Ryzen 9 3900X vs. Core i9-10900X

[scan80269]

I'm curious to know which of these two CPUs can yield faster iteration times at the current wavefront (FFT 5120K):

1. AMD Ryzen 9 3900X - 12 core, 3.8GHz base, 64MB L3, dual channel DDR4-3200, 105W TDP

2. Intel Core i9-10900X - 10 core, 3.7GHz base, 19.25MB total cache, quad channel DDR4-2933, 165W TDP, AVX-512

The Ryzen has much larger total cache while the Cascade Lake-X supports quad channel DDR4 and AVX-512.

[nomead]

I don't have access to it anymore, but a friend ran mprime benchmarks on his 3900X when he received it - a single run, no tweaks, DDR4-3000. One worker at 5120K = 2,44 ms / iter.

I would be interested in a comparison too, as even with the processor price cuts, the i9-10900X system will be more expensive (motherboard plus four sticks of memory).

[ATH]

My guess would be the i9-10900X because of the quad channel RAM and AVX512. Memory speed is the most important thing for Prime95.

I wish I had access to computers like this to test.

[scan80269]

There's a post in the perpetual benchmark thread for Core i7-9800X Skylake-X with DDR4-3600 memory showing approximately 2.7ms/iter for 5120K FFT.

Assuming i9-10900X with 3600 memory would be close to this, it looks like 3900X may have a slight edge, possibly because of the FFT fitting into the L2/L3 cache, along with power & cost advantages.

[Mark Rose]

Originally Posted by scan80269

There's a post in the perpetual benchmark thread for Core i7-9800X Skylake-X with DDR4-3600 memory showing approximately 2.7ms/iter for 5120K FFT.

Assuming i9-10900X with 3600 memory would be close to this, it looks like 3900X may have a slight edge, possibly because of the FFT fitting into the L2/L3 cache, along with power & cost advantages.

The L3 cache cache on the 3900X is split in four 16 MB chunks, one per CCX, and it's a victim cache, and not unified, so it's unlikely the FFT is fitting. The 3900X is basically two 3600 put together, but they'll share memory bandwidth so I don't expect the 3900X would be much faster than a 3600 for that matter. The 3600 with 3600 MHz memory is probably a sweet combination.

[nomead]

Originally Posted by Mark Rose

The L3 cache cache on the 3900X is split in four 16 MB chunks, one per CCX, and it's a victim cache, and not unified, so it's unlikely the FFT is fitting. The 3900X is basically two 3600 put together, but they'll share memory bandwidth so I don't expect the 3900X would be much faster than a 3600 for that matter. The 3600 with 3600 MHz memory is probably a sweet combination.

Well the R5-3600 (with DDR4-3600 memory, too) is 5.01 ms/iter at 5120K... so yes, there's quite some difference at that FFT size. The Ryzen 3600 chip is quite sweet for double checking though (I have one), but there's a big drop off in performance above about 3456K FFT size, again somehow pointing to the FFT plus some extra data fitting inside that 32 MB L3 cache. Whether it's in chunks or not, it seems to matter.

See the comparison plot in another thread.

[Mark Rose]

Originally Posted by nomead

Well the R5-3600 (with DDR4-3600 memory, too) is 5.01 ms/iter at 5120K... so yes, there's quite some difference at that FFT size. The Ryzen 3600 chip is quite sweet for double checking though (I have one), but there's a big drop off in performance above about 3456K FFT size, again somehow pointing to the FFT plus some extra data fitting inside that 32 MB L3 cache. Whether it's in chunks or not, it seems to matter.

See the comparison plot in another thread.

Interesting! Thanks for the link.

[simon389]

I don’t understand how the AMD 3900X has a faster FFT benchmark than my Intel 9800X (see benchmark thread) considering the 9800X has quad channel memory and the 3900X has only dual channel.

Does a faster FFT mean it can crunch PRPs faster as well?

[xx005fs]

Originally Posted by simon389

I don’t understand how the AMD 3900X has a faster FFT benchmark than my Intel 9800X (see benchmark thread) considering the 9800X has quad channel memory and the 3900X has only dual channel.

Does a faster FFT mean it can crunch PRPs faster as well?

Yes, though it really depends on the FFT size. Since Zen 2 have massive L3 caches, it can somewhat compensate for the dual vs quad channel memory. However, as FFT size becomes larger, the L3 cache can hold less and less portion of the FFT size, and the rest has to be stored in memory. That's why Zen 2 will be faster than 9800x on current wavefront.

[Prime95]

Originally Posted by simon389

I don’t understand how the AMD 3900X has a faster FFT benchmark than my Intel 9800X (see benchmark thread) considering the 9800X has quad channel memory and the 3900X has only dual channel.

Does a faster FFT mean it can crunch PRPs faster as well?

Be careful in what you are measuring. The benchmark thread may have data on which CPU could finish a single PRP test fastest, while most users care about throughput - how ll/PRP tests can the CPU finish in a given unit of time.

[scan80269]

Originally Posted by simon389

I don’t understand how the AMD 3900X has a faster FFT benchmark than my Intel 9800X (see benchmark thread) considering the 9800X has quad channel memory and the 3900X has only dual channel.

Does a faster FFT mean it can crunch PRPs faster as well?

@simon389 for your 9800X results posted to the benchmark thread last year, was AVX-512 enabled for Prime95?