AMD Ryzen 9 7950X3D Linux Performance

[zboszor]

But ECC UDIMM does that even exist for DDR5? I think that on-chip ECC will be confused alot with real 72-bit ECC UDIMMs.

ECC DDR5 is 80 bits, not 72 as it contains 10 chips. Look at https://www.crucial.com/memory/serve...2085s1ec48ba1r

Also: https://www.youtube.com/watch?v=Fc7qFzawKzw
AsRock X670E Steel Legend supports ECC with the latest Agesa / BIOS upgrade.

[GruenSein]

We might be headed for an era of really complicated schedulers. AMD land has its more-cache-vs-more-clockspeed tradeoff, and Intel land (as well as a few generations of ARM) has its performance-vs-efficiency tradeoff. And our OS is supposed to know how to do the right thing? Shiit, even I don't know what the right thing is, other than "try both, then measure the difference."

Actually, the "try and measure"-approach is not uncommon in HPC for example. Some CFD solvers perform these kinds of heuristics with respect to nodes and domain decomposition because similar domain sizes don't necessarily mean that the computation time will be the same depending on what's going on there and how well it converges etc. This can also deal with non-homogeneous node setups. So a similar approach might be possible for schedulers with respect to threads and cores. However, there are also some differences. One is that measuring performance in a general sense independent of the type of workload might be tricky. In CFD, it is relatively simple. You can measure the computation time per time step. In the real world, the workloads be subject to some kind of intermittency causing them to benefit from cache only sometimes. Another concerns the fact that there is not just one workload being distributed but multiple ones that might have different priorities. So, while I think that this heuristic approach has some benefits it is still far from simple.

[unwind-protect]

I hear that the scheduler in Windows for the Intel performance and efficient cores is partially driven by hardcoding best hand-measured results for select popular applications. Willing to believe it, graphics drivers do that all the time.

[Mahboi]

Thanks a lot for the SIMDJson in extended testing.
It's odd how inconsistent some of it is, like sometimes a 5800x3D obliterates a 5950x but the 7950x3D will barely do above a 7950x. I expect AMD will do as AMD does and FineWine their microcode in the future.

However there's a bit I don't get, the Gb/Watt throughput largely favors Intel, since the 13900K apparently averages around a paltry 50W. How's that possible? Isn't this one a chonkster that's expected to average around 135W? Any idea how this result came forth Michael?

[Michael]

Thanks a lot for the SIMDJson in extended testing.
It's odd how inconsistent some of it is, like sometimes a 5800x3D obliterates a 5950x but the 7950x3D will barely do above a 7950x. I expect AMD will do as AMD does and FineWine their microcode in the future.

However there's a bit I don't get, the Gb/Watt throughput largely favors Intel, since the 13900K apparently averages around a paltry 50W. How's that possible? Isn't this one a chonkster that's expected to average around 135W? Any idea how this result came forth Michael?

Last check IIRC the simdjson upstream benchmark test itself was predominantly single threaded (and can leverage AVX-512) hence the low overall CPU power consumption there.

[Mahboi]

Last check IIRC the simdjson upstream benchmark test itself was predominantly single threaded (and can leverage AVX-512) hence the low overall CPU power consumption there.

Would make sense that some results didn't go on the 3D Vcache cores and some did then. That would explain the whole "5800x3D beats 5950x by far but 7950x3D barely beats or equals the 7950x".

Strange that SIMDJson's benchmarks do not use multicore. But oh well. I got the info I wanted anyway. Thanks!

[Mahboi]

It is pretty wild to watch my 5950X getting smoked like that so soon. Although contemporary CPUs blew out the power budget and shoved in as much cache in as you could fit, so they probably won't be able to keep up this pace of improvement in subsequent generations.

We shall see yet. Zen 4 is ultimately a Zen 3 upgrade and a bridgehead for AM5. The goal of Zen 4 is to not be terrible, to stabilise the 3D Vcache beta that was the 5800x3D, and to put some AM5 motherboards in computers.
Once the AM5 motherboards guarantee a number of return users for Zen 5 and quite certainly Zen 6, they can focus on truly novel stuff. The rumor mill has been alive with Zen 5 for a long time, with words such as "true revolution" and "great leap forward", or perhaps that was Mao.

I wouldn't be shocked if Zen 5 shatters expectations, and if it finally passes that 16 core bridge that they've been sitting on since Zen 1. Time will tell.

[Mark Rose]

The luxury, my 1541-clone only had 170kb floppies

Such extravagance, my Apple IIe had 113.75 KB floppies.

[Rabiator]

I am not too worried that Windows is getting a noticeable advantage out of having this.

From an admittedly cursory look at this article and a Windows launch review of the 7950X3D, it seems that Linux is actually the better platform for this CPU.
On Windows the AMD trades blows with Intel's 13900k, being better at games but somewhat behind on applications. On Linux the big picture looks like a clear win for the 7950X3D, with some minor wins for Intel but massive wins for AMD.

[F.Ultra]

Such extravagance, my Apple IIe had 113.75 KB floppies.

And it was an Apple you get doubly my condolences