Hacker News Reader

Comment by rational_indian | Hacker News Reader

I have a Ryzen 2700X. Is Ryzen 9 3900X going to be a good upgrade for me?

Id wait for the next set of chips. There is IPC and clock gains to be had, but unless you are often cpu-bound, I think your money would be better spent or saved

Depends on what you do. You'd gain 4 cores and 8 threads and also the single core performance is higher. If have some direct benchmark comparisons under https://www.pc-kombo.com/us/benchmark/apps/cpu/compare?ids%5.... The blender benchmarks do not show a big difference, but things like 7zip and AES-Encoding do, so it really depends on the workload.

That Blender result is nonsense, they screwed up the test somehow. This is a better comparison:

https://www.anandtech.com/bench/product/2125?vs=2519

The 3900X has 50% more cores and Blender is embarrassingly parallel, so it's more than 50% faster, as expected.

That's the real question for whether it's worth it. For single threads the 3900X will be 10-20% faster. For well-threaded applications it's >60% faster.

Keep in mind that the 2700X still has a resale value of over $200, so if you're selling one and buying the other it's not a very expensive upgrade. But the value proposition is a lot better if you can use the extra cores. (Of course, if you can't and you just want the single-thread performance improvement, there's always the 3700X.)

It's a presentation issue. The benchmark data agrees with you, the transformation from "lower is better" to "higher is better" makes the result misleading. That needs a better solution, showing that the difference is actually quite high.

It's not a presentation issue, the pc-kombo.com site is just a garbage. In one of the Blender graphs it has the 3900X as "10839" and the 2700X as "10732" (very close together), but it links to the Anandtech results where the actual numbers are 161 and 268, not close together at all and completely different numbers. In another graph it links to a "source" where the 2700X isn't even on the page.

Their graphs are completely fictional, probably as a result of bad data scraping code since it looks like they're just pulling results (inaccurately) from other sites.

Look at the worst result in the anandtech benchmark. Remember what I said about converting from "lower is better" to "higher is better". It is a presentation issue.

If the 2700X is not on the source page that page either got changed or it's a combined benchmark that was presented in multiple articles. In which benchmark do you see that?

This is that page:

https://www.pc-kombo.com/us/benchmark/apps/cpu/compare?ids%5...

This is the Anandtech benchmark that it links to from the first Blender graph (you can click on the graph):

https://www.anandtech.com/bench/CPU-2019/2247

The numbers on that page for the 3900X and 2700X bear no resemblance to the ones in the graph that links to it.

Compare to their first 7zip graph which links here and contains the actual numbers for the 3900X and 2700X, because they evidently scraped that one correctly:

https://www.anandtech.com/bench/CPU-2019/2240

This is the page they link for the second Blender graph:

https://www.computerbase.de/2019-11/amd-ryzen-threadripper-3...

The numbers in their graph don't match what that page has for the 3900X and the 2700X doesn't appear on that page at all.

Computerbase hides some benchmark results by default. Click at the top right "+29 Einträge" to see 29 processors more, including the 2700X.

I understand that the number in that specific benchmarks is confusing right now. Again, that's from the conversion done for the "lower is better" benchmarks. That's why it's not the same number, just the same distance. I hope a better solution can be used soon.

It's not the same distance though. The way to convert "lower is better" to "higher is better" is to invert the results, e.g. convert seconds per render to renders per second. You take 1/X where the original number was X. Or do renders per day if you don't want results as small fractions, which just means multiplying the inverted results by a constant, e.g. the number of seconds per day.

For both the Anandtech and Computerbase Blender results, the 2700X takes ~60% longer in reality, which is ~66% more renders per second (which is the same percentage more per day), but the difference in the length of the bars on pc-kombo.com is <10% for Computerbase and ~1% for Anandtech. Something has gone wrong there.

It's the same numerical distance. Look how small the bars are at the anandtech page, https://www.anandtech.com/bench/CPU-2019/2247 also does not show you the one is 66% faster - same problem, just inverted. But the original is easier to get, and 1/X style comparisons could work - or in general just "how much faster" bars. Thanks :)

> It's the same numerical distance.

...

What people are going to care about is the difference as a percentage. :)

> Look how small the bars are at the anandtech page, https://www.anandtech.com/bench/CPU-2019/2247 also does not show you the one is 66% faster - same problem, just inverted.

Actually the bar for the 2700X on the Anandtech page is 66% longer than the bar for the 3900X, though it's admittedly harder to eyeball when they're both that short because of the much slower processors at the bottom of the page. You can imagine how much easier that page would be to see if they didn't include those two Celerons at the bottom. (You can also imagine how that would affect the results using the math you were using; if adding or removing a CPU changes not just the length of the bar but the actual numbers for an entirely different CPU, something's fishy.)

Doing 1/X for timed benchmarks helps that for the modern processors though, because it makes it so the bars that are short and hard to eyeball are the slowest processors rather than the fastest ones.

Reply on news.ycombinator.com