New on LowEndTalk? Please Register and read our Community Rules.
All new Registrations are manually reviewed and approved, so a short delay after registration may occur before your account becomes active.
All new Registrations are manually reviewed and approved, so a short delay after registration may occur before your account becomes active.
Comments
Intel is like Apple, its still expensive but they lost the edge, you can find a better and cheaper alternative, called AMD
FMV isn't very valuable in my field. Main reasons: I often have to do more than x86 (e.g. P9, div. arm) plus in most of my code what is but a couple of hot spot functions my code is those functions which leads (a) to quite significant size growth and (b) a big mess because I have to deal with different architectures too.
Also gcc often simply isn't an option because I have to use clang.
I often can't use C or C++
TL;DR I love AMD Zen but in my work it adds to the mess we already have.
Oh and of course the elephant in the room: You can compile per target but most OS code doesn't use FMV and much OS code does not even differentiate between multiple sub targets (like Haswell vs Nehalem) so it's basically down to setting compiler switches for a given system which helps a lot but can't even get close to having different function versions.(which is yet another reason why benchmarks (incl. mine) do not show the full capabilities of a given platform, especially not a new one, which again leads manufacturers to provide benchmark code which however of course is selected to make shine whatever they wish to make shine ...).
That is largely not true in the low end segment. Reason: Going AMD Zen requires you to buy new (read: way more expensive) nodes.
For end users/consumers though it's certainly attractive and usually cheaper too to go AMD Zen.
Ah that rules it out then.
Indeed - so choice of AMD or Intel is always be relative to ones own usage requirements and usage environments. But AMD EPYC/Ryzen usage and adoption can only improve and grow in the years to come
Though I find it strange that both Intel and AMD have record breaking demand for their products but still can't drum up the supply side to meet it ! If only other businesses had such a problem
One segment which Intel still has the lead over AMD is server class Xeon E3-12xx with relatively low TDP, high clock speed >3.5Ghz and lower prices. AMD EPYC server class don't have anything in that segment for high clock speed and competitive pricing yet. AMD EPYC 7371 gen1 https://www.servethehome.com/amd-epyc-7371-review-now-the-fastest-16-core-cpu/ was first cpu was close on high clock speed but pricing was much higher than Xeon E3-12xx at the time.
This of course is probably going to be based on what the user is using the processor for but I'm nearly certain for many (not all) of the cases, price is the biggest influencing factor.
Price wise, the latest AMD stuff is cheaper with similar if not higher consumer grade benchmarks (based on cpu benchmarks website).
For utilization of which I have limited knowledge of newer AMDs, I would assume the higher benchmark scores would help with my video post production / conversions before uploading to video platforms. This is why I would choose the newer AMDs (based on that assumption for encoding video).
Surprisingly a few forums I've did research on with the relatively latest AMDs like the 2200G, 2400G and even the newer 3xxxG's seem to have an issue doing encoding with Streamlabs (for live stuff like Twitch) and this has currently encouraged me to wait before getting them. Surprisingly Intel seemed to have more compatibility in area - in this case at this point, it's probably safer to go with Intel.
This is just some of what I meant in the first two lines of what I thought about the question. I see LET votes show a massive swing to AMD which I totally get.
I think price is another factor that is often misjudged.
Of course, for us consumers price is of high importance. In fact I know of quite some people who buy pretty much price-based only; they buy the best they can afford, simple as that.
For companies though the situation is different, mainly for two reasons (I know of; possibly there are more):
Just caught up with your excellent posts. I guess maybe AMDs might make better sense for smaller hosts rather than DCs currently? There is a certain unique selling proposition offering AMD VMs at current state (better performance and security at similar pricing) and smaller hosts can probably get a Ryzen dedicated to resell much more quickly.
Also, the dedicated market probably is not affected. Mainly it is the VPS market with the security vulnerabilities that may be of concern.
AMD is the superior potassium, it's Linux where Intel is Windows.
Have a look at one of the tables that list the Epycs. You'll see the laws of nature: core count goes up - power draw does too while clock rate gets down. Same at intel.
Funny true story: Cisco designed one detail that probably next to nobody would even notice really smartly. This detail allows for better airflow at a critical spot (the back end of a blade). Power savings? Roughly 5% ... and the engineers get crazy over it. The trick? Their design allows for fans to spin slower. FANS! The bloody most primitive part in a blade (or server).
Let me wake you up: Those funny little fans in all servers (dual socket, intel or AMD, no matter) typically draw 35 -50 W and up to 70 W. Each! There are usually 5 fans, so a servers fans draw something in the region of what some processors draw. Yet nobody I know of even looks at them when guesstimating the power draw of a system.
Lesson to be learned: Do not underestimate the simple little things and never ignore the power a system needs.
THAT is why people in DCs, providers, etc. often smilingly pay more for an 'L' (low power variant) processor that is slower than the normal one. DC translates to performance per rack - not per server and certainly not per core.
Next question: Who's the culprit, the big power gobbler? Clock speed. Many don't know that but performance does not rise linearly with clock speed. Which is why all, intel like AMD have lower clocks with higher cores. 4 or 6 cores? No problem, you can high clock speed. 16 cores? That's where it starts to get tight. And north of 20 cores there is but one way: lower the clock quite brutally.
The other point I want to address is that many seem to consider intel beaten and possibly soon going belly up. That however is very unlikely. intel has lots going for it, for example many fabs. intel can and do produce themselves while AMD needs a third party fab plus only very very few fabs on this planet can do 7 nm.
Let's have a look at quanta, a taiwanese giant who is the real manufacturer of many Dell, HP, etc. brand name servers. How many Epyc boards do they produce (albeit with slight OEM variations)? Two. 2 Epyc main boards. One with one socket and one with two sockets. That's it. Or look at Lenovo; all they have so far if I'm not mistaken is 2 Epyc systems.
"But I know that Gigabyte alone produces 7 (or whatever) main boards" I hear you say. Yes, they do - but those are consumer boards. Server boards are way more complex and must be way more reliable.
Short version: Epyc won't kill intel anytime soon. Ryzen might, but I guess that intel will fight hard to survive - and they have the resources. And btw, intel has barely some small scratches so far. They are by no means limping.
What IMO really happens is something quite different. What really happens is that intel - thanks to AMD - is forced to not anymore think and behave like a mix of king and thug. They were forced to stop "whatever I want" pricing, they are forced to care about customers and their needs instead of an "you'll eat whatever we please to serve" attitude, they can't afford anymore to simply ignore or lie away vulnerabilities because there was no competition and everyone had to play on intels game board.
So we indeed owe AMD a BIG FAT "Thank you!" and we have damn good reasons to buy those Ryzens but let's not act as if intel was all but dead and as if AMD only produced superstar processors.
Oh, and btw, 97 out 100 servers (loosely guessing) simply do not need more than 16 cores because we have only so many e.g. mid to high end database servers. Funny side note: good luck finding a 4 socket high availability Epyc server.
TL;DR Let's laud AMD, let's thank AMD, but let's also keep our feet on the ground and let's stay realistic.
Btw, I personally think that the real reason for intels problems are inside intel itself and less with Zen. They had to fire their long-time CEO, they cluster-f_cked up their 7 nm plans, they carelessly bought to many companies, etc. And IMO that is also the main reason why they look like paralyzed right now and try to push ridiculous new "super Xeons" that look pale.
One difference particular in AMD Ryzen 3000 and EPYC Rome 2nd gen is that their Turbo Boost frequencies on average end up holding that state for longer compared to Intel cpus. So at all core max load, AMD Ryzen 3000 and EPYC Rome 2nd gen will either have a higher all core cpu frequency or closes the gap with Intel's all core max Turbo frequencies. Lots of folks overlook that Intel's Turbo boost max cited frequencies are usually for 1 or 2 cores and drops as you load up the cpu cores in terms of work loads to max cpu usage.
For my Intel Xeon and AMD Ryzen testing I have also been logging and charting their cpu frequencies per thread for idle and under load tests some examples at https://community.centminmod.com/media/albums/cpu-frequency-charts.14/
Example with Intel Xeon E-2288G and AMD Ryzen 3900X
With performance governor enabled, under respective all core load, their max cpu frequency ended up close to each other - AMD Ryzen 3900X has 24 cpu threads turbo pegged at 4.0-4.1Ghz while Intel Xeon E-2288G had all 16 cpu threads turbo pegged at around 4.1Ghz under CentOS 7.
Some of the proper benchmarks out there that measure cpu frequency under load too also report the same AMD Ryzen and EPYC Rome 2nd gen under all core/thread loads on some cpu SKUs/models catch up to higher clocked Intel Xeon/Desktop parts because of under all cpu core/thread load, the AMD cpus narrow the gap with Intel's all-core cpu frequencies Turbo max.
Of course, you have to make sure the distro, kernel version and system settings are properly allowing AMD or Intel cpus to boost to their rated frequencies under full load.
Example AMD Ryzen 3900X with all 24 cpu threads peaked around average of 4.0Ghz during all-core cpu thread workloads and with less than all-cpu thread workload's frequency peaking around 4.5Ghz
+1 indeed, AMD is forcing Intel to do things it usually wouldn't do which is a good thing for all customers when competition exists.
True, though all it takes is a few more serious Intel specific cpu Spectre/Meltdown like vulnerabilities to pop up in next 3-12 months to put an even more sour taste in traditional Intel loving/favouring folks.
Maybe
Cheap, more cores, good thermal without overclock,
Best cheap cpu, ryzen 5 3600
Can't say about the servers, but using a laptop with ryzen 5 mobile in it and it works flawlessly. Great performance.
@eva2000
"clock", "gap" - Yes, but again: try to measure power draw too and you'll see what I was talking about. Every processor not only draws more power with increasing clock but that increasing is not linear in a quite ugly way. In other words: Beyond 2 to 2.5 GHz you get increasingly less performance per Watt which translates to sharply increasing el. power costs at the high speeds.
Also keep in mind that those "only x cores of all can turbo at some high speed" games are due to electronics/physics. One can only pump so much energy through a die (and one can carry away only so much heat from a processor).
Look at my hosting primer. Electrical power is a major cost factor and it has two sides, (a) the power itself, and (b) the heat produced - and the equation gets very ugly with processors driven beyond a certain point (somewhere around 2 - 2.5 GHz for many core Xeons and Epycs).
Beyond a certain level it simply makes sense and is cheaper to run 2 nodes at lower speed than putting those cores into 1 node and run it at high speed.
Next add the memory question. The difference between say 2600 MHz and 3000 MHz in both price and power consumption is significant plus, with Epyc, you want to have memory and processor at the same nominal speed.
All of this are reasons why I warned of being too impressed by phoronix, anandtech, etc. benchmarks. Those are largely focused on max/top speeds and if that means a need for more cooling, so what then use liquid cooling with a larger radiator, etc. - but things are quite different in a DC where systems usually are densely packed, el. power is a very significant cost factor and where node lifetime is a significant factor.
But again, clearly YES, AMD Epyc is closing the gap and their design highly likely allows for further improvement. Specifically it seems to me that Epyc allows one to get somewhat more performance out of a given el. power budget and/or more cores.
TL;DR I'm not contradicting you generally but only with a focus on the DC/hosting. And I can't be all wrong, just look at what AWS did: they designed their own processor and it's not high clock and it's based on Arm, so obviously their focus is to optimize along the cost/performance/power axis.
Thinking about Intel vs AMD, lately everyone is praising AMD, which AMD earned it.
However, I would imagine Intel would destroy AMD if they both are using the same feature size, i.e. 7nm
Well, I recently got a "new" Xeon node handed over for NanoKVM.
Frst run with Unixbench, gave me 5155.8 points.
After, microcode updates => 4491
After, newest kernel from Proxmox 6 => 4177
Premium migrations, nah don't touch intel.
can microcode updates be undone?
Yes and no. Yes, 7 nm allow for more performance or lower el. power consumption and hence for more cores within the same power envelope.
BUT: There is much, much more to processors and manufacturing them. I don't think that intel can and will destroy AMD as soon as they have their 7 nm process working. If intel ever again destroyed AMD then via market power, cost and price. I doubt though that that will happen.
Sorry, no. Yes, losing about 20% of performance due to mitigation is ugly, sure.
BUT: You and your clients/users will run an OS and plenty applications (PHP crap anyone? OpenSSL?), so everyone will stay vulnerable.
Plus all the processor vulnerabilities are just the ones so far known. You can bet on new ones coming up and AMD won't stay relatively clean. And again, even if they did stay clean, so what, that does not translate to VPS users not being vulnerable.
Plus, and that's an ugly one: AMD, just like intel, is condemned to x86 architecture conformance and to stay backward compatible. Trust me, that's a deep hole with plenty poisonous snakes inside.
The reason they decrease performance is because they're the security vulnerability fixes for the speculative execution vulnerabilities publicised in the last couple of years. If you want to disable them you can but on your head be it.
Obviously a VM is not 100% bulletproof.
Same goes for everything.
But that's the Intel/x86 tax you pay, to keep it secure enough.
Yeah I did end up measuring power draw via turbostat output for E-2288G and Ryzen 3900X at peak all core loads these were the numbers
Intel Xeon E-2288G 8C/16T
AMD Ryzen 3900X 12C/24T
whether turbostat is accurate in it's reporting not sure. I don't have physical access to these servers - only remote access. But clearly AMD's 7nm process is in play here.
Of course for hosting providers, only way they will know the true numbers is to test and evaluate servers themselves and measure their actual power draw.
The software level kernel mitigations definitely impact performance - especially on Intel side. For my benchmark comparisons for below cpus, I also did PHP 7.4 vs 7.3 vs 7.2 vs 7.1 tests similar to ones I did at https://community.centminmod.com/threads/php-benchmarks-7-4-vs-7-3-vs-7-2-vs-7-1-vs-7-0-php-fpm.18741/
Guess which of those cpus had a more dramatic drop in performance due to only having software mitigations as opposed to hardware level mitigations during benchmarks
Only way is to test cpus for your intended use to see really. But as already mentioned not all web hosts and DC users would prioritize performance to that extent anyway. To some extent they should as faster performance generally means more performance per watt in terms of efficiency - you can do more/process more within the same power envelope. To say using older Intel cpu with 'good enough' performance, maybe at a power draw that a more efficient cpu i.e. AMD might be able to do at same or lower power draw level too.
For such yes definitely don't look at the top end cpus from Intel or AMD - 200+ watt cpu probably isn't suited to some DC usage. Sweet spots for AMD EPYC Rome 2nd gen are EPYC 7302P and 7402P single socket with 155 and 180 watt respectively and dual socket AMD EPYC 7282 and 7302 with 120 and 155 watt respectively for best bang for the buck in terms of cpu frequency, core count (all are 16C/32T based except 7402P is 24C/48T based) and power usage.
other hosting providers should learn from this and make other useless threads like pepsi vs. coke and google vs. duckduck go and so on to spam their sigs. anyone with any technical knowledge knows that amd is good for some tasks, intel and arm for others. it all depends on the application. there is no one best cpu manufacturer.
Coincidentally, Phoronix have posted AMD EPYC gen1 vs gen2 power consumption benchmark numbers highlighting what I said above gen2 just does perform more per watt than previous generation https://www.phoronix.com/vr.php?view=28588 and sweet spot for performance per watt is AMD EPYC 7302
That's a good example for my reasons to not care about Phoronix and similar test numbers.
They show total system numbers rather than processor (power consumption) numbers. Their system is not at all comparable to a server system. Don't get me wrong, as a consumer I like Noctua coolers but to see some consumer power supply and a funny Noctua cooler is certainly not inducing confidence in this context. And then their power measurement system ... reading the power once per second is, pardon me, just ridiculous. Also note the 3 120 mm fans plus another 3 80 mm fans of unknown make, model, consumption, etc. and all that in a case that is much higher (4 HU) than 2 HU and has an airflow that is totally and grossly different from a servers airflow ... and then they laud the Gen2 Epycs for staying nicely cool.
And of course they don't tell us much about their testing. Example: "The CPU temperature was also being monitored" without telling us how - and that can be really important and make quite a difference.
That's just ridiculous.
And of bloody course they use the usual test suite that is virtually utterly worthless to judge a server.
Of course - those are serious people, right? - they also show us a pseudo screenshot of their compiler parameters (of whatever) with ridiculously incompetent parameters. One gets the feeling that they think "the more parameters and values the more professional it looks!" ...
Also note that the gen 1 Epycs are made to look like sh_t compared to the oh so wonderful new Gen2 Epycs. But then Gen2 is the current hype ...
Professional measuring and system evaluation is something quite different from what the Phoronix hype-boys have on offer.
It would for example measure power at the supply rails and with at least millisecond resolution. It would load and test a server board in a server with loads that are typical for a server that is, database, http and https both multithreading and event driven (async), etc., it would look at and tell us about memory speed, etc. etc.
Final point: Those fan boys don't pay for their equipment, they get at least most of it from the manufacturers. Providers at LET however must buy and pay for their equipment which may well translate to 500% to over 1000% the price for 20% better performance at best (and realistically more like 10%)? Bad deal, very bad deal. Simple reason: Now with gen 2 out lots of gen 1 servers are available second hand. And well noted, those "old" Epyc server can hold their ground against most of intel Xeon based servers incl. new ones.
TL;DR we get to see the result of what gaming and overclocking fans achieve when they play - for free, well noted - "test engineer" with what they consider a "server".
True their test setup isn't 100% indicative of server production loads but it's relative testing of 2 cpus on same server/environment. That's all a comparison can do within the confines of what you'd have available for testing.
Though I agree I'd like them to use HTTP/HTTPS web server/database type loads focused on more. But all these high core count servers have intended uses beyond what I'd usually use so guess that's what is tested.
Is it really just hype for EPYC Rome gen2 ?
Which would you choose
AMD EPYC 7351P 1st gen 2.4Ghz base and 2.9Ghz turbo with 16C/32T 155W
with 8MB L2, 64MB L3 and DDR4-2666mhz
at $750
or
AMD EPYC 7302P 2nd gen 3.0Ghz base and 3.3Ghz turbo with 16C/32T 155W
with 8MB L2, 128MB L3 and DDR4-3200Mhz
at $825
For me the AMD EPYC 7302P is a better deal even at +$75 more.
But yeah actual retail prices differ from RRP. I've seen AMD EPYC 7351P between $770-840 while AMD EPYC 7302P at $999-1050
Indeed true depending on which Intel Xeon cpus you stack them against though the AMD EPYC 1st gen loose out of cpu frequency - especially for all core turbo boost max.
AMD can provide more cores with the same volume, which means it can produce more VPS. So you understand
If one can't do server testing with typical server loads then one stays away from it, simple as that.
Nope, that's virtually worthless, except for rather exotic cases.
As a LET provider I would choose the Gen1 based node because I can get it at a fraction of the price of a Gen2 based node, which means I can earn my investment back much faster and earn money, I can buy more nodes for a given investment sum, I can offer VPSs for a lower price.
If on the other hand I were a (not hosting) company and needed a new powerful system and new I'd highly likely go for the Gen2 server. But this thread was about a provider and one that also addresses the low end segment.