Ryzen 5 7600X part 1: initial tests

To give some more advanced testing the time and effort it deserves while not delaying the info I’ve got any longer than necessary, I’m splitting this into several parts (part 1, part 2, part 3).

I don’t have any kind of pre-release access, had to spend a while driving to buy it, still have to write this, and the gaming numbers I wanted to put in part 1 got too GPU-bound to be useful, so tests don’t go too deep yet. I’ll dive into memory and infinity fabric details in part 2, gaming performance in part 3, power and core overclocking details in part 4, and possibly more after that. (Edit: I’m going through these points in a different order.)

An R7 5800X is what I have handy to test against. In some cases (labeled 5800X 6C on charts) I’ve run an extra round of testing with it cut down to six cores to be a better point of comparison for the 7600X.

I’m focused on a different market segment than most reviewers implicitly are. Usually, high-end CPUs with weak value propositions steal the show, and even the mid-range CPUs are supported in testing by beefy coolers, high-end motherboards, and the fastest RAM that takes little tuning (3600 14-14-14 for DDR4 recently). There’s nothing wrong with building that kind of system if it’s what you want, but it’s not very cost-effective for the average gamer. I want to know how the value stuff holds up.

With DDR5 and motherboard pricing as it is, the value situation isn’t very good right now, but it’ll get better and most of this should stay relevant. The 7600 non-X could also give it a big boost whenever AMD decides to release that, assuming it as usual doesn’t lose much performance versus the 7600X.

The gaming test situation

(I’ll put gaming results in part 3.)

Having only an RX 6600 XT handy to test with, I made some extreme preparations to get cleanly CPU-bound numbers, but it wasn’t good enough and a lot of my numbers got tainted by GPU limits anyway. That’s despite game settings far more biased towards the CPU than almost anyone would realistically use, including in some cases barely playable low resolutions. I tried to leave a particular margin of GPU performance for the faster CPU when figuring out the settings, and the games that got GPU-limited mostly did so by getting bigger gains than the games with clean results. Therefore, if I went ahead and published the clean numbers I’ve got, it’d undersell the CPU. I’m just going to redo all of this for part 3.

I don’t think this makes the 7600X’s gaming performance pointlessly high even with this GPU. The gaming scenarios that are most realistically demanding of a CPU don’t tend to be easy ones to test. I had no hope of covering those trickier cases in the time I’ve had to test so far, but should be able to manage at least one for part 3.

Memory and infinity fabric

(I’ll expand on this in part 2.)

I’ve got the 7600X paired with a Corsair 2x16GB 5600 36-36-36 kit run at XMP, which looks like the cheapest around that should use Samsung dies (but I haven’t confirmed the dies yet). The 5800X is paired with some 2x8GB Samsung C-die rated for 3600 19-20-20, again run at XMP. The C-die is a touch slower than realistic for a 2022 value build’s DDR4, but should be much closer to correct than I can emulate with B-die.

Zen 4 has memory bandwidth problems. It’s still not entirely clear where they come from and how impactful they are. The most obvious explanation for the moment is that the infinity fabric is too narrow (which looks to be true to some extent regardless), but measured numbers don’t line up quite as expected for that. I’ll do my best to get to the bottom of it in part 2.

I’m testing with fclk at 2000 despite the RAM being at 5600 because that’s the motherboard default and therefore how the average gamer would run it. (AMD has been saying it should be 1866 in this case.) The higher fclk seems likely to be the better default.

Heat, voltage, and frequency

(I’ll expand on this in part 4.)

Both CPUs are tested under a be quiet! Pure Rock 2, a 120mm tower cooler of fairly average performance. The fan curve goes to 100% at each CPU’s soft throttling point (90°C for the 5800X and 95°C for the 7600X) to keep as close to ideal performance as possible on workloads that get that hot, but drops quickly below that, staying very quiet in most scenarios. (This is the fan curve I run for normal use, not just testing.)

The 7600X never hits 95°C anyway in the tests I’ve run so far. Cinebench and Blender are the hottest; in these, it does temperatures in the upper 80s, 1.15 to 1.20V, and 5.2 to 5.25 GHz. In lighter workloads, it spends a lot of time as high as 5.45 GHz. It idles at only 1.25V. That tuning feels nicely relaxed in contrast to desktop Zen 3’s high-strung idle behavior.

While the stock tune is undemanding of cooling, that could change quickly when overclocking.

Motherboards

The motherboard situation is lousy for the moment, to the point it completely ruins the value proposition. It should get much better when B650 and B650E launch on October 3rd. In the meantime, boards start at $250-ish and go way up from there, without delivering much of anything to justify those prices. Don’t buy in this week like I did.

The ASRock X670E PG Lightning I’m testing with was the cheapest option available. It should be far overkill in ways that affect performance, but is severely lacking in attention to detail and has multiple major coil whine issues. The performance-affecting default settings seem sane (after enabling XMP), and I’m sticking with those defaults for initial testing.

It takes about 30 seconds to train the memory on every boot. Hopefully BIOS updates will cinch this down.

Web responsiveness tests

There’ll always be some poorly coded website out there to bring any CPU to a crawl. Even at more normal levels a disproportionate amount of computing happens through a web browser, so anything that can be done to speed it up has disproportionate effects on how snappy the overall computing experience feels.

Speedometer benchmarks changes to the displayed page rather than pure Javascript.

JetStream’s separate treatment of average speeds, first-run speeds, and worst-case speeds is welcome. Averages alone miss a lot of ways that Javascript can be slow.

x264 encoding

This is re-encoding a 22-minute 1080p 60 fps Planetside 2 recording. The settings are something you might use for streaming if CPU encoding on a separate system: 6 kbps at the fast preset.

This workload can mostly but not quite entirely saturate these core counts, so the hexacore 7600X being able to beat the octacore 5800X is a good result.

Cinebench R23

This makes at least a good reference point, since basically everyone publishes Cinebench numbers.

As above, beating the 5800X while short two cores on the multi-threaded portion is a good score, but 14232 isn’t a very strong 5800X showing. This particular 5800X can be undervolted into the 15000s trivially on worse cooling, but with the stock tune thermal throttling is responsible for a lot of this gap.