AMD Ryzen Threadripper 7980X & 7970X Linux Performance Benchmarks

[edwaleni]

One thing the review didn't say is that Zen4 Threadripper lacks Pluton just like Zen4 Epyc lacks it:
https://www.phoronix.com/forums/node/1416513

Pluton has an OEM problem right now. Dell, Lenovo arent going for it.

As for Pluton....here is what Dell thinks;

“Pluton architecture depends on support within the SoC – versus a separate “chip” – and is a different approach to hardware security capabilities. At this time, Pluton does not align with our hardware security approach in our commercial PCs. You’ll see new Dell commercial laptops with 12^th Gen Intel Core processors coming soon – these will continue to include Trusted Platform Module/TPMs (TCG certified and FIPS 140-2 level 2 validated) to protect the device. But with all new technologies, we will continue to evaluate Pluton to see how it compares against existing TPM implementations in the future.

And when it comes to endpoint security, our holistic security approach includes both software-based, “above the OS” protections and hardware-based, “below the OS” capabilities to protect against traditional/emerging attacks and threats at the deepest levels of a device,” the statement added. “These investments over the last decade allow us to provide the industry’s most secure commercial devices to businesses. “)

[JEBjames]

Michael

typos page 2

"One important bit that cane up" should be "came"

"Threadripper 7 000 series" should be "7000"

typos page 13

"peak of 378 Wats" should be "Watts"

p.s. Thanks for bringing us this crazy amount of testing

[Michael]

Michael

typos page 2

"One important bit that cane up" should be "came"

"Threadripper 7 000 series" should be "7000"

typos page 13

"peak of 378 Wats" should be "Watts"

p.s. Thanks for bringing us this crazy amount of testing

Thanks!

[F.Ultra]

Interesting, good to know. I guess the surprise for me is that the 7980x lost out to the 7970x in lower-threaded scenarios. But, the boost clock is supposed to be lower on the 7980x (5.1GHz) than the 7970x (5.3GHz - is that only for all-core boost, though?).

As for the comment about scaling cores with threads, I am well aware of that, which is why the allmodconfig kernel compile looks like one of the few that can actually saturate the threads on the 7980x...

The base clock is also higher on the 7970x, considering that it also have the same area to release heat with fewer CCD:s probably means that the 7970x have a much easier time hitting the boost (and for more cores at the same time) than the 7980x.

[Keats]

What strikes me the most is how well the 7970x stacks against 3990x even in benchmarks that scale well with more cores.

[coder]

Incredible code compilation performance. I have been waiting to see how these would perform for months, because the high end Ryzen 79xx series has been so potent. Interestingly it appears the CPUs are either power starved (by this specific motherboard, maybe?) or just thermally constrained so that they can't really stretch their legs.

Probably I/O bottlenecked, since the 1 TB WD SN850 isn't exactly a top performer.

Michael, how many SSDs do you use for running a benchmarks like these? Is it one per machine, or do you just use the same drive and just move it from one machine to the next?

[coder]

I don't see any unusual performance issues.

Take another look at the compilation benchmarks. They're not scaling very linearly.

For each die you add, you're losing a decent chunk of performance due to the added latency.

BS. Plenty of other benchmarks, like rendering, don't seem to have a problem scaling much more linearly.

Schedulers aren't always so perfect, a lot of tasks can only scale up to a certain amount of threads, and not all tasks that can scale indefinitely will do so linearly.

Compiling these large codebases should scale very linearly. There's oodles of concurrency.

[Michael]

Probably I/O bottlenecked, since the 1 TB WD SN850 isn't exactly a top performer.

Michael, how many SSDs do you use for running a benchmarks like these? Is it one per machine, or do you just use the same drive and just move it from one machine to the next?

Same drive moving around

[aviallon]

Same drive moving around

I was wondering, could you test an allmodconfig kernel build with some of the best CPUs from this benchmark, but with the kernel sources and the build output ("everything") in a tmpfs.
It should be interesting.

[schmidtbag]

Take another look at the compilation benchmarks. They're not scaling very linearly.
BS. Plenty of other benchmarks, like rendering, don't seem to have a problem scaling much more linearly.
Compiling these large codebases should scale very linearly. There's oodles of concurrency.

C'mon now, you're not dumb. You know damn well that not everything scales linearly. You also know reasons why that might happen, such as (but not limited to):

• Scheduler issues or glitchy power governors
• Added latency from inter-die communication (an issue TRs in particular have historically had great issues with)
• Memory and/or cache bandwidth limitations (compared to Epyc, TR is quite crippled in its memory specs)
• Thermal throttling (some tasks are more power-hungry than others)
• VRM quality
• Amdahl's Law

While I agree a large codebase is one of the few tasks I would expect to scale linearly, some of the things I mentioned could affect it more than other benches that did manage to scale linearly.