AMD - Zen chitchat

[scaryjim]

I think something along that line as a while ago I went to see how expensive dual Xeon is while messing around with uk.pcpartpicker.com and was a bit confused with the options even when only looking at Broadwell-E parts.

Yeah, I just went and poked ark and it's not very clear at all.

A couple of google searches did turn up both single-socket and dual-socket S2011 v3 boards for workstations using the C612 chipset, as well as single-socket X99 boards. So Intel give you options, at least!

Although while I was pondering this another thought occurred to me - in all of the information we've got about Naples so far, we've not heard anything about storage/USB capacity. It left me wondering if Zeppelin's "southbridge" is actually just a cluster of PCIe lanes? I know it's already possible for PCIe connections to flexibly negotiate both SATA and USB traffic instead (this is what SATA Express and many mini-PCIe connectors do in existing motherboards), so I think that means it's possible that the southbridge is actually just a collection of general purpose PCIe ports, some of which can also be used for SATA/USB?

If that's the case, it could be that using the advertised 128 lanes of PCIe express on Naples (and note I've seen them described as 128 lanes of I/O in some interviews) would mean the SoC wouldn't provide any storage or USB...

[Corky34]

Not really related to much but as this is Zen chitchat i thought Anandtech's list of caveats seems to sum up all the problems with Ryzen nicely.

• Windows 10 RTC disliking 0.25x multipliers, causing timing issues,
• Software not reading L3 properly (thinking each core has 8MB of L3, rather than 2MB/core),
• Latency within a CCX being regular, but across CCX boundaries having limited bandwidth,
• Static partitioning methods being used shows performance gains when SMT is disabled,
• Ryzen showing performance gains with faster memory, more so than expected,
• Gaming Performance, particularly towards 240 Hz gaming, is being questioned,
• Microsoft’s scheduler not understanding the different CCX core-to-core latencies,
• Windows not scheduling threads within a CCX before moving onto the next CCX,
• Some motherboards having difficulty with DRAM compatibility,
• Performance related EFIs being highly regularly near weekly since two weeks before launch.

Most of the caveats seem to cover what was already known or suspected but i wasn't aware of Windows 10 RTC disliking 0.25x multipliers or the 240 Hz gaming issues.

[scaryjim]

... Anandtech's list of caveats ...

I'm not convinced they're right about the Windows scheduler issues they've listed, and they've missed the important one - the Windows scheduler treats all 16 threads as equal whereas the physical and virtual threads have very different performance.

Plus I have no idea what "towards 240Hz gaming" actually means. My guess would be that they mean situations where you deliberately underload the GPU using low resolutions and IQ settings to acheive high framerates, but frankly that's far more likely to be a symptom of the other performance issues than a specific issue of its own...

[Corky34]

This article I've just read seems to explain what's going on and why, while not the cause, the Windows scheduler doesn't seem to be helping, it seems like AMD have designed the equivalent of a two socket system on a single substrate but Windows is treating it as a single chip.

[CAT-THE-FIFTH]

It looks like the other articles talking about FO4 performance being at IB levels or a bit less were correct!

I still hold out hope,when motherboards have better RAM support,performance will be better,but at least I certainly know Ryzen 2 will be better than what I have in this scenario!

[scaryjim]

... it seems like AMD have designed the equivalent of a two socket system on a single substrate but Windows is treating it as a single chip.

I know, I've posted about that design choice several times including ways in which AMD could possibly address some of the issues with it (including admitting that it's only really an 8MB L3 cache size and improving the latency between CCXes).

However that's only part of the story, and another factor is that Windows will schedule equally between the physical cores and virtual threads. The virtual threads are apparently several times slower than the physical cores, so in lightly threaded loads Windows should schedule all threads on the physical cores, but will actually schedule them equally between physical and virtual, reducing performance.

CAT's linked some charts around here somewhere of the effect on game performance of turning various cores on and off and SMT on and off and no one factor is causing all of the variation. AMD need to address a number of issues. The main one they can sort out with the Windows scheduler, though, is going to be preferring the physical cores over the virtual threads.

[DanceswithUnix]

It looks like the other articles talking about FO4 performance being at IB levels or a bit less were correct!

Odd, the Ryzen seems to always have one core pegged at 70 to 90%, then Intel chips are way more even.

That 3570k is really maxxed out, giving it's all

[CAT-THE-FIFTH]

Odd, the Ryzen seems to always have one core pegged at 70 to 90%, then Intel chips are way more even.

That 3570k is really maxxed out, giving it's all

Its seen in the sweclockers review,and someone on a Czech forum,tried disabling SMT,running one CCX and windows 7 and it was similar.

What I find the issue now,is in my current play through I actually started building larger settlements,started using the build elements from the contraptions DLC(logic gates and factories) and started using mods,and I find one core of the Xeon E3 1230 V2/Core i7 3770 pegged at 70% to 85% usage with the other cores seemingly less pushed,when I start to see slowdowns. I can easily hit 25 to 30FPS at qHD using a GTX1080 with the GPU at like 60% usage.

To put it in context - The Witcher 3 at qHD including Hairworks to the max and maximum tessellation enabled does not go that low.

If you look at the full video,see how the slower CPUs can hit under 50FPS at 1080p with an overclocked GTX1070.

The framerates over 60FPS are pointless since the engine links physics with framerates,so you will end up with things going wonky anyway and they are only good to see the relative positions of CPUs.

But all of them show one thing - SKL/KL absolutely destroying earlier Intel CPUs,often by anything from 30% to 60% and from what I gather bandwidth actually helps improve performance,since the upper end of the scale seems to be mostly more down to the CPUs running higher speed RAM more than the CPU overclock.

Its why I am thinking once we can get better motherboards which can use higher speed RAM for Ryzen,I think it might be possible for it to get some decent gains. Most reviews had issues running RAM at anywhere near the 3000MHZ which the kits AMD provided were rated for.

Edit!!

The Far Harbor DLC is even worse - especially due to the effects and some of the NPC spawn areas.

I kind of avoided building anything large in Far Harbor!!

[kompukare]

The HT4U.net review is finally out:
https://www.ht4u.net/reviews/2017/am...1800x_im_test/ (Google Translate)

Massive review, haven't read it yet but they've got the R5 1600X in there which is interesting. Results are just between the i7-4790K adn i7-6700K (for apps), just above the i5-6600K for games.

[CAT-THE-FIFTH]

The HT4U.net review is finally out:
https://www.ht4u.net/reviews/2017/am...1800x_im_test/ (Google Translate)

Massive review, haven't read it yet but they've got the R5 1600X in there which is interesting. Results are just between the i7-4790K adn i7-6700K (for apps), just above the i5-6600K for games.

They seem to have simulated it using the R7 1800X.

[kompukare]

They seem to have simulated it using the R7 1800X.

Yes, saw that now.
Their direct comparison tool is interesting. Seems by default they ran the 1800X with DDR4-2133 memory, but their database tool also has the results for different speeds: DDR4-2400, DDR4-2666, DDR4-3200.
The stock vs DDR4-3200 results are interesting:
Lowest application difference is +1.3% (Microsoft Word)
Highest application difference is +19.6% (7Zip - no surprise there!)

Lowest game difference is +5.9% (Assassins Creed: Syndicate)
Highest game difference is +18.2% (Rise of the Tomb Raider)
(Mind you, even with 3200 memory, RotTR is still only as fast as the i7-3960K/i7-4770K, but then stock memory speed is about the same as i5-3570K.)

Seems for the memory tests they didn't only ran games at 1080P (without AA) though. Another strange thing: their simulated Ryzen 5 1600X was given DDR-2400.

Also fun with the compare tool is comparing Ryzen to FX8370.

[Phage]

Hmmmm so I'm not seeing the case for mving to Ryzen yet then. Id be happy with any gaming gain over my 4770k, but it isn't there yet.

[CAT-THE-FIFTH]

Yes, saw that now.
Their direct comparison tool is interesting. Seems by default they ran the 1800X with DDR4-2133 memory, but their database tool also has the results for different speeds: DDR4-2400, DDR4-2666, DDR4-3200.
The stock vs DDR4-3200 results are interesting:
Lowest application difference is +1.3% (Microsoft Word)
Highest application difference is +19.6% (7Zip - no surprise there!)

Lowest game difference is +5.9% (Assassins Creed: Syndicate)
Highest game difference is +18.2% (Rise of the Tomb Raider)
(Mind you, even with 3200 memory, RotTR is still only as fast as the i7-3960K/i7-4770K, but then stock memory speed is about the same as i5-3570K.)

Seems for the memory tests they didn't only ran games at 1080P (without AA) though. Another strange thing: their simulated Ryzen 5 1600X was given DDR-2400.

Also fun with the compare tool is comparing Ryzen to FX8370.

I think they used 2400MHZ for the R5 since its a lower end part and I suspect the lower end motherboards will have worse support for highspeed RAM??

If not it seems rather weird.

[kompukare]

I think they used 2400MHZ for the R5 since its a lower end part and I suspect the lower end motherboards will have worse support for highspeed RAM??

If not it seems rather weird.

Actually, it seems they had it running at 24000MHz due to an error:

Erst nach den Benchmarks fiel uns auf, dass wir hier mit DDR4-2400-Takt vermessen haben, daher die Abweichungen beim Speichertakt

That is, Only after [doing] the benchmarks did we notice that we had done them with DDR4-2400 hence the differences with memory clocks. Guess they weren't going to re-test again.

The other thing they say kept the same clocks as for the 1800X and just disabled two cores (and someone The Stilt's Strictly Technical thread that 4+2 is not possible and that 3+3 is the only valid hex core you can simulate in the BIOS). But the last rumours I saw of R5-1600X had it's clocks at 3.3/3.7 which is closer to 1700X.

Oh, anyone else tempted to automatically use R3, R5, and R7 rather than the mouthful Ryzen 7 1800X etc.? The alternative number-modelnumber like 7-1800X just doesn't work even with the hyphen.

[CAT-THE-FIFTH]

Actually, it seems they had it running at 24000MHz due to an error:

That is, Only after [doing] the benchmarks did we notice that we had done them with DDR4-2400 hence the differences with memory clocks. Guess they weren't going to re-test again.

The other thing they say kept the same clocks as for the 1800X and just disabled two cores (and someone The Stilt's Strictly Technical thread that 4+2 is not possible and that 3+3 is the only valid hex core you can simulate in the BIOS). But the last rumours I saw of R5-1600X had it's clocks at 3.3/3.7 which is closer to 1700X.

Oh, anyone else tempted to automatically use R3, R5, and R7 rather than the mouthful Ryzen 7 1800X etc.? The alternative number-modelnumber like 7-1800X just doesn't work even with the hyphen.

I think AMD should have called them R4,R6 and R8 because its one number higher than i3,i5 and i7!!

But I seriously hope the motherboards and some windows updates are in place. Even a 10% uplift would make the R5 1600X look much better in benchmarks.

Also AMD confirmed 3.6GHZ to 4.0GHZ clockspeeds for the R5 1600X.



What I would have liked to see is 6C/6T performance against an Intel 4C/4T part.

Outside the golden oldy engines,I expect anything newer a Core i5 will start to lose.

[scaryjim]

... anyone else tempted to automatically use R3, R5, and R7 rather than the mouthful Ryzen 7 ...

Yup, already have been doing Problem is we're only a year past when AMD R7 was a lower mid-range GPU - now it's a flagship CPU...