Amd Ryzen Threadripper & X399 MegaThread! FAQ & Resources

Page 15 - Seeking answers? Join the Tom's Hardware community: where nearly two million members share solutions and discuss the latest tech.
Status
Not open for further replies.

goldstone77

Distinguished
Aug 22, 2012
2,245
14
19,965
Looking at all the linux benchmarks Intel CPU's, 6 test vs. 1, get hit more often, and at higher margins than AMD. Intel is clearly taking the majority of the impact from patching in comparison.
embed.php

Here is a complete list of linux 4.14 benchmarks here.
https://www.phoronix.com/scan.php?page=article&item=linux-retpoline-benchmarks&num=5
Also, note that version 4.15 is supposed to be released Sunday. And this will be a continuous updating process for both Intel and AMD that will not end until they develop new secure CPU designs.
 

juanrga

Distinguished
BANNED
Mar 19, 2013
5,278
0
17,790


It was a way of saying "forget the myth about AMD performance not being affected".
 
Plus, from what was said by the author of the i3-8100, the patches out there now are hotfixes and the real, complete patches are still in the works. So all the numbers we see now are likely not the final result.

I didn't believe that AMD was totally unaffected, just didn't seem plausible. However, AMD says that they are not affected by one of the vulnerabilities as it exploits one of Intel's design flaws.
 

goldstone77

Distinguished
Aug 22, 2012
2,245
14
19,965

James Prior
‏@cavemanjim
15h15 hours ago
More
Replying to @IanCutress @ProperSystems
Technically you can have upto 32 pcie devices attached to a TR system, but you need pcie clock buffers after you get to 9 devices. Bios needs to support >7 gpu device enumeration, ask the mobo manufacturer for support
https://twitter.com/cavemanjim/status/1002360587349786624
Good to know!
 

genz

Distinguished


According to Linus (and I'll take the mastermind behind Linux and Mac* over any Intel/AMD Rep, and let the rep convince him). the patches are very biased in the way they work, and written to be as debilitating as possible to all parties by Intel. Makes sense as far as driving sales if anything else.

Actually, they apply patches liberally to things they shouldn't, including patching AMD chips for bugs they don't have. Please understand that a bug fix for a security issue on this level is not a bug fix, it's turning off the part of your CPU with the bug and doing it's work in emulation. Intel (who are issuing Spectre/Meltdown bugfixes period) is turning off bits of AMD chips for no reason.

I can tell you now that I know that AMD could not have been affected by the same side channel attacks, because Spectre's SC method relies on prefech/pred 'guessing' sum results or memory addresses and prefetching them before they are 'needed'. Intel was miles ahead of AMD at pred prefetch in the day of the P4 and still is now. The P4 was actually reliant on it because it's pipeline was so long that it would cost several times longer than a K6 to flush buffers after a cache miss. Obv Intel planned for P5 or better steppings to reach the clocks that the core was built to make.

All the way to Sandy, AMD was using bascally the same prefech from Athlon XP/Sempron.

What this does do however, is give Intel the perfect opportunity to make all those people still enjoying Sandy Bridge and 8 core Xeon X's a massive speed bump that might make them upgrade. It's not like CPU's are getting slow every weekend like they used to.

(*because Mac runs on BSD Unix and that's foundationally a Linuxified Unix)

There is no long term bug fix because you have to change the design of the processor. If you accidentally make one part of the Cpu wrong, there is no spare part to work around. Everything else has something to do or it wouldn't be there, and obviously the issue is in the metal and switches not the 1s and 0s. That line was Intel saying that basically Coffee Lake is still as vulnerable as any before as it was not redesigned to eliminate the bug.

Looking at just how long the bug apparently existed, it's possible that the first CPU that Intel design without this bug will be slower than the last CPU with that bug. It has to be at the core of the CPU logic to go through probably 100s of releases, steppings and iterations without being spotted or tripped over, or simply disappear with design updates.

EDIT: Read back and I seem to be making statements that are a little unbelievable so I will update below with links to where this info is from. Italics is Linus verbatim words.

https://techcrunch.com/2018/01/22/linus-torvalds-declares-intel-fix-for-meltdown-spectre-complete-and-utter-garbage/

Meanwhile, a bunch of other things are added in the same patch that Torvalds points out are redundant with existing solutions, for instance adding [performance harming] protections against an exploit already mitigated by Google Project Zero’s “retpoline” technique.

Why do this? Torvalds speculates that a major part of Intel’s technique, in this case “Indirect Branch Restricted Speculation” or IBRS, is so inefficient that to roll it out universally would result in widespread performance hits. So instead, it made the main Meltdown fix optional and added the redundant stuff to make the patch look more comprehensive.

Is Intel really planning on making this shit architectural? Has anybody talked to them and told them they are f*cking insane?

"They do literally insane things. They do things that do not make sense. That makes all your [i.e. Woodhouse’s] arguments questionable and suspicious. The patches do things that are not sane.

…So somebody isn’t telling the truth here. Somebody is pushing complete garbage for unclear reasons. Sorry for having to point that out"


Woodhouse (who in a long-suffering manner asks they “be done with the shouty part”), later in the thread acknowledges Torvalds’ criticism, calling IBRS is “a vile hack” and agreeing that “There’s no good reason for it to be opt-in [which means that OS must choose to turn it on at boot, and current implementations do not check if the CPU needs the patch or not, and this was written entirely by Intel including that bit].” But he but notes some points that are, if not exactly in favor of Intel’s approach, at least explain it a bit.

Intel, for its part, offered the following statement: “We take the feedback of industry partners seriously. We are actively engaging with the Linux community, including Linus, as we seek to work together on solutions.”
 

goldstone77

Distinguished
Aug 22, 2012
2,245
14
19,965
MSI MEG X399 Creation Flagship 2nd Gen Ryzen Threadripper Motherboard Previewed – 19 Phase PWM Design
By Hassan Mujtaba
1 min ago

MSI-X399-MEG-Creation-Motherboard_3.jpg

"These heatsinks are interconnected via heat pipe which runs from the top VRM heatsink to the heatsink alongside the I/O. The rest of the 3 VRMs are dedicated to the SOC. Total of 19 DrMOS phases is found on the motherboard which delivers up to 250Ws of power to the CPU alone."
https://wccftech.com/msi-meg-x399-creation-motherboard-2nd-gen-ryzen-threadripper-detailed/
 

genz

Distinguished
250W TDP is mental. They must have made a lot of headway with the spreader.

A 250W chip is much more doable on an MCM than a single chip in theory, due to the core clusters being physically further from each other and therefore the heat being spread across the heatsink better. Not saying that overclockers aren't already pulling 250W from some X299s, but the fact that this chip will probably run a lot cooler at the same power draw because it will be more effectively sinked.
 


Considering that it's 250w for 32 cores as opposed to 220w for an 8 core FX 9590...They have come a long way in efficiency and performance. 250w isn't that bad for such a massive chip, though cooling will likely take a good water setup.
 


Well lets remember that the fx 9590 is an overclocked FX 8350. Those piledriver cores were operating way out of their most optimal performance to efficiency range at 5ghz.

But still, either way the performance out of a 250W chip today is mind blowing vs what a 250W chip (whether more efficient or not than the FX 9590) in 2011 was. I believe the best comparison would be 4 2011 octo core Xeons in a quad socket server board.
 

goldstone77

Distinguished
Aug 22, 2012
2,245
14
19,965
Ryzen Threadripper 2 (2990WX and 2950X) Review: AMD Unleashes 32 Cores
by Paul Alcorn August 13, 2018 at 6:00 AM

It Starts With 12nm LP
AMD's Threadripper 2 processors are manufactured on GlobalFoundries' 12nm LP process technology. The ported-over design helps boost transistor performance, but does not affect die area or transistor density. As a result, the Zeppelin die's ~4.8 billion transistors and 213mm2 area remain similar from first-gen Ryzen. The dual-die X-series models feature a total of 9.6 billion transistors and 426mm2 of silicon, while the quad-die WX processors feature 19.2 billion transistors over 852mm2.

Lower leakage current does enable 200 MHz-higher clock rates or an 80-120mV core voltage reduction at any given frequency compared to 14nm manufacturing. All told, AMD claims the 12nm design enables up to 11% less power consumption than 14nm-based Threadripper CPUs at the same clock rates, or up to 16% more performance at the same thermal design power. AMD also adds other nuanced refinements, like lower L1 (15%), L2 (9%), and L3 (8%) cache latencies, along with reduced memory latency (2%).
https://www.tomshardware.com/reviews/amd-ryzen-threadripper-2-2990wx-2950x,5725.html
Tons of information to sort through!
 

goldstone77

Distinguished
Aug 22, 2012
2,245
14
19,965
AMD Threadripper 2990WX Linux Benchmarks: The 32-Core / 64-Thread Beast
Written by Michael Larabel in Processors on 13 August 2018. Page 1 of 11

Overall, the AMD Threadripper 2990WX is a terrific offering for those engaging in heavily threaded workloads whether it be for a software development workstation with frequent code compilation, CPU-based rendering in Blender or other professional applications, carrying out OpenMP/MPI-threaded scientific workloads, or a range of other multi-threaded workloads or a lot of multi-tasking in programs like GIMP, Darktable, and other workstation applications.

While the Threadripper 2990WX may seem like a lot -- and it is -- it's actually with great value as shown by our performance-per-dollar results. For 32-cores / 64-threads at $1799 USD, consumers are getting a lot of performance potential and right now it blows the Core i9 7980XE out of the water in not only raw performance and performance-per-dollar. Intel will definitely need to respond before potentially losing much market-share in the HEDT space. The Threadripper 2990WX has also proven to be very competitive in performance-per-Watt to round out this offering.
https://www.phoronix.com/scan.php?page=article&item=amd-linux-2990wx&num=1

The 2990WX can be seen winning around 40 test, I lost count!
 

goldstone77

Distinguished
Aug 22, 2012
2,245
14
19,965
AMD Announces Threadripper 2, Chiplets Aid Core Scaling
David SchorDesktop ProcessorsAugust 7, 2018 Tagged 12LP, 12nm, AMD, HEDT, Ryzen, Threadripper, Threadripper 2, x86, Zen+

Going purely by the introductory price, second-generation Threadripper is now $12.50/core cheaper for the 12-core parts and $6.25/core cheaper for the 16-core model.
tr2-tr1-cost-comp.png

tr2-2-die.png

tr2-4-die-768x640.png

https://fuse.wikichip.org/news/1569/amd-announces-threadripper-2-chiplets-aid-core-scaling/
 

jaymc

Distinguished
Dec 7, 2007
614
9
18,985
AMD's New Threadripper 2990WX Much Faster On Linux Than Windows 10:

Linux is kicking ass when compared Windows 10 Pro in identical benchmarking test's with Threadripper 2.

" The site also measured FLAC audio encoding (Linux won, but barely), H.264 video encoding (Linux won by a large margin) and a healthy assortment of other benchmarks. "

That's because Windows has poor support for Non-Uniform Memory Access setups. The benefits of NUMA, however, are restricted to certain workloads. In discussing the Phoronix results, NotebookCheck observes the following:

"In all Blender tests, the performance on Linux was clearly better by at least 15% compared to Windows 10. However, in single-threaded tests such as FLAC Audio Encoding and BLAKE2, the Threadripper 2990WX performed almost equally across Windows 10 Pro and Linux. FFmpeg also favored performance under Windows 10 compared to most other Linux distros. Therefore, these results seem to indicate a scalability issue of Windows 10 Pro across the many CPU cores in heavily multi-threaded workflows."

"I will obviously take this moment to point out that the selling point of Threadripper is its ability to chew up a workload with that extreme thread count. At this stage in the game, it's beyond clear that Linux extracts the best performance for those types of workloads. Probably not news to the Linux community, but it's an eye-opener for me."

Check it out here:
https://www.forbes.com/sites/jasonevangelho/2018/08/15/want-faster-performance-with-amd-threadripper-2-use-linux/#5f9efe7139c9
 


It makes sense Windows would suck ball a NUMA; unlike Linux, Windows will happily allow threads to jump between cores in order to keep them alive after they've been bumped. On a NUMA system, that will result in significant overhead due to increased memory latencies.

In theory, if thread(s) were locked to a specific core after creation, that downside would vanish. Windows doesn't have a great way to handle it currently though, and most developers won't manually do this as you can easily tank performance on non-NUMA architectures.

So yeah, pro's and con's.
 
Status
Not open for further replies.