AMD Ryzen 5 7600 / Ryzen 7 7700 / Ryzen 9 7900 Linux Performance | Hacker News Reader

Ryzen 9 7900 Linux Performance

phoronix.com

102 pointsmfiguiere40 comments

Power consumption summary

> The Ryzen 5 7600 / Ryzen 7 7700 maintained around a 70~71 Watt average with a peak of 91 Watts while the Ryzen 9 7900 maintained a 73 Watt average and similar peak of 92 Watts. This is much lower than the Ryzen 7000 X-series parts and much lower than the Intel competition -- the Core i5 13600K meanwhile had a 95 Watt average with a peak of 167 Watts.

Performance / efficiency

> If taking the geometric mean of all 300+ Linux benchmarks conducted, here is the positioning of this assortment of Intel and AMD CPUs. The Ryzen 9 7900 was at 95% the speed of the Ryzen 9 7900X (while on average at 66% the power consumption) and was still fast enough to come just ahead of the Core i9 12900K Alder Lake processor. The Ryzen 7 7700 was at 96% the speed of the Ryzen 7 7700X and nearly matching the Core i5 13600K performance. The Ryzen 5 7600 lastly was at 96% the performance of the Ryzen 5 7600X while tieing with the prior generation Core i5 12600K and the prior Zen 3 Ryzen 9 5900X. When factoring in the CPU power consumption of these "65 Watt" parts, the performance is all the more impressive especially compared to Intel's real power figures.

Both X (AMD) and K (Intel) lean heavily towards "gib last 5 % perf" in order to win benchmarks in reviews across their line-ups. The non-K 13th gen SKUs aren't out yet, but in the largely similar 12th gen they also enjoyed substantially lower power consumption.

Unless you are using a laptop (as these processors aren't mean for) or are running a datacenter (which these processors aren't meant for) does anyone really care about 65 vs 150w differences?

Yes, to avoid getting a space heater when you turn on your computer during the summer. In winter it's not really that bad, as you get both compute and heating from the same device.

Lower power bills help too. And let's not forget the ridiculous power numbers that new GPUs are consuming.

if only 100W would produce any meaningful amount of heat, that would be wonderful. A simple space heater runs at 1500W and can only barely help heat a medium sized room as supplemental heat. Unfortunately (for heat purposes), computers waste a bunch of that electricity on actually doing computations.

The electricity spent doing computations is still turned into heat.

High-end GPUs have been around 250-300 W for a very long time (with the odd xx95 X2 Ultra Titan model consuming more). "Ridiculous power numbers of new GPUs" is a pure-bred meme.

GTX 1080 was 180W. I think everyone agree that 1080 was high-end at that time. What's happening now is that xx80 series now become ultra-high-end silently, as we can also see price rising.

Even if power consumption is not a concern, there's a need to account for power dissipation. Using 65W allows for a low-noise air cooler or a nearly-silent AIO, and it also allows for smaller form factors with compromised cooling solutions.

Actually 65W allows for fully passive cooling solutions. See e.g. https://streacom.com/products/fc8-fanless-mini-itx-case/

I personally do not care that much about peak power, if I am actually crunching numbers I might as well do it as fast as possible.

However I care a lot about idle power consumption, with a desktop computer turned on most of the day, these small inefficiencies really add up.

As a concrete example, my desktop GPU has a well known bug that makes it consume 30 W more idle than it should. For an always on desktop, that's 260 KWh wasted per year.

in europe that's easily 130 euro nowadays, crazy

I am running my 7950X at 105W, because I get 97%-99% of the performance on my own workloads and extremely quiet operation. At 65W I'd lose too much performance and wouldn't gain anything meaningful; at 170W the PC would be audible.

As others have said, there are dissipation considerations in a desktop, too. I would absolutely hate having to work next to a jet-engine or having a fridge-sized rig to fit some absurd cooler. Others don't care as much. My 135W-rated Xeon does get audibly loud despite the big-ass Noctua, especially when the GPU next to it is also working hard. If it wasn't during gaming, when I can't hear it, I would hate it.

But I also think the actual usage profile should be considered. Typically, my CPU waits around doing nothing. Occasionally, I ask it to build whatever project I'm working on. I can tolerate some noise during this time if it's not too long, and if it noticeably shortens the time I have to wait for it. So if the idle consumption is similar, but it can get very fast when needed, I'm all for it.

> I would absolutely hate having to work next to a jet-engine or having a fridge-sized rig to fit some absurd cooler.

Even SFF cases can fit 240mm or 280mm AIOs these days.

Old standby heat sinks like the Noctua NH-D15 can cool 200-300W CPUs quietly, and they fit in common cases. They have a 120mm tall version for slightly smaller cases that doesn’t give up much in benchmarks.

I think the cooling challenges have been overblown, honestly. It’s kind of trivial to quietly cool powerful CPUs these days.

> My 135W-rated Xeon does get audibly loud despite the big-ass Noctua,

Unless there’s something wrong with airflow or your mounting, your fan curves are almost certainly too aggressive.

A lot of motherboards, especially workstation and server boards, will ramp fans to 100% by default. You probably don’t need that. Adjust the curves until it’s quiet, don’t worry about getting the absolute lowest possible temperatures. This is what OEMs do. If you build your own, you’re expected to do it yourself.

The NH-D15 is what I actually have, and all other fans are Noctua save the GPU which has whatever the manufacturer figured was good enough. Now the PC isn't outrageously loud, but the fan hum can become tiring after a while if the CPU is busy. It's almost inaudible when idle.

Maybe there's an issue with general airflow in my case (an older Fractal Design Define R something).

Important for small form factor PCs, which are basically ITX PCs in the smallest possible case, often below 10L in size.

In my experience computers last longer when cool, and even more important is a minimum number of heat cycles. Things like capacitors on motherboards use chemicals that break down over time, and more quickly if they are hotter. Sure you can buy capacitors that will last 10 years in challenging conditions, but I doubt any consumer motherboards do.

Ideally my desktop lasts 10 years, even if it ends up in some kind of secondary usage scenario. To that end I want it silent, well cooled, and to experience minimum temperature swings.

I usually end up with 2x 140mm intake fans, a 120mm fan on top of the CPU, and one of the decent PSUs that go to zero RPM when at low loads. In a case like this a 65w TDP CPU (that's normally going to idle at way less) is MUCH easier to cool than a 150w-300w (the common range for the alder lake Intel chips). No exotic cooling (water, water blocks, piping, reservoir, pump, etc) needed, and not a big deal if the bearings loose some lubrication or dust accumulates on a filter.

Sure would some CPU intensive workloads go 10-20% faster at 2x-3x the power? Sure. Do I care, no. I'd rather spend the extra purchase price, cost of electricity, cost of cooling, and related on ECC memory. I don't really want to be able to hear when my desktop is busy, should be silent all the time.

I'm using a 5950X, and in the BIOS there is a setting where you can enter the power limit in Watts. So while the "X" CPUs consume more power as configured out-of-the-box, you can still set your own limit as desired.

I have a 3200G and I limit it to 35W(of 3 built-in values, 64, 45 and 35) since I run it off a picoPSU. Haven't done much testing without the power limits, but it's plenty peppy at 35W.

These lower power limits reduce performance a lot on chiplet CPUs (unlike your monolithic APU). The impact is maybe a little less crass on Zen 4 because the 7nm IOD seems to use less power than the old 12nm one used in Zen 2/3.

For example, a 5600X in CB R20

   PPT   Freq      Points (nT)  Core power
   82 W  4500 MHz  4350         54 W         (stock)
   65 W  4200 MHz  4000         43 W
   38 W  2500 MHz  2400         17 W         (lowest allowed)

The SoC/uncore consumes around 15-20 W on Zen 2/3, so at 38 W you only leave 15-20 W for the actual CPU part of the CPU. Meanwhile the low-load and idle savings are essentially nil because all of the idle power is the IOD/uncore.

According to https://www.anandtech.com/show/17641/lighter-touch-cpu-power... Intel is much more power sensitive than AMD is. In other words, limiting Intel power consumption will hurt performance a lot more than limiting AMD power consumption does.

I just read someone posting in the Intel subreddit who was concerned about the temps he was seeing on his 13900k that was being overclocked and was hitting 100C on an All In One Water cooler. Someone chimed in for the amount of power it was consuming (350+ watts) that it was normal. Yikes. All that power for a few extra FPS.

Good example of "input=voltage^2*quantity*size; output=voltage*quantity*size" as the napkin math of transistor processors.

Are there any idle power measurements in that article? For 95%+ of the time my home computer isn't doing a lot (text editing, simple browsing, playing radio or music).

https://www.tweaktown.com/reviews/10309/amd-65w-zen-4-ryzen-...

> Looking at our CPU power draw alone, the 7900, 7700, and 7600 all pull around 20 watts at idle.

For comparison, the Intel Core i9-13900K seems to draw just 10 watts, and the Intel Core i5-13600K only draws 6!

https://www.tweaktown.com/reviews/10222/intel-core-i9-13900k...

That's interesting.

According to TFA, the i5-13600K has the best "performance / cost" ratio. So, all other things equal, for a "regular PC", the intel part looks the most interesting, since it won't be any slower when you need it to actually do something for a few minutes (like compile something).

Though - now I see on the first of the links I posted, they show much higher idle power draw for the Intel parts, around 22-24W. I wonder why they have different numbers for the same parts on different articles.

Good catch. Loaded figures for both and idle for AMD seem the same between the two articles...

It could be the different motherboards. Phoronix used an ASUS for their Intel tests while TweakTown had a Gigabyte.

Lots of tweaks on motherboards, i.e. give the CPU a bit more voltage than standard and it will boost more and your motherboard gives users more perfomance.

Also just differences in measuring, was it from the wall or via voltage pads on the motherboard.

The code compilation benchmarks on page 3 of this article [1] make me think about building a 7950X box as a CI runner. Just a massively brute force way to reduce continuous integration build and test times. :)

Has anyone benchmarked these against typical CI cloud providers, or deployed self-hosted runners like this? I’m curious if it’s truly possible to reduce CI times -50% to -75%. If so, huge win for developer happiness… and hardware and power costs would be pretty insignificant compared to the happiness and productivity gains.

[1] https://www.phoronix.com/review/ryzen-7600-7700-7900-linux/3

At my company we rent a couple of dedicated Ryzen 9 5950X boxes from Hetzner to host our Gitlab CI runners, for a bit over $100/month each.

These rip through our pipelines so much faster (and cheaper) than any managed CI offering we've tried, that we didn't even bother doing proper benchmarking.

I would assume 7900X or 7950X would be even better :D

And unfortunately a similar instance at AWS would cost 2000$/month, even with spot instances it’d be 500$ at most which is still huge. I am aware of why cloud exists, but this is just sad. Have been using Github Actions and Bitbucket Pipelines, they both suck and no good observabilit exists for them.

AVX-512 for the masses! It's going to be interesting to see what we can do with this now that many will have it in the coming years.

One of the big near term beneficiaries is going to be RPCS3. Wonderful deep dive into what AVX512 means for it here - https://whatcookie.github.io/posts/why-is-avx-512-useful-for...

Any word on affordable motherboards and RAM for these CPU's? It seems kind of silly to put a $229 CPU into a $350 motherboard...

The B650 motherboards are starting to show up at around $165-170 (though plenty that are $230+)

https://www.newegg.com/asrock-b650m-pg-riptide/p/N82E1681316...

https://www.bhphotovideo.com/c/products/Motherboards/ci/1986...

https://pcpartpicker.com/products/motherboard/#s=41&c=160

I don’t think so. I’ve always bought /really good/ motherboards and then adequate CPUs etc to abuse. And upgrade later.

I've always considered the upgrade route when purchasing motherboards, but I've never done so over the 32 years I've owned PC's. In the beginning things were moving too fast but lately it was because I was really lucky picking my CPU 10 years ago. Sandy Bridge was the last 25%/generation improvement from Intel, we've had 10%/generation improvement ever since. So that Sandy Bridge has held up surprisingly well. Everything but the CPU & MoBo has upgraded - 5 SSD's, 3 GPU's, RAM, monitor, keyboard, even the case, but it's still the original CPU & MoBo. This year I'll complete the Ship of Theseus and replace the CPU & MoBo and whatever else has to get dragged along (aka RAM).

I have a question as I have no experience with it but does the Kernel play an important role regarding performance? So does Kernel 5 perform a lot better than 4?

Reply on news.ycombinator.com