Comments Locked

48 Comments

Back to Article

  • Mookid - Thursday, September 26, 2019 - link

    I’ve been satisfied with Intel cpus for ever, but this might be a lot better upgrade than i9 9900(s).
  • jgraham11 - Thursday, September 26, 2019 - link

    Yeah me too up until recently! First was the patch for Intel's Meltdown bug (Huge slow down), then patches for spectre which Intel CPUs are vulnerable to all of them. Then patches for the Zombieload bug (an Intel exclusive). According to some benchmarks, my Intel sandy bridge is running at half the speed it was. I bought my first AMD Ryzen system last year, no CPU bugs and no regrets!

    I read on Anandtech a little while ago that most big server companies are shutting off Hypertheading because that's where the bugs are from, so you only get half the performance you paid for!
  • jgraham11 - Thursday, September 26, 2019 - link

    On yeah forgot Foreshadow, Intel processors are also vulnerable to that one. The best part of all these bugs is that Intel knew about these bugs when they released their 9th Gen Skylake refresh refresh processors, their 8th Gen Skylake refresh, and the original Skylake processors were continued to be sold with a known massive security vulnerabilities in them. Great company eh!
  • Reflex - Thursday, September 26, 2019 - link

    So you are hosting a large virtualized SQL server? Can't see any other way you are seeing perf hits like that. For regular desktop use and gaming, the perf hit of the patches ranges from 0-5%, and the 'half speed' you claim would require some very specialized usage profiles that are not in the realm of typical users.

    Plenty of reasons to go AMD these days, don't have to make things up.
  • rahvin - Thursday, September 26, 2019 - link

    To get that 0-5% hit you have to disable most of the patches. Even with the later meltdown patches that reduced the impact you still see upwards of a 10-15% hit with all the patches enabled, and if you got the extra step and disable hyperthreading because of it's intrinsic vulnerabilities on Intel you can end up with a near 25% hit depending on model.

    For a normal desktop usage scenario it's assumed you don't need most of the spectre patches because you aren't running VM's or server software with untrusted users so the patches don't get enabled leaving you vulnerable. The sneaky thing is most desktop users on windows aren't aware the patches aren't being enabled, it's being decided for them by windows. For the most part desktop users don't need most of these patches to be relatively safe, but if you want to be for sure safe or you are running potentially dangerous software you need to enable them and the real performance impact for the older processors (7th gen and older) can be pretty significant, especially with all the patches.

    Though a 50% hit on Sandy or Icy Bridge is IMO out of the realm of possibility with the current patches. Early on the big hit on these processors was the meltdown patch, but once windows moved the patch to the method that the Linux Kernel used (retpoline) this impact was significantly reduced. I think the biggest you can get on those two processors with all the current patches is around 20-25% and IIRC that's with disabling hyperthreading as a precaution against future spectre variants.

    The scary thing with the spectre variants is they don't know how far this is going to go and some of the exploits are remote capable. As this type of attack was only proved capable a little while ago it's concievable that as more and more variants are discovered that "drive-by" exploits could end up in javascript on the web, and that's the scary part.
  • msroadkill612 - Friday, September 27, 2019 - link

    Some of us have lives outside of figuring what is slowing our computers. It happens at the best of times w/ bloat etc. Juggling which patches are pertinent etc. on top of that, complicates matters exponentially.

    I think the naysayer poster is viewing it from the perspective of a tightly run IT dept, rather than plain folks and smaller entities.

    We sure know they hit some workloads...discerning which affect us is time we would rather not waste, ta very much.
  • Reflex - Friday, September 27, 2019 - link

    Again, this is not true at all. Multiple sites have profiled the perf impact of these fixes. For typical user and gamer workloads, the impact even with all mitigations applied is negligible. That picture changes in certain other specialized workloads, such as virtualization and databases (especially SQL) the hit can be in some specific situations as high as 30%. Those situations are rare and not at all within a normal desktop user's workflow.

    While you are correct that Spectre is the gift that keeps on giving, that specific series of exploits is nearly identical across Intel and AMD, and has no easy mitigation for anyone (it also impacts non CPU hardware, btw).

    If you are going to keep claiming this, I request that you actually back it up with benchmarks from reputable reviewers.

    Source: Both websites and the fact that I was personally profiling this impact *before* the CVE was publicly released due to my job. But again, you don't have to take my word for it, here is an initial overview from this very site from the very first set of patches/mitigations which had not yet been performance optimized (it's improved drastically since): https://www.anandtech.com/show/12566/analyzing-mel...

    Yes, drops, but very would be user detectable, and the worst one (storage performance) was long ago fixed.
  • Reflex - Friday, September 27, 2019 - link

    And just because I happened to pull up a search on the topic, here is a more recent test, focusing on server workloads: https://www.extremetech.com/computing/291649-intel...

    As can be seen there, worst case with HT disabled and all mitigations applied, sees a 25% hit in very specific workloads. On average a 15-16% hit is taken across the tasks. This is a big deal, but again these are servers, consumer workloads aren't seeing hits like this or even close.
  • Reflex - Friday, September 27, 2019 - link

    And just because I happened to pull up a search on the topic, here is a more recent test, focusing on server workloads: https://www.extremetech.com/computing/291649-intel...

    As can be seen there, worst case with HT disabled and all mitigations applied, sees a 25% hit in very specific workloads. On average a 15-16% hit is taken across the tasks. This is a big deal, but again these are servers, consumer workloads aren't seeing hits like this or even close.
  • Alexvrb - Saturday, September 28, 2019 - link

    The benches are performed on systems that may or may not have every available mitigation actually applied. I highly doubt any of these testers verified their test system wasn't actually vulnerable to the latest side channel attacks. They just install the latest OS updates and say "yep, it's patchz0redz". Intel just sends the microcode update out to vendors and let them decide what to do with it... for a consumer platform it's not "necessary" to enable some of them by default. Heck, they may not even do THAT much any more for Sandy.

    Recent Windows benches are also performed with both Retpoline AND Import Optimization now enabled by default, which helps tremendously regardless of what mitigations have been applied.
  • Alexvrb - Saturday, September 28, 2019 - link

    With that being said, the performance hit varies a great deal by load. Typical desktop usage even worst case is probably 5-10% on average. Less for newer architectures.

    However, storage benchmarks are going to show a different story. Any heavy storage workload suffers quite a bit - don't measure FPS, measure loading time on a game that is notorious for disk thrashing. First loading up a No Man's Sky save when you last parked yourself in a complex base, for example.
  • Reflex - Saturday, September 28, 2019 - link

    The storage story has been improving as well, its a moving target and for typical users its negligible.

    To be clear, I am fine with talking about these issues. My issue comes with exaggeration. That helps nobody at all aside from fanboys trying to make their holy wars seem more valid. Nobody at home is seeing a 50% perf drop, even on an old architecture with all mitigations. No home users is seeing even a 20% drop for any typical task.

    The real issue from my perspective is less about hypothetical perf drops and more about how these issues will impact security down the road, as well as what other devices have unknown side channel attacks. I know it has been demonstrated on some network interfaces, for instance.
  • MDD1963 - Friday, September 27, 2019 - link

    We don't need to experience the performance hits to simply quote (spew) the 'sky is falling' statistics with all the worst case scenarios we read on 'the internet', but, instead simply treat them as gospel, even for still 'hypothetical' exploits. To quote Dick Jones..."Who cares if it works!"
  • rahvin - Friday, September 27, 2019 - link

    Timing attacks (disovered in 2010 IIRC) were only hypothetical until Spectre was demonstrated. There is working exploit code for each Spectre Variants, that's demonstrated not hypothetical.

    I don't know if these exploits have been found being used in the wild yet, but with demonstrated exploit code it's simply a matter of time until these exploits make it into the active exploits.
  • svan1971 - Thursday, September 26, 2019 - link

    Great, I hear the 3900x will be available also, only 3 months after it was released.
  • Gothmoth - Thursday, September 26, 2019 - link

    better late than intel.

    i hear their 10nm process is finally coming into gear after what 3-4 years?

    what is 3 month compared to that. and amazon has the 3900x here.. onyl 12 but it is available.
  • AshlayW - Thursday, September 26, 2019 - link

    +1 you nailed it with this comment.
  • Slash3 - Thursday, September 26, 2019 - link

    Unfortunately Amazon in the US hasn't had the 3900X in stock (vs preorder) since August 16th.

    https://www.nowinstock.net/full_historydetails/112...
  • Korguz - Thursday, September 26, 2019 - link

    a local store i go to, has 20 3900x's in stock :-) like i tried to explain to some one else who said the 3900X isnt available, there ARE other stores OTHER then new egg and amazon.
  • rahvin - Thursday, September 26, 2019 - link

    That's not true, I saw it available for an hour or so a couple weeks ago, around the first week of September. That website's not doing a very good job monitoring. It comes in stock every so often and then is sold out again within an hours of being made available at the MSRP.

    I think anytime is becomes available bots buy up the stock and relist it at higher prices.
  • msroadkill612 - Friday, September 27, 2019 - link

    It is disgraceful. You could have bought Intel's 7nm $500 PCIE 4 12 core 3 months ago and been so much better off.
  • msroadkill612 - Friday, September 27, 2019 - link

    Seriously tho - AMD has not lost sales over the delay - in the sense that intel has no substitute.
  • intelati - Thursday, September 26, 2019 - link

    This is my dream...
  • mjz_5 - Thursday, September 26, 2019 - link

    Had no issue purchasing my 3900x up here in Canada
  • AshlayW - Thursday, September 26, 2019 - link

    I think AMD is just showing off at this point. This could very well be replacing my 3700X in my desktop, (not that the 3700X isn't adequate, but having 24 Threads in a 65W envelope at the height of Zen2's efficiency sweetspot would be just too tempting). I wonder if it would maintain similar boost clock speeds to 3700X in lightly threaded workloads (I.e gaming), but only drop to lower speeds when all threads are loaded up. Either way, the performance per watt will no doubt be higher than the already impressive 3700X, as processors tend to be more efficienct with wider and slower approaches. Very much welcomed for my overnight video render projects and WCG.
  • Reflex - Thursday, September 26, 2019 - link

    I keep saying I won't upgrade my 2700X SFF build until a new socket is announced, but damn if AMD isn't making it increasingly tempting with SKUs like this one.
  • hesido - Friday, September 27, 2019 - link

    AMD keeping the socket and motherboard compatibility is a great thing!
  • msroadkill612 - Friday, September 27, 2019 - link

    You could say they have added a lot of fun for enthusiasts - a kick ass power upgrade & onsell is as doable as a treat in the same league as a decent dinner for two.

    A whimsical facet of it imo, is it works as a pyramid scheme - each onsell situation creates a salesman for another am4 platform to join the AMD ecosystem.
  • haukionkannel - Thursday, September 26, 2019 - link

    Yeah... very sensible product. It does not try to get those high boos frequences when Ryzen sweet spot seems to be 4.2GHz, so this is almost as fast as 3900x (like 3600 vs 3600x) so it will offer much better value than 3900x that very rarely meet those very high speeds.
    We just have to hope that we Also get 3950 that have those 3.1 to 4.2 speeds. It Would be much better that upcoming 3950x cpus!
  • AshlayW - Thursday, September 26, 2019 - link

    A 16-core with 65W TDP would likely be the pinnacle of the AM4 Zen2 performance per watt overall. I'd dig that chip. Theoretically should be able to still do 4 GHz+ on =< 8 core loads, just like the 3700X does, so gaming won't suffer a lot.
  • rocky12345 - Thursday, September 26, 2019 - link

    Great to see more choices on the CPU front but it is just my opinion that if these are actually gonna be released they are probably 3900x CPU's that did not make the cut and instead of throwing them out AMD is making use of them at a lower performance point. It could also be a CPU made for OEM's that want to build prebuilts and because it is at a lower wattage they will get away with using slightly lower spec coolers. Just an opinion nothing more.
  • supdawgwtfd - Thursday, September 26, 2019 - link

    So the same as every single other CPU ever made in all time that isn't the top spec CPU then?

    Your comment is pretty worthless lol.
  • Korguz - Friday, September 27, 2019 - link

    rocky12345, its something called die harvesting, which amd and intel have been doing for years now.
  • Threska - Thursday, September 26, 2019 - link

    "Meanwhile, the Ryzen 9 Pro 3900 adds various security technologies, enterprise management, and reliability enhancements for business and corporate computers."

    Platform Security Processor. I can already hear the fear-mongers rattling their cage.
  • imaheadcase - Thursday, September 26, 2019 - link

    huh? Its business security stuff. Not sure details but normally means a hardware option vs software needing to use something.
  • close - Thursday, September 26, 2019 - link

    It's the PSP, AMD's equivalent of the ubiquitous Intel ME. A controller that can run stuff in ring 03 (higher privileges than the SMM/CPU itself so to speak). This has a black box firmware so it's opaque enough that it's hard not to think that it can be very easily exploited for nefarious reasons.

    Of course, besides fear mongers there are people like Threska who despite several very eye opening hacking and spying revelations still decide to pretend like all is good in the world and nobody would try to pull a fast one on them.

    @Threska, since you obviously live by marketing statements I have a bridge to sell to you.
  • bananaforscale - Thursday, September 26, 2019 - link

    PSP has been a thing for a few years now.
  • MattZN - Thursday, September 26, 2019 - link

    Just take your 3900X, go into the BIOS. AMD CBS -> NBIO -> SMU, and set the cpu socket power to whatever you like (within your mobo and cooling capabilities). Its supposed to be watts but it doesn't seem to be very accurate, so try '80' first and see how it goes. Its fun to play with.

    Presumably the '3900' will simply allow AMD to use lower-binned dies that don't turbo as well.

    -Matt
  • Dug - Thursday, September 26, 2019 - link

    I'd be curious if you can get same with 3900x as this 3900. If true, then 3900x seems like a better buy as you can adjust when you want. I wouldn't expect the 3900 price to be that much different than 3900x
  • MattZN - Thursday, September 26, 2019 - link

    Ok, I have a few minutes and a 3900X that's idle. Lets find out.... socket power set to '65' under CBS -> NBIO -> SMU. This is not running windows so my workload is basically a bulk compile... ahhh, lets see. I'll compile chromium. This is obviously a non-AVX workload so YMMV.

    On this workload under full load I get 3.0-3.2 GHz all-cores. Most of the time its sitting at 3.1GHz or so. Power consumption at the wall is 118W. Machine configuration is: 64GB of ECC ram @ 2400, a little radeon 6450 for the screen (just in VGA mode), tower cooler (43C), and three SSDs.

    Normally I run this machine with SMU set to 105W, which is around 165W at the wall. Its in a rack so everything is on a power budget. p.s. 3900X's are real monsters when it comes to doing bulk compiles and things like that. Just unbelievable horsepower and power efficiency for the price. It's only limited by having 4 unbuffered memory slots (64GB w/16G DIMMs or 128GB w/32G DIMMs).

    -Matt
  • yannigr2 - Thursday, September 26, 2019 - link

    Thanks for the info Matt
  • Slash3 - Thursday, September 26, 2019 - link

    Not bad! I'd love to see a site do an in depth test of power curve efficiency on these chips (I may have missed one if it exists, it's been a hectic summer).
  • Spunjji - Friday, September 27, 2019 - link

    Tidy! It'll be interesting to see whether they specify a tighter voltage curve to hit slightly higher speeds on the 3900 at that TDP, or just hard-code the lower TDP and let the CPU sort itself out.
  • NICOXIS - Thursday, September 26, 2019 - link

    449?
  • bananaforscale - Thursday, September 26, 2019 - link

    Could be cheaper as it won't have to bin as well.
  • yannigr2 - Thursday, September 26, 2019 - link

    I think this model and a 3950 non X could help with the lower availability of 3900X and the delay of 3950X. Many want those models for their number of cores/threads, not specifically for their boost speeds and they would be probably more than happy with a somewhat lower speed, lower wattage model that will be available now.
  • Danvelopment - Friday, September 27, 2019 - link

    I haven't owned an AMD since my Phenom II X2 but looking forward to setting up my new server with an R5 when I'm close to release. Current server is an E3-1275 v3 (just switched my main workstation from an E5-2670 to an E3-1230 v5).

    It's pretty cool that I can buy an R5 now, then upgrade to a 16 core later without having to redo the whole machine when my project releases.

    If I ever need more than that, money will no longer be an object.
  • peevee - Friday, September 27, 2019 - link

    Oh Dog. This is just stupid. I'd rather have 3800x.

Log in

Don't have an account? Sign up now