Intel Core i9-13900T CPU benchmarks show faster than 12900K 125W performance

The benchmarks themselves aren't really deceptive, but wccftech's reporting on them is misleading. These aren't claims being made by Intel, but just random user benchmarks that appeared online.

For all benchmarks, context matters. These days, most in-depth reviewers understand how to work with transient power and thermal limits with respect to time. For example, measuring the MacBook Air M1's peak speed/time vs. its lower sustained speed after it hits thermal throttling around the 5-7 minute mark.

It's kind of hard to compare performance/watt meaningfully when the wattages aren't at all similar.

If you limit power on current generation Intel/AMD processors, the perf/watt is pretty good; depending on the specific load, sometimes better, sometimes worse than M2. When you allow more watts, perf goes up, but perf/watt goes down as you see diminishing returns. (With some loads, you may also have a point where more watts reduces performance, whoops)

Geekbench isn't great as it doesn't even attempt to capture power usage, but that may not be available or accurate on all systems anyway.

In Linux you want powercap tools. You can tune all three parameters in terms of power and time constant, so you can do exactly what you suggest. You can change it on the fly without rebooting so you could have a ridiculous setup like mine where I have the power limits wired up to two unused buttons on my mouse.

If Vim isn’t responsive in a machine with hundreds of MHz something is very wrong with your setup. Vim was fast on machines with dozens of MHz. Either a bad terminal emulator or some plugin making it slow, I’d guess.

If you're running on Linux, almost all of this is completely configurable using ryzenadj for AMD laptops or powercap and thermald for Intel chips. Here's a document on how you might go about adjusting this w/ powercap: https://github.com/junaruga/framework-laptop-config/wiki/Imp...

For anyone looking for a good summary of the knobs available for AMD: https://github.com/FlyGoat/RyzenAdj/wiki/Renoir-Tuning-Guide

Or for Intel: https://github.com/lhl/linuxlaptops/wiki/2022-Framework-Lapt...

CPU throttling happens on many different timescales. That's why Intel processors have multiple power limits (PL1, PL2, PL3, PL4), in addition to current and temperature-based throttling mechanisms. The time constant for PL2 is usually either 28 seconds or 56 seconds. At those timescales, the concern is usually not with power delivery or CPU die temperature but rather with exterior case temperatures that the user may be directly touching.

Based on the reported benchmark performance, it seems very unlikely that temperature-based throttling kicked in, and it's clear that the chip was operating well above 35W for at least a large portion of the benchmark run. So the PL1 and PL2 turbo power limits are the relevant controls at play here.

I was curious about exactly how much burst heat could be absorbed, so I asked WolframAlpha [0]. In a 15" workstation laptop, I think the CPU could quite reasonably pull +100 watts over steady-state TDP for 30 seconds. (100-gram aluminum heatsink absorbing 100W * 30sec -> ∆T = +33ºC)

[0] https://www.wolframalpha.com/input?i=%28100w+*+30+sec%29+%2F...

It still feels insane that a CPU can draw that much power. I mean that's like over 8 amps at 12 volts for doing... well nothing really. It's 100 watts of pointless resistance losses.

Something that could've powered like half of an electric scooter at top speed. And GPUs are like 3-4 times worse.

You may also be interested in the DB4 [0]. I can't justify the cost for the aesthetics of fanless but perhaps you can. :)

0. https://streacom.com/products/db4-fanless-mini-itx-case/

And a ton of people are cross-shopping more expensive newer parts with still-available 5000 series CPUs which are much cheaper. There’s an interesting question about whether absolute performance is more important or if you should weight that against cost and/or power.

The GN benchmarks recently have done a good with explaining both the cost perspective and the performance-per-watt.

Edit: for reference, Newegg currently has the 5950x and 7950x at $500 and $600 respectively. And that’s before the higher platform cost for AM5/DDR5.

The 7950X3D is not going to be faster in non-game benchmarks. Most likely it will be a bit slower in most synthetic tests and non-game workloads (rendering, encoding, FEA etc.)

It's a fair point of reference for those of us with a 5950X, which was a top performer a couple years ago and is still a damn lot of CPU no matter how you slice it.

Living in a place where electricity costs are very high, such as California, the AMDs have much better price to performance when compared to the Intel CPUs over a period of 3-5 years (which is how long I normally wait before upgrading CPUs).

Also to mention, 5950X has long been out of stock in many places.

Makes sense. T series chips generally aren't as widely available as individual components as their mainstream higher power counterparts. To your point, they are available from OEMs as complete systems, who seem to get priority.

There is advice to just buy the non-T series and limit power. Interesting to see that, at least in this example, they aren't quite equal.

I see this frequently, but what is “better silicon”? Physically something is off about the manufacturing which impacts performance, but does not kill the chip. What are those defects?

There are 0.1% bad transistors instead of 0.2%? Heat output is more uniform? The bad transistors are clustered in such a way that signal routing is more efficient and leads to a measurable throughout difference?

I saw a HN comment that chips that perform well at lower voltage don't scale as well to higher power, and vise versa. By that logic, T-binned chips would perform better at a low TDP compared to a downclocked 13900.

Yeah, exactly. I was able to get a decent ~50W drop and stayed at a higher than base clock rate (1920 MHz @ 900 mV). It's still a space heater, but it's better.

I also frame rate limit myself in a lot of games - I don't need my MMOs running at 160 fps, so many games I'm at 150-200 W. I still wish it was less, but that's much more reasonable than 400 W.

> (I've undervolted AMD cards that come with ridiculous core voltages without having to drop clocks at all in past)

It's often even worse than that. There are plenty of cases where you can undervolt so far that you now have enough headroom in the power delivery and cooling to allow you to run at substantially higher clock speeds.

Depends what you mean. It generates more heat, it’s just more effective at getting that heat off the die and into your room.

Edit: since it’s performance per watt is higher, if you’re capping frame rates then you can get less heat out of the 4090. Like I said, depends what you mean.

For the most part it seems like in Intel are throwing a ton of power into their CPUs to get these results.

So it's more like putting a lipstick on a long dead design. Sure may look fresh from afar, but it still stinks.

What I am trying to say is that at this level of power consumption Intel should be many times faster than M1 - that is offering substantially better performance per watt.

Sure there are niche applications where power consumption doesn't matter and only the single core performance counts, but could older tech be overclocked and achieved the same result?

Not sure. Point being is that what Intel does is increasing pollution and causing people to bin perfectly working machines just because there is a "new" CPU on the block, where in real life they probably won't see a difference.

It's really irresponsible thing to do for Intel. They should go back to the drawing board and stop releasing meaningless products until they actually get something worth upgrading to.

In a Laptop its quiete stupid to stay at 100w just to have a slightly better Performance. I mean its only a significant margin as Long as you can get the heat away, which is often the biggest problem in most laptops…

(M1 Max uses half the power for geekbench 5 and ~1650 and being a 2 years older chip. If the m2 max can hold the Power limit and still get to ~2k it would be fastly superior, i.e. the normal m2 is already at ~1850 With Less than half of the Power of the 13900t)

Each brand has its strengths and weaknesses. AMD isn't dominating anymore, but they are still ahead in areas such as AVX512 (absent in Intel Core), the ability to mount more than 8 performant cores in a chip (Intel's intra-core communication architecture is difficult to scale, hence their new P+E approach, which fortunately seems to work well), and their ability to take advantage of the multi-chip process to mount absurd amounts of cache. The multi-chip approach is also a tech advantage that allows them to cut costs.

> What confuses me is that AMD still has a huge lead for server CPUs. I have not seen massive adoption of AMD in the cloud yet.

This was years ago so the tide may've shifted but: part of it could still be vendor experience and "it works"-experience?

When EPYC gen 1 & 2 came out, we were shopping for a bare metal megaserver spec at work. I got a budget and free rein (except it had to be Dell), and I really wanted to pick EPYC for the better core clocks (which were significant in our build architecture) and more cores and better cost.

With one EPYC spec and one Xeon Gold spec built (EPYC was I think $13k vs $15k?), work was a bit uneasy about AMD processors just yet. Our workload was MSVC compilation but they were concerned about architecture differences, since all of our workstations & laptops were Intels. They preferred paying more because "we already have Xeon servers and they're proven".

So, we ended up getting the Xeon Gold spec instead.

(Disclaimer: I work at Intel, but not on server products. Opinions are my own.)

There's a lot more that goes into a server platform than just the cores, for instance: the BIOS code, the BMC support, the maturity of the motherboard designs, etc. These are all areas where Intel seems to still have an edge -- but I'm also very excited about our upcoming server architectures on the core/compute side.

Intel can reach equal or better performance than AMD only at higher power consumption.

At equal number of active cores and equal power consumption, an AMD CPU has a much higher clock frequency than an Intel CPU. At equal number of active cores and equal clock frequency, an AMD CPU has a much lower power consumption than an Intel CPU.

For desktop and laptop CPUs, Intel can win with this higher power consumption strategy, especially in single-threaded applications, where the fact that Raptor Lake has a higher IPC than Zen 4 also helps.

On the other hand, for server and workstation CPUs, Intel cannot benefit from this trick, because those CPUs are used in multi-threaded applications where they run constantly in thermally-limited conditions.

The Intel Sapphire Rapids CPUs have from many points of view a superior design in comparison with AMD Genoa, but they are severely handicapped by the inferior manufacturing process used by Intel.

Because of that, the Sapphire Rapids CPUs have low clock frequencies when all cores are active and small cache memories. This ensures that they are no match for AMD Genoa and they barely match the previous generation of AMD server CPUs.

The manufacturing process disadvantage is compounded by the stupid Intel policy of randomly disabling various features in most SKUs, with the exception of those that are ridiculously expensive.

This policy ensures that the benchmarks published for Intel are excessively optimistic, because they show the top SKUs, while the SKUs that most people have are crippled, so they may have a much lower performance.

So in servers the Intel CPUs are not competitive at all, except for various special applications where the CPU performance matters very little, e.g. when the cheapest SKUs are good enough for providing a big memory and many PCIe lanes, or when it is acceptable to pay Intel to enable the included accelerators, because the application can benefit from them.

It's all about the "With enough thermal headroom". If this isn't specified in the benchmark, then the benchmark is useless because you can't compare with your setup. It's even truer in laptops, but this is true of servers and desktops as well.

In addition, recent CPUs will use this thermal headroom to the limit.

> Complex devices need complex benchmarking, unfortunately. You won't get a simple, single number that shows how powerful a cpu is.

This is the gist of my comment (see after the first phrase), so I think we agree.

Yeah, I think having both is exactly what should be done (in addition to talking about the thermal setup).