AMD Ryzen 5 3600XT / Ryzen 7 3800XT / Ryzen 9 3900XT Linux Performance In 130+ Benchmarks

[r1348]

Fine, they're overclocking the better silicon now that they have better yields.
Still no reason to move out of my 3700X, but they could be interesting for those upgrading their systems now, still I'd wait for the 4000 series to be out.

[brunosalezze]

Wow, the 3950x is perf-per-watt champion, didn't expect that.

3950x is best of the best silicon, in order to fit 16 cores in a 105w tdp

[mppix]

"Over the entire span of benchmarks, the Ryzen 9 3900XT saw a peak power draw of 144 Watts with an average of just 75 Watts compared to the Core i9 10900K having an average power draw of 133 Watts and a peak of 379 Watts."

nice amd

14+^+ on overdrive

[mlau]

3950x is best of the best silicon, in order to fit 16 cores in a 105w tdp

I was more surprised that it even beat intel, which in my opinion has far better cpu power management. And that makes it even more impressive.

[mczak]

3950x is best of the best silicon, in order to fit 16 cores in a 105w tdp

That may be true, but it would only contribute a tiny bit to it being a perf/w leader. The truth is all these cpus are operating so far above their ideal (as far as efficiency goes) operating points it's not even funny. With 16 cores the 3950X, having the lowest "tdp per core" among the tested cpus, is just forced to get a bit closer to that point, increasing efficiency quite dramatically.

[pal666]

The chiplet with worse silicon goes in a Ryzen 5 or 3 or whatever. That's the one of the main benefits of not making a single die.

they already make those from chips which couldn't make say 3800x. so now your suggestion is they take 10% best 3800x chips and make 3800xt, and from 90% worst make ryzen 3(ryzen 5 is still two chiplets)? doesn't look like good busisness.
my solution is simpler: they just improved process performance so that it can do 100 mhz higher with same or better yield than before

[caligula]

"Over the entire span of benchmarks, the Ryzen 9 3900XT saw a peak power draw of 144 Watts with an average of just 75 Watts compared to the Core i9 10900K having an average power draw of 133 Watts and a peak of 379 Watts."

nice amd

Just wow Reminds me of the Pentium 4 days. Intel hasn't had this shitty silicon for a long time. Why would anyone buy this piece of sh*t hardware even if it was 2% faster in some CPU intensive single thread games? After all, buying Intel basically means that you don't want any progress in CPU design ever again. Handing over your money to Intel means, please don't develop any new process nodes in the next 10 years and yes, please invent gaping new on-chip security holes that require mitigations with double digit % perf impact.

[starshipeleven]

they already make those from chips which couldn't make say 3800x. so now your suggestion is they take 10% best 3800x chips and make 3800xt, and from 90% worst make ryzen 3(ryzen 5 is still two chiplets)? doesn't look like good busisness.

My suggstion is that they bin the chiplets and assign them to a product line depending on quality AND production quota. The chiplets can go from the higher end CPUs to the lower end ones.

At the end of the day, it hurts your bottom line to flood the market with high end parts so there is a fixed quota of that, once that is fulfilled, everything is moved to midrange or low end.

[Teggs]

chiplets yo.

The chiplet with worse silicon goes in a Ryzen 5 or 3 or whatever. That's the one of the main benefits of not making a single die.

Aren't they the same all the way up and down the stack? A Ryzen 3 3100 is just a failed EPYC 7742 chiplet, isn't it?

Apart from that, I'm surprised that the scaling was so consistent. Either AMD sent Michael three golden samples, or the advertised increase in performance is actually there in every application for XT vs X parts.

[morydris]

This is a brilliant article, I especially love the "per-watt" and "per-dollar" views.

Only one gripe: PLEASE drop CoreMark. It's completely unrealistic code and you'll find all academics railing on it as it strongly favors simple naive cores, cores that won't do well on real workloads.