512 in general is useful only for some json, limited simulations and PS3 emulation and benchmarks. The gains in daily user tasks are minuscule as desktop CPU ain't really meant for that, AVX2 was more or less enough.

 

[phanbuey]

It doesn't make sense to compare an E core to a logical thread, if you disable SMT you are inevitably making that core less efficient, I don't mean just power wise but also performance wise. It's a serious regression architecturally. The whole purpose of SMT was to increase the utilization of resources that would otherwise be idle.

That's true but it adds a ton of latency, so while you're increasing the utilization, you're actually decreasing the ability of the core to respond, and decreasing the single thread IPC -- if you have a ton of efficiency cores sitting idle while that's all happening, then you're sub-optimizing for the hybrid architecture.

Turn off SMT on the 7800X3D, and games that aren't threadbound get 5-10% boost.

So if you can limit the wait times, make a more efficient and responsive P core, while offloading the HT style work that would happen in-between high priority instructions to e cores, you're actually making the overall process much faster. -- less theorhetically efficient at full thread load from the point of view of a single P core, but more specialized and more responsive overall.

If eliminating HT hardware also allows you to fit more P cores on the die with a space and a TDP reduction then that win is compounded further.

 

[GhostRyder]

I mean, I am all for it for the overclocking enthusiast! Give us a chip that can handle crazy power limits. My biggest concern is going to be cooling because its already a struggle on the 14900K. Even with a really good 3x120mm AIO it will throttle when told to unlock max performance so I can imagine needing heavy custom cooling for this (Which is really what you should run anyway on something like this). My question will be how good these will be on the P-Core overclocking and how much further without ridiculous power scaling they will go.

 

[bonehead123]

Spend more, burn more, yea, ok, right....hehehe

 

[bug]

Well who's using water cooling can try connecting it to room heater and use 14900KS to make through winter lol

Neah, winter's almost over. By next winter, we'll have a newer, better heat source

 

[AnotherReader]

That's true but it adds a ton of latency, so while you're increasing the utilization, you're actually decreasing the ability of the core to respond, and decreasing the single thread IPC -- if you have a ton of efficiency cores sitting idle while that's all happening, then you're sub-optimizing for the hybrid architecture.

Turn off SMT on the 7800X3D, and games that aren't threadbound get 5-10% boost.

So if you can limit the wait times, make a more efficient and responsive P core, while offloading the HT style work that would happen in-between high priority instructions to e cores, you're actually making the overall process much faster. -- less theorhetically efficient at full thread load from the point of view of a single P core, but more specialized and more responsive overall.

If eliminating HT hardware also allows you to fit more P cores on the die with a space and a TDP reduction then that win is compounded further.

The cost of SMT is far less than its detractors think. The designers of the unreleased EV8 with 4-way SMT reported a 6% increase in die area over a single threaded equivalent. This was over twenty years ago and today the cost is likely to be even smaller. SMT is useful in the case where IPC is low due to low cache hit rates or other non-execution related bottlenecks.

 

[Chrispy_]

The cost of SMT is far less than its detractors think. The designers of the unreleased EV8 with 4-way SMT reported a 6% increase in die area over a single threaded equivalent. This was over twenty years ago and today the cost is likely to be even smaller. SMT is useful in the case where IPC is low due to low cache hit rates or other non-execution related bottlenecks.

Yeah, from what I read of Zen 1's architectural deep dive (where SMT was closer to Intel's HT than Bulldozer's huge amount of pipeline sharing) the actual implementation is almost free in terms of die area.

It should be something the OS scheduler can turn off and on at will, rather than this dumb all or nothing approach we have now where we have to choose to cut our thread count in half just to get some latency gains. We only need those latency gains on ONE thread

 

[HBSound]

ITX motherboards with 24-pin and 8-pin connectors shouldn't have issues powering this CPU at full capacity.

 

[dalekdukesboy]

So much for new and improved in reference to efficiency….

 

[marios15]

You can run a 128core EPYC at 360W, that's 16x more (real) cores