All new Registrations are manually reviewed and approved, so a short delay after registration may occur before your account becomes active.
Xeon vs Epyc and Ryzen, an excursion with a surprise
I have a couple of systems myself plus access to some more via a couple of friends, colleagues, and clients, and I occasionally look through the many benchmarks I've done.
That's how this started, or more precisely, it started when I looked over some result sets involving (or in LETese "involucrating") diverse processors and I suddenly noticed something strange: a "boring" Xeon with a performance number significantly above the rest, in particular above diverse cool and supposedly unbeatable, fast Zen based systems.
Huh?
A boring 14 nm Xeon beating a 7 nm Zen processor? Strange!
Well, some of the benchmarks were done with an "old" version (2.0.5), maybe that was a factor? I was sure, and I mean absolutely certain, that the relevant code hadn't changed, but hey, maybe I should carefully pick some known to be good Epyc and Ryzen VPS and run the current, and in particular the same benchmark version on all, just to exclude any even tiny doubt. And that's what I did.
I picked
- a NexusBytes / @seriesn VPS (Ryzen 3900)
- a Contabo / @contabo_m VPS (Epyc 7282)
and ran a dozen benchmark rounds on all 3, including the InceptionHosting / @InceptionHosting Xeon E-2278G based VPSs, using the most current version 2.1.0a.
Here are the results:
AMD EPYC 7282 16-Core Processor [4C, 8G]
2.8 (3.2) GHz, L1 32Ki,32 Kd, L2 512K, L3 64M (4M/C, 2M/vC), 120W (3.75/vC)
ProcMem SC [MB/s]: avg 251.4
ProcMem MA [MB/s]: avg 509.9 (202.82 % SC)
ProcMem MB [MB/s]: avg 634.3 (252.31 % SC)
______________________________________________________________________
AMD Ryzen 9 3900X 12-Core Processor [4C, 4G]
3.1 (4.3) GHz, L1 32Ki,32Kd,L2 512K, L3 64M (2.66/vC), 65W (2.7/vC)
ProcMem SC [MB/s]: avg 439.0
ProcMem MA [MB/s]: avg 1109.2 (252.67 % SC)
ProcMem MB [MB/s]: avg 1157.5 (263.67 % SC)
______________________________________________________________________
presum. XEON E-2278 8 core Processor [2C, 1G]
3.4 (4.6 to 5) GHz, L1 32Ki,32Kd, L2 256K, L3 16 M (1M/vC), 80W (5W/vC)
ProcMem SC [MB/s]: avg **495.6**
ProcMem MA [MB/s]: avg 830.4 (167.55 % SC)
ProcMem MB [MB/s]: avg 814.1 (164.27 % SC)
Huh? A boring workstation Xeon with an onboard GPU (which makes it look toy-like in the eyes of many hardcore server pros) beats a Ryzen? How come? What's the explanation?
(Note that the Xeon system only had 1 GB memory which is less that the test set and the other systems)
Before diving in let me first quickly throw in my own workstation based on a Ryzen 4750G (just a single non virtualized run)
Ryzen 7 4750G Pro 8 core processor [16C, 48G]
3.6 (4.4) GHz, L1 32Ki,32Kd, L2 512K, L3 8M (1M/C, 0.5M/vC), 65W (4W/vC)
[PM-SC] 457.14 MB/s
[PM-MA] 1.44 GB/s (315.9 % SC)
[PM-MB] 1.56 GB/s (341.3 % SC)
The 4750 having a slightly better single core performance than the Ryzen 3000 is easily explained by the 3000 being measured in a VM and also being slightly older. The significant performance difference in multi-core though brings up questions ...
But first let me provide a "legend".
The '[xC, yG]' shows the number of vCores seen by the benchmark (or any other) program as available. Note that each header also mentions the "real" cores (as opposed to SMT vCores). Also note that a providers VPS specification may have "vCore" or "vCPU" to mean pretty much whatever they please, so a VPS with say 4 vCores (sometimes sloppily called "cores") might actually mean anything between a fraction of a real vCore and 4 (full) vCores (and in very rare and exotic cases even physical cores. Contabos VSDs are an example).
The line right below each header lists (in this order) "base clock" "(turbo clock)", L1 (instruction and data) caches, L2 cache, L3 cache, and "(L3 cache per vcore)", plus the (official) TDP and the (official TDP per vcore).
At this point I could end by saying something like "Zen processors are not (always) faster than Xeons". 495.6 is more than 439, period, have a nice day.
But of course I won't. I'll try to explain more about it.
The first noteworthy observation is that the Xeon really leads only in single-core; in multi-core the Ryzens just blow it out of the water while the Epyc trails.
Let's look at that. First have a look at the number in pharantheses behind the MC performance number; it tells you how much more you can get done using multiple cores (relative to the single-core result). See it? Now the Epyc, while being slower than the Ryzens and the Xeon, doesn't look that bad anymore. In fact, it gets more work done when multi-threading than the Xeon (ca 200% vs ca 165%).
Plus, at least in data centers, there is often the question whether single-core performance is really that critical. After all, servers tend to have a certain type of workload, namely serving, which translates to "lots of waiting" (doing IO) where a high single core performance isn't worth a lot - but multi-core performance is. Why? Because it matches typical server loads well.
But there are of course also jobs that do need high performance; crypto is a good example. (Hint: SSL handshakes, in particular RSA ones, are very serious performance gobblers).
But the playing field is largely defined by the providers and those have to look at the cost factors. If you don't know the hosting world well my primer may be helpful.
Wrt processors TDP (and TDP per vCore) is probably the most important factor. That is one very significant point where Ryzen leads over Xeon which uses almost double the electrical power per vCore. This alone can decide whether one can make an attractively priced offer or not.
But things are always more complicated. In this field one factor that compensates for the higher power consumption of the Epyc is how many cores one can pack into a box (or a rack). Ryzens are single socket only but Epycs allow for 64 and even up to 128 cores (256 vCores!) in dual sockets in a box which translates to the cost of the whole box being spread over more cores (and hence VPSs) which boils down to lower cost per VPS. Plus it's cheaper in terms of switches, which, if professional ones are used, are considerable.
A small example: 12 cores per box vs 64 cores per box. With the smaller box one needs 6 switch ports while with the big box one just needs 1 (albeit a larger ~ more expensive one). Plus, there are dozens of boxes in a rack and if many switch ports are needed one will have to use 2 HU TOR switches.
In conclusion the three main points I see wrt intel vs AMD (in the DC) are these
- Nope, AMD is not really light years ahead in performance. One reason for that is that intel processors are not bad; once they can produce them in 7 nm (or at least 10) the situation will change. The other reason I see is the fact that many, probably even most, VPS users do not really need the performance. For 80+% of use cases e.g. a E5-26xx v4 is plenty powerful enough. Oh and btw, the Ryzen is not meant for the DC but for the desktop.
- Density. Where AMD indeed is significantly ahead is density and density price. AMD can offer a 64 core processor at a relatively attractive price. intel otoh can't offer that - now. But that will change.
- TDP. that's a problem and a part of it is the fact that intels TDP numbers are worth next to nothing while AMD's TDP numbers are at least ballpark correct - and lower, which (see above) can be a killer factor in the DC, especially now with all the green initiatives.
But, and that could be a killer 'but': intel also has their own fabs and plenty. The day intel has a 7 nm process working and production ready with good yield will be a very bad day for Lisa Su/AMD.
Comments
Is this an ad of the mentioned providers?
Are you paying for those servers or just got them for free?
Those happen to be mine (payed for).
I don't get why you think that this might be an ad. From my perspective it's a simple matter of honesty and transparency to say what it is and where I got it from what I'm talking about, plus that it's real, even normal products and not some weird exotic systems.
And my point was about Xeon vs. Zen (in hosting), not about this or that provider.
You don't need to write from which providers you were testing then this article might interesting.
Otherwise, I see this is a shameless ad for those providers.
the contabo vps costs 5 eur + VAT
nexusbytes VPS costs 16 USD + VAT
Inception hosting vps costs 2,50 EUR + VAT
Rest assured that I'd never shill for @InceptionHosting. Nor have I the slightest interest in selling or renting out my workstation.
You'll probably call this shilling too but NexusBytes just happens to be the be the provider from whom I purchased the best of my very few Ryzen based systems. And I wanted stable system for the comparison.
Sorry, but I'm not interested in that kind of conversation, nor do I like "games" where someone throws whatever weird allegation at one and expects his victim to explain and defend himself. We've had enough of that kind of ugly behaviour here during the last couple of days.
.
^^
You looked like a tech guy but come on, you're comparing two different CPUs in a virtualized environment.
Does it make sense?
Maybe 5.00 GHz > 4.6 GHz > 3.2GHz
Does it make sense? What is the point of this comparison then?
Epyc and Ryzen CPU have more cores and you can gain performance only by parallelizing processes.
OK. Let me try to explain that to you so I won't look like throwing some shit again.
You've three different VPS and surprise one of them was better than the others.
Why?
There a lot of factors that might affect the performance of your tests:
From inception hosting tos:
Nexusbytes tos:
And contabo didn't mention any limits or I wasn't able to find so basically CPU is restricted.
So basically your test at inception was bursting full core which specs your tests showed.
Looking at these two providers it's suitable only for serving HTML pages, normal applications would be suspended.
That means comparing apples to oranges doesn't make sense at all.
You could compare that by getting those servers and running tests even that doesn't mean you'd get a different result because speed by CORE is different between them.
@alexvolk
(a) my topic was processors, not VPSs
(b) accordingly I did not compare VPSs but processors
(c) My original question was whether Xeon is really "not useable", not comparable, or even "dead" (as quite a few speculate).
(d) a part of my benchmark does test speed per core.
But thanks for reminding me where I am and what seems to be considered attractive content around here.
Besides the difference in environment mentioned about, the benchmark results also depend on how you benchmark it, i.e. which instruction sets you are using.
Moreover, in your results, the ratio between multicore and single core is much less than the number of vcores, especially on your workstation it is much less than 8, I highly suspect that your code has some problems.
Now do the same tests on dedicated.
Aa for Intel doing 7nn or even 10nm they have said themselves that it's going to happen in 2023. By that time TSMC is already down to 3nm.
Regardless of anything else, if he paid for them then it can hardly be an ad. Surely that would also apply to anyone who commented about a service they purchased.
And I suspect that you don't know what you're talking about. It is a well known fact that the performance of multi-threading all cores is considerably lower than 'number of cores x single-core performance'.
I'd love to, but I don't have a private rack full of reasonably current systems. But based on what tests I could do (not many) the difference (on KVM virt.) seems to be small, something about 3% to 5%.
Yes indeed. Plus, if I had not mentioned the providers there would have been some complaining about that.
A poor multi-thread code will be much lower than 'number of cores x single-core performance', but a good multi-thread code will be close to 'number of cores x single-core performance'. If you cannot get close to such value, it means your code is poor.
Of course it depends on how easy it is to parallelize the jobs, but for benchmarks purpose you should adopt the jobs that can be easily parallelized.
Long time ahead. Intel is barely at 10nm now. Intel also has HUGE power states at sometimes 250W and more.
Intel will not use their own FABs anymore. Intel now buys capacity external. Thats how bad their nodes are.
The ONLY thing Intel beats AMD at is AVX512 - And that only at INSANELY lower clocks and extreme heat. So they don't beat them there either.
Not sure what you try to prove - in ANY synthetic benchmark an EPYC will beat any Intel CPU, up to a Platinum 8180 at less power use.
Thus, performance per watt is higher on AMD (and cheaper), and this entire thread is pointless.
God damn, I need to join the bashing here - comparing CPUs by using VMs from different providers, architectures, rules set and not controlling the node is totally useless... and I think most of people here knows that you know that - that is why everyone is asking who paid you for this.
Source:
https://www.phoronix.com/scan.php?page=article&item=epyc-7003-linux-perf&num=7
I'm too lazy to handpick pictures.
。。。。what surprise？？
Usually i'd say Intel, considering they offer me CPUs for free all the time just to not buy AMD at my day job, but i doubt anyone pays him. The testing methodology is just flawed from start already and tries to prove a point that does not exist even in theory.
I will not run the math now, but i assume even if you double AMDs power consumption (so a rough node change from 5/7nm TSMC to Intel 10nm+++/10nm++++) they still come out ahead, at less price, so i don't see what is there to prove at all - Right now AMD is just punching Intel left and right, deservedly so.
And this thread ignores entirely that AMD simply has the better tech right now (multiple CCX, IO die on cheaper 14-10nm node, Infinity Fabric, PCIe 4.0, more memory channels at higher clock, paths to 8 socket systems while Intel barely can do 4 anymore...)
how dare you! ;-)
@jsg interesting topic, especially if one keeps in mind, that power consumption and such is only of interest for the provider but rather not for the customer in the end.
I have seen comparable results in the more common geekbench with some VMs with E-2276/78. so these are a good alternative for clients if the provider manages to keep a comparable balance ... with ryzen they more likely cram even more people on the same box, so more cores only benefit the provider, not the user 🤷♂️
however with that I also have to agree with what others said about the uncontrollable situation within a virtual environment... you can't say for sure, if there are any limitations in place that keep the VM you're getting from utilizing full power. or if the nodes are already so full, that you simply cannot get close to dedicated performance specifically for your benchmarks.
from the view of a potential customer it'll always be kind of a gamble if you don't know the provider already ;-)
Also, for simple comparison - This shows everything:
You can guess which one is which.
The market has spoken.
(Disclosure: I hold both AMD and Intel stocks. I also have short positions on Intel.)
As a product AMD stock holder, the 2nd one
Yeah, Yeah, you are right. Old habits ... sorry
Repeating BS doesn't make it true. Try again when you have learned and understood the basics of processors.
Yes, intel is behind. Yes. intel has f_cked up big time.
BUT: For a start those numbers don't really mean a lot. A 10 nm process can be similarly small as a 7 nm process. Besides those numbers almost never really match the gate size.
But if you want to stress it, let me ask a question: What is AMD's process size? Answer: None; it's TSMC's process size, not AMD's. GlobalFoundries, which was spun out of AMD and still is closely linked has 14 nm, even worse than intel.
I don't fully agree. As '7 nm' and soon '5 nm' is THE process to have, thanks more to marketing than to technical factors, then intel, already having lost quite some time, needs to switch to 7 or 5 nm NOW and can't afford to wait till their fabs are modernized.
Yeah, yeah, I know, that's what the tech gadget "journalist" say, and I don't care; neither for their blabla, nor for AVX512.
What I see and care about is numbers and facts, and those show that intel does have some processors that have higher single core performance than Epyc (both from about the same time, 2019)
Not in mine and not in some others. intel knows the x86 architecture and has lots of experience.
That said, you are right though re intel processors tending to be power hogs plus providing very questionable TDP numbers. Btw, I did mention that.
Those benchmarks look at Epyc 3 - not 2. And those benchmarks are utterly worthless for this discussion. Example: "compiling linux kernel" (or whatever) includes a lot of stuff (like type and amount of memory and disks) that waters down the relevant numbers.
Thank you. And: depends. As I know from painful experience it does become of interest with dedis. And it is of course of interest with VPSs too albeit indirectly
I don't care. At all. For "the market" (even important) technical factors are but one ingredient among many.
@all
Kindly note that I did not say "intel is better than AMD". In fact, I didn't even say "intel processors are good" or that I like them. What I said basically was "intel processors are not sh_tty and light years behind AMD".
Part 2, general
Do you think that providers like NexusBytes / @seriesn or Contabo / @contabo_m don't think, and think hard, about which systems based on which processors they buy and sell? You bet that they think hard and a lot about that and that they diligently calculate all relevant factors.
Ryzen is a simple choice for small and mid size providers. Simple reason: everybody and their dog want AMD Zen based VPSs plus Ryzen is attractive in term of low power consumption.
So why doesn't everybody offer Ryzen based VMs? One major reason is that one can't get the nodes that cheap as demand is high and the processors are relatively young which means such nodes can't be purchased dirt cheap (unlike e.g. Xeon 2600v4), so especially smaller providers hesitate.
Large providers on the other hand often have a very different approach along the line "most VMs per rack" which also means "most cores per node". Epyc is an obvious choice. They usually gobble more power than Ryzens but still less than comparable Xeons and they offer a basis for the goodies that business customers like, e.g. multiple NVMes and an amount of PCIe lines that gives room for dual 100 Gb adapters, etc. And all of that cheaper than a comparable Xeon based node.
As for "intel having fabs isn't worth a lot" I suggest you ask Huawei, who lost billions and a major market segment because they don't have a fab. If TMSC, for whatever unlikely reason, cuts the link to AMD the nice Zen pony ride finds an abrupt end.
Do I like or prefer intel? Mostly no. But I've understood that NO monopoly is good for us users, while quite a few here seem to simply have exchanged an intel monopoly for an increasing AMD monopoly.
Finally, I also see what I really need. I didn't get my Ryzen and Epyc VPSs because I really need the performance. If all one really needs is to handle a couple of hundred or even thousand requests per second - which is the case for 80+% of users incl. many companies a decent - and much cheaper - Xeon system will do fine and it won't gobble up more power than an AMD.
I see advantages with both AMD and intel (granted currently much more so with AMD) and I want both to survive and prosper because that is the best for us buyers/users.
@jsg you what mate? That was just ONE benchmark from Phoronix. And even more it doesn't matter if it's zen "3" or not, it's still going rampant over anything Intel has. Including their "platinum" offers.
Just admit it, Intel has fucked the pooch on their startegy. They never thought AMD would catch up.
Actually, he doesn't seem to know this.,hence all the people telling him that these results are useless.
Processor X is processor X, no matter the node. And I intentionally carefully selected good providers. Plus I ran quite a few tests with servers that I have here in my lab and consistently found that, provided that a server is not brutally oversold, VM performance clearly follows server performance.
What are you talking about? He's right. It's completely dependent on the code and what needs to be serialized and what can be done in parallel. You'll never see a fixed performance ratio between single and multiple cores and that's why you test with different types of real world test loads. Basing something on one synthetic test from your non-validated program does makes it look like you don't understand processors.
Oh, are we back again at you, knowing virtually nothing about programming explaining to me who actually wrote a benchmark software, how things work?
Plus, you even got the guy you are defending wrong.
@jsg thanks for taking the time to put something like this together, nice one.