AMD Ryzen CPU Core Scaling Performance

[Mark Rose]

Other reviews showed better performance for x+0 configuration instead of x/2 + x/2. That used to make sense, as data doesn't have to be transmitted in between the two CCX/L3$ pools (L3$ is still shared, so all cores have access, even the ones on the other CCX). However, some of Michael's benchmarks show the opposite, which is surprising.

Not surprising, if they're not running Linux 4.10+. Microsoft hasn't patched Windows yet, so the Windows scheduler is unaware of the L3 cache split, bouncing threads away from the cache with the needed data.

[schmidtbag]

Of course it works. It's the mediocre code created by incompetent game developers that won't scale.

Well to my understanding, HT works similarly to SMT, and HT was not known to match the performance of an entire CPU core, which is why I'm so impressed AMD did such a good job.

[dungeon]

Well Zened product all around, except gaming spike






This picture summarise everything fine, you have plus ~50% all around to previous top FX, beside gaming which is more like 22% Gamers probably expected there would be 50% also but it is not really

Now add about 15% on top of 6900K for gaming scenario and you have 7700K picture also

[indepe]

Not surprising, if they're not running Linux 4.10+. Microsoft hasn't patched Windows yet, so the Windows scheduler is unaware of the L3 cache split, bouncing threads away from the cache with the needed data.

Another (perhaps additional) explanation would be that, if the game threads don't share data, then 2+2 is better than 4-0 because the total L3 cache size available is twice as large (each CCX group has its own L3 cache).

[andrebrait]

What about the power governor? Performance was much better than ondemand the last time.

[eydee]

The Dota 2 Vulkan numbers are really strange. SMT aside, 4-0 is slower than 3-0 and 2-0. This doesn't seem to make sense. And all cores enabled is the slowest result of all? I'm not sure the explanation would be SMT, in this case.

RADV maybe. It is in kind of an alpha state, only is in Mesa because of demand and for more testing. Performance has always been low so far. Dota should be tested with both AMD's and nvidia's proprietary vulkan drivers.

[leipero]

Joker productions ?

https://www.youtube.com/watch?v=nsDjx-tW_WQ&t=1s

Interesting... Joker is running memory at 3000 MHz, that might help to explain the difference in gaming results. I noticed the "MEM" usage on the HUD is different between 1700 and 7700K, not sure how to interpret that.

Hmm... I wasn't able to find any kind of summary of benchmark numbers - all of the comments *about* these videos on other forums suggest that the numbers were pretty close, but there has to be a better way than watching the video and scribbling down numbers.

Yeah, Joker Productions, that's the name, I didn't even search for any specific, but by video it showed direct overlay of games and MSI Afterburner with FPS, mem usage, CPU usage, GPU usage and so on. From what i saw, it falls in line "similar IPC with different frequencies", same is produced by Guru3D review with CPUz benchmark, so dunno.

[indepe]

...
One video I watched shows performance of [email protected] vs. 7700k@5GHz with 720p resolution and low graphic preser (to eliminate GPU bothleneck), and it seems that Ryzen falls exactly where it should be, at 1.1GHz disadvantage,
...

The overclocking potential seems to be one of the many open questions. Some appear to think it is limited to +100 MHz (maybe that's just non-X), while elsewhere one reads the 1800X has been overclocked to 5.8 GHz (cooled with liquid nitrogen).

[indepe]

The performance numbers I'm seeing do smell a bit like they are memory bandwidth related, but some of that may just be that the BIOSes out there don't have all the fine-grained timing values for different memory speeds yet. Some of the review systems seem to be running memory more slowly than I would have expected (eg Michael mentioned running at 2133) but from what little digging I have had time to do at least one of the cases seems to be the BIOS having timings for two speeds, one fairly slow and one just a bit too fast.

That would fit with what I'm seeing - that 2666 sticks sometimes do better than 3000's right now because the 2666's actually run at 2666 (or whatever the number is, 2666 isn't looking right when I type it) while the 3000's actually run at something lower because they can't run at 3000 yet and they don't have the in-between timings yet. It would be nice if I could find some benchmarks against a 6900K running dual channel rather than quad channel - my guess is that gaming performance would be a lot closer and faster RAM clocks would help to close the gap with quad channel.

Again, take with a big grain of salt. I don't have a system to play with yet (CPUs are in stock around here but not mobos, which is the opposite of what I expected) so all I can do right now is look for patterns across a bunch of reviews like everyone else.

G.SKILL has announced DDR4 memory modules designed specifically for Ryzen CPUs.

[mmstick]

The performance numbers I'm seeing do smell a bit like they are memory bandwidth related, but some of that may just be that the BIOSes out there don't have all the fine-grained timing values for different memory speeds yet. Some of the review systems seem to be running memory more slowly than I would have expected (eg Michael mentioned running at 2133) but from what little digging I have had time to do at least one of the cases seems to be the BIOS having timings for two speeds, one fairly slow and one just a bit too fast.

That would fit with what I'm seeing - that 2666 sticks sometimes do better than 3000's right now because the 2666's actually run at 2666 (or whatever the number is, 2666 isn't looking right when I type it) while the 3000's actually run at something lower because they can't run at 3000 yet and they don't have the in-between timings yet. It would be nice if I could find some benchmarks against a 6900K running dual channel rather than quad channel - my guess is that gaming performance would be a lot closer and faster RAM clocks would help to close the gap with quad channel.

Again, take with a big grain of salt. I don't have a system to play with yet (CPUs are in stock around here but not mobos, which is the opposite of what I expected) so all I can do right now is look for patterns across a bunch of reviews like everyone else.

It's been discovered that Ryzen's equivalent of a Northbridge, Data Fabric, is designed to run at 1/2 the speed of the memory, so the higher the frequency of the memory, the higher the NB frequency, which directly results in performance improvements for the CPU.