Hacker News Reader

Comment by neogodless | Hacker News Reader

I do think TechSpot[0] (aka Hardware Unboxed) has a good chart[1] that highlights some differences in power consumption utilized to achieve performance.

Notice at 185W, the Ryzen 9 7950X is churning out about 36% higher performance than the Core i9 13900K. The Intel CPU, however, is then able to continue drawing up to 335W!

At their respective peak scores, the i9 is using over 70% more power.

[0] https://www.techspot.com/review/2552-intel-core-i9-13900k/

[1] https://static.techspot.com/articles-info/2552/bench/Power_S...

View on HN

I'm mostly interested in numbers under 100W, maybe even 65W, since I plan to down-clock my (future) Zen 4 machine.

AMD is killing it with efficiency there.

I've seen several posts showing that the most minor of underclocks and undervolts produce substantial power savings and minimal performance differences.

Apparently Intel's trying really hard on an old process to get the performance crown, forcing AMD to clock higher than they'd like to compete. It's ridiculous that they double the power for 10% performance. But the nice thing is you can half the power and take the 10% hit.

Also note that the 7900/7900X are dual chiplet design that has a significant power impact. The 7700x and below have a single chiplet and saves power, and also increases single thread perf. If you look at the gaming benchmarks the 7700x is often better than the 7900x, not to mention all the Intel chips except the top of the line i9.

So if you buy the 7700x and underclock just a bit you should have a very efficient desktop.

> the 7900/7900X are dual chiplet design

This is a minor terminology mix-up. The term you want here is "CCX". Each CCX holds multiple chiplets.

See https://www.tomshardware.com/reviews/amd-ccx-definition-cpu-... for more details.

EDIT: It does look like my own information on this was rusty.

A "chiplet" is not an individual core, but a trimmed down die that does not have a memory controller. A "CCX" contains cores (within a chiplet) and L3 cache.

https://www.hardwaretimes.com/amd-ccd-and-ccx-in-ryzen-proce...

> The basic unit of a Ryzen processor is a CCX or Core Complex, a quad-core/octa-core CPU chiplet with a shared L3 cache.

> For the Ryzen 3000-series processors, CCX refers to a four-core grouping inside of each core chiplet die (CCD). For example, an AMD CPU with 8 cores will have a CCD with two 4-core CCXs.

From that link, your statement does not appear accurate. Multiple CCX's appear on a given Core Chiplet Die. There are indeed two vs one core chiplets on different Ryzen designs (plus the IOD).

check out https://arstechnica.com/gadgets/2022/10/intel-i9-13900k-and-... - they power limited both 7000 and raptor lake models and surprisingly, amd doesn't respect the power limits - at 65W it's burning 90W and intel is respecting the 65W limit, 7.868Whr vs 8.58Whr to complete the same task (h264 encoding in handbrake), so interestingly intel on a 2 node disadvantage is more power efficient if you set your power limits.

Yes - comparing the above 65W score from AMD to Intel's 145W score highlights the stark difference!

That techspot review triggers me. Why does the guy go on and on about how impractical the 13900 is compared to the 12900, since the 13900 is "throttling"? Yeah, the 13900K hits a steady state of 5300MHz with all cores active and stays that way forever, which is great! The apparently "more practical" 12900K hits 4900MHz on all cores and does the job slower but it's "not throttling". This makes zero sense.

The popular tech press understanding of how closed-loop control systems work is just REALLY annoying.

And it is important to remember the AMD IOD is still on N6 or similar to current Intel 7 2nd gen. ( i.e On I/O level AMD and Intel are on the same Node ) Even if Intel was on Intel 5 or TSMC N5 equivalent it still likely wouldn't be using similar power as AMD.

So I am actually surprised at how well AMD Zen 4 is doing. It used to be if you want MT you go with AMD, at the expense of ST. Now this is no longer the case.

If you care about energy efficiency, just buy AMD Zen 4.

I think this mix of processes is testament to the fact analog is not scaling, and high speed IO has a lot of analog/phys. So having an IOD on 3nm would cost more and probably consume similar power, at least for the phys. Of course you will always get sub systems on that die that would benefit, but overall its likely a net loss.

but I do think Raptor lake consunes less power than Zen4 when idling e.g. web browsing, code editing etc. I would want to see power usage for VS code + Firefox etc.

I use high performance only for seconds when compiling.

Zen's idle performance is absolutely atrocious due to the I/O die that constantly draws 20 watts. Pretty sure Intel's stuff in total draws less than half that.

Intel invested a crapload of money in laptop CPU's and low idle which is one of the reason they dominated that market for so long - I would guess that it bled over into desktop as well.

Are you talking about last years zen3? That was a common complaint.

From https://www.tweaktown.com/reviews/10195/amd-ryzen-5-7600x-ze...

"Power consumption is measured directly from the dual eight-pin connections. At peak, the 7600X pulls an impressive 130W, and during idle, it was down to 18w."

Sounds pretty good, especially considering its at the "dual eight-pin" power supply connection. Not laptop class, but pretty good for a high performance desktop.

Anecdotally, my Zen 3 5900X rated at 105W TDP, sitting kind of idle (with a dozen or so programs open, task manager showing 1% usage) is drawing about 11W.

I do seem to recall a AGESA update (motherboard firmware) that did managed to reduce the IOD power usage when idle. Sad that it was months after most of the reviews came out.

12900K idles at 8.5 watts. 18 watts is awful.

Measured where? At the power supply?

Not sure what exactly is included in the 18 watts via the power supply connector. Wasn't sure if it was all CPU, or CPU+Dimms, or maybe the entire motherboard.

The 8 pin connector goes to the CPU. The big ATX 12V connector goes to everything else.

For idle power consumption, watt usage from wall is better measurement.

Benchmark with same power limit like this is what I'd like to know. I also want to know wall watt usage for them, to verify it's limited correctly, and how much chipset/other components uses power.