@@ServeTheHomeVideo I'm glad that you and your team are dedicating a ton of time to stuff like this. Having this information in an easy-to-digest video is difficult by itself, let alone the logistics & testing. Thank you for all you do!
I think single socket will be the best way forward for most businesses. The core count and performance of a single socket is already fantastic, and like you said the licensing costs more than the hardware. Everyone I've talked to about Azure Stack HCI I'm saying to go with 16c32t nodes until they hit 6-8 nodes in a cluster. 16c per node is the minimum to license with DC on an EA, and you have to account for N+1 node redundancy. Even if you were pushing it a bit more and went with 6x 32c64t nodes, one socket can handle that just fine. Obviously if you're looking at 4 racks full of servers, then going dual socket and consolidating the number of nodes could save you some money. But most businesses I've talked to don't need more than 8-16 nodes in a single location. Really my wish would be for Intel to release a platform in-between LGA4677 and consumer that's a single socket with 16c32t chips and 6-8ch ECC that's "cheap" and low power.
Very impressed to see Intel coming back into the game with their lowered idle power consumption. Interesting to see where Granite Rapids and eventually Diamond Rapids lands down the line in about 2-3 years.
Great Video!! I know that you are highlighting the single socket SPR and EMR processor platforms but I must say that the 2 socket systems are also very phenomenal systems. In our SAP HANA testing for vSphere support we have observed user throughput levels that match the 8 socket Cascade and Cooper Lake platforms. These systems consume way less power too. The only TCO bummer is the use of DDR5. In sizing these systems for HANA, the memory is a big part of the cost. However, I for one am very excited with the direction that Intel is going. I can't to get my hands on the new Sierra Forest line and see how that performs.
@@zbigniewmalec4816 Indeed. It also gives you the option to go with 256gb rdimm too. But these are very expensive. We tested HANA using 128 gb rdimms, so 2TB per socket (60 core). These awesome little systems handled almost the same amount of current users as the older 8 socket cascade lake system with similar response time for OLTP and OLAP processes. I am very amazed at these processors.
Totally agree with you Pat. Most folks should really save on core count and invest the savings in more RAM or storage. I run my own private cloud on a dual socket 12 system with 24 cores total. Despite having some 20 VMs (including AI inferencing and in-memory databases), the average CPU utilisation tends to be under 10%. Also, higher GHz tend to be better overall.
In what way? Emerald rapids showed Intel with slightly higher core counts with slightly more performance, but less power consumption across the board compared to the AMD CPU tested here.
Interesting! A couple of suggestions: since the BMC is running even when the server is off, you could have powered down the server, measured the wattage used, then subtracted that from the total. Also, there are USB cables you can get with a built in LED that shows how many watts of power is being used that you could have used to find the power draw of the USB NIC. And given the different core counts, I think Watts/Core probably makes for more comparable numbers rather than just total system power.
Like others have said, I find myself hunting for pcie lanes and ram channels for homelab, more so than higher core count. At decent prices. I look forward to getting one of these EMR cpus maybe in 3 years when cheaper and mated to a cheapo board. Great video, highly educational 👍 Thank you.
I'd be more impressed if the idle power was below 20W. I have a 12-core Haswell-EP that idles at under 2W without any special setup. Sounds like modern Xeons could be a lot more aggressive with power saving.
True, got E5-2650Lv3 as my offsite backup server. Barely sips power, but man, server boards at those days really doesn't have much wiggle room for standby like S3/4. Either on or full on system halt 😅. Need those pricy workstation board for better idles power
The CPU's tested have about 3-4x as many cores... and additional components such as additional accelerators for "AI" more PCI-e lanes just more everything. That's going to add up in terms of idle power draw.
@@jolness1 yup, like I pointed out in my first comment, in Broadwell/Haswell-E era, motherboard is also an contributing factor. Even with dual socket Asus Z10PE-D16 WS has proper idle and standby mode implementation (note that the motherboard still using server grade C612 chipset).
2W sounds like BS figures that some internal metrics report but are always far off from actual power consumption.. I'd be surprised to see any S2011 system draw less than 80W from the wall. The lowest I've seen is 16W from the wall on C236 and that's rare AF.
I can't speak for all HomeLabbers, but for me I needed PCIe Lanes more than raw compute for my HomeLab. So in January I went with an Epyc 7313P on a RomeD8-2T board with 256GB of DDR4 3200 RAM. I found the Milan chip for $470 shipped, bought the mobo new ($666), and the memory was $381. I couldn't find a better value for Milan or newer platform (Intel or AMD) so looking back I'm still pretty satisfied with my choices.
Totally. In 3 or so years when EMR is the same age as Milan is now, I would expect more value options for home labs. When that happens, a point of this was showing the difference in power consumption and performance when picking SPR or EMR
This is interesting but I think AMD is still winning here. You can get a lot more cores and a lot more cache with them and they are clocked a lot higher than what Intel is providing here. I would be interested in seeing a similar comparison on idle power draw with them as well but they do have a lower TDP and if their consumer chips are anything to go by they will likely draw less power. That being said it is also interesting that Intel is actually the "budget" option now. The closed competitor to the 8558U would probably be the 9454P and that thing cost 4,598 USD.
@@ServeTheHomeVideo Yes but it is only 4MB different which is negligible. Again I tried to find the closest part. If I wanted the fastest part I would have compared to the 9474F which has a base clock of 3.6Ghz or almost double that of the xeon. Comparing top to top we could have compared to the 9684X which has 2x the cores at 96 and even with all those cores the base clock of each core is still 550Mhz higher. It also has a whopping 1.1GB of L3 cache. AMDs chip even gets higher single core performance as far as I can tell thanks to the massive cache and higher clock. So in my mind the real story here isn't so much that this thing is beating AMD because it probably doesn't even manage that in the power consumption you are measuring. The story is that they are actually the budget option compared to AMD now being like a full 30% cheaper. They are better than their last gen and I do think it is worth evaluating these processors. I just don't know if your power efficiency or even software argument hits it here because if software licensing is done on a per core basis then I probably would go with that 9474F that is quite a bit faster. If power consumption is the thing here then that 9684X is going to be the better option because it still has faster single core and still has 2x the cores. Yes the TDP is quite a bit higher at 400W but as you said that is only part of the power draw of the whole server and additionally is only part of the power draw of the whole rack. A single server with 1 9684X or the 9754 is going to consume a whole lot less power than the 2 or 3 servers you would need to build around the 8558U.
@@chaosfenixwell it is different in that the L3 on Intel is ‘more’ continuous than on that AMD processor. In the Intel processor that’s 260MB spread across two chips, or 130MB per chip, versus 256MB spread across 8? Chiplets at 32? MB each (not exactly 100% on AMD’s product line). And so you could have 128MB of data entirely in L3 cache on each chip for Intel, whereas that would be spread across 4 chips on AMD. And I’m pretty confident that Chiplets on AMD don’t benefit as much from cache on other chiplets, otherwise AMD would would offer CPUs where they replace CPU chiplets for just the 3D Vcache chiplets (double the l3 of a cpu tile), or even double vcache chiplets (4x the L3 of a CPU tile and 32MB more than even the X3D chips have).
@@levygaming3133 Yes my point isn't that there are differences. They are different companies. I just tried to find the closest competitor in the stack and at similar as possible specs AMD is going to win overall in most areas but price. The number of workloads where that 4MB of cache or smaller cache latency is going to win in performance vs a chip that has almost 2x the clock speed has to be negligible at best. Again, it isn't going to win on performance or power here. But it is winning on price which is where I think this thing is the most interesting.
That's a great improvement from Intel for a single socket solution (while still great for AI because the included accelerators and such)! Why didn't more people talk about this??
Here in Europe these efficiency gains really do matter a lot. I wonder how efficient these are compared to Milan if you're able to utilize the onboard accelerators
@@xl000large swathes of Europe abandoned nuclear power in favor of tying their power grid to Russian oil, natural gas, and (German?) coal in the vein hope that this would reduce Russian aggression and not, as actually happened, emboldened future aggressions. As a result of recent trade embargoes on Russian energy sources, and some members of the EU starting to finally realize coal is bad, power in Europe has increased in price dramatically. I’m pretty sure I saw reports out of the UK that people’s power bills were eclipsing rents. And even before this, power in Europe was already more expensive, for one reason or another. For example, in the US, 6 cents per kilowatt aren’t uncommon, with 13 cents being relatively high. In the EU, I’ve seen figures ranging from 21 cents to 38 cents per kilowatt.
Guys will you able to make some compatibility tests on tinyminimicro machines with 96GB RAM modules? Youve tested it with HP 600 9gen and mentioned that you also tested it with other 1L pc's but you did not mention which ones. Such compatibility list in some article would invaluable for all folks who want to buy miniPCs for their first homelab. And as we know RAM is the most important limitation for them.
some people are physically incapable of making the S sound I wonder how this works in other languages, like Arabic Japanese or Spanish, with a very different set of of base sounds. People in those cultures are made fun of for not being able to produce other arbitrary sounds I guess
How much of the power consumption is coming from the 12x 40mm fans, + the PSU fans though? How does this stack against 2u dual socket systems when you factor in the watts wasted on blowing air? (Also hearing damage)
If you don't need a lot of PCIE Ryzen cpus have been the go to for a homelab for several years now. Im using a 5900x with 128gb RAM as my homelab and it is working great with low power consumption because of slightly undervolting.
@@ServeTheHomeVideo That is true, but if you need more than 128GB of memory you probably need more PCIe and have a larger budget also. Been following your website and channel for many years now and I love the content. Greetings from Norway :)
Lower IDLE power makes perfect sense. It's a newer node, more modern package, and probably more advanced power management. It's maybe not something you should expect with every generational upgrade, but occasionally it does happen, especially after several generations.
"Occasionally it does happen." Well . . . Intel 10nm Enhanced SuperFin n.k.a. *Intel 7* per July 2021 company press rel. Loved that moment in Intel history ❤ I think it was called "Accelerating Process Innovation." Exactly, change name 😂
Licensing software on a per core basis is the STUPIDEST decision I have ever seen. It gets the automatic response of, well somebody else provides something better. When I saw it in windows server, I was like. guess I'll make it work on linux.
i think it's important to think about the average load you plan on running on your servers when considering these sorts of numbers. obviously idle and fully loaded both need to be included in this sort of review comparison, but when you're making a purchasing decision one or the other of those numbers might have very little relevance to your future power bills
My 4th gen Xeon is both fast and power efficient (E3-1285L v3). I could buy a Z97 board and a 5th gen Broadwell Xeon (faster and more power efficient than Haswell and with L4 cache), but that would be too expensive at this point.
i have a quick question about that idle power thing. do cpus not in idle neary use 0W. i mean the mainbord it self uses power too. i have a r630 it uses 150W idle. i know it can go low as 60W but i dont have only 1 cpu and one stick of ram and running of a usb stick ok. my system has 4 sata ssds in it and 4 2tb sas12g hdd, all connected to then perc h330 mini in two virtual drives, then i have 2 xeaon E5 2699v4 in it and 768gb of ram with all 24 slots populated, also i have the dvd drive and the display also i have the idrac8 enterprise and a 2x2 10gbit nic. so year this adds many power hungry components. 20W alone for the idrac, each ram module uses 2w to 3w each so this is another 48w to 72w, the h330 mini uses also 20w so this adds up fast. so i think cpu idle power is nearly 0w what is shown by the system monitor. i guess this is correct. i dont kow for sure what happen if you enable dynamic power but much deeper you cant go and i thinks this impacts the performance and would lead to bad performance response and at some point it would not mater because the system wont go deeper because of no need for c states accordingly to the system utilization. if you optimize the components around and throw away everything you dont need this is possible. ass i sad if you go with the minimum requrements you can see impressive numbers but at some point this does not longer mater.
Desktop Intel CPUs can dip down into the single digits because they are fully monolithic, and have relatively few memory channels and I/O connections to keep active. These chips all have to power a lot more cores (48P instead of 8P+16E), 4x the memory channels, many time the I/O connections, and they spend some power on the connection between the CPU tiles. You can see some of those idle power losses looking at Zen4 on desktop, where the 7950X often sits above 20W where an otherwise more hungry i9 will happy sit at 10W.
It would probably beneficial to know what C/P states were enabled and if they were the same accross testing. And how influencial they are int saving idle Power.
"Our servers burn less power when not doing anything" isn't the win this ad wants to think it is. Maybe my business is just doing it wrong. We tend to deploy servers to do things. 😅
Huh? We showed the 300W TDP CPU out-performing the 350W TDP CPU and over 80W less at the wall. So this is idle up to 100% load. Also, we have to say sponsored even if the chips are marked as non-retail confidential. Very odd comment.
@@ServeTheHomeVideo You're right, the generational gains at load are noteworthy. I just found the heavy focus on idle usage very strange for a server bound product. As for the rest, sponsored is sponsored. For the duration of this video you are an extension of the sponsors marketing department. It doesn't make the video "wrong" or not worthwhile, but it is what it is. It should be watched in that context.
Give it a few years and I'm sure there will be a plethora of cheap ail-express boards and a flood of these on the used market as Sierra Forrest and the like take over the high-performance segments.
So, I'm looking for a replacement for the E5-2630L v4 (a 2016 10c/20t processor) for my homelab. It has 40 PCIe lanes, and as far as I'm aware there's nothing like it left on the market. Neither AMD or Intel seems to make these types of processors any more, where you get low wattage (55W), but a decent amount of PCIe lanes. Anyone know upgrades?
You've pretty much fallen out of the enterprise range at that TDP unfortunately. Something like a Ryzen 9 7900 65w, ECC support, 24 lanes of PCIe gen5, combined with an X670 board that provides an additional 12 lanes of PCIe gen4 would likely be the way to go. Supermicro, Asrock rack etc for boards with a BMC for IPMI. If you can't find a suitable board with IPMI, a Ryzen Pro 7945 could be ok in a consumer board combined with the AMD Pro tech for Out of Band management.
My production system has three hosts with 10 cores each (Scaleable V2), but then again I only have 5 windows servers and a few debian thrown in. Cost was the factor at the time or I would have gone with 20 cores each. I didnt think about memory channels, only 2 modules installed in each host, probably should have filled all the sockets for a little faster performance. Used a 2x64 config. so I had room to grow.
Great video! It’s a bummer (in this case) that there is no eps12v for the cpu to just put an amp clamp on. I think this is the best way of isolating given the constraints. I’ve collected data like this before and it is very time consuming. Appreciate the work you guys do,
Thanks. We had the package power in this one but I generally do not love using it. The real number that matters is at the PDU since that is what drives rack sizing and overall consumption
Single socket systems are the only ones that really make sense now that a single socket can cram all of the things into it. Sod non-uniformity, it complicates things. Maybe the gen after sierra forest intel might be competitive again. It is amusing having to talk idle power consumption just to give them a positive spin.
@@ServeTheHomeVideo I didn't say just idle, it was just the most prominent example. The 8558U being better than the 8458P is good, but as the 8458P is from intel's "struggle to keep up by pumping power" generation, it doesn't really say much. The e core gen after sierra forest is the one I'm looking at for intel getting back in the game, sierra forest is already a bust as the main benefit of moar cores has already been handled by the competition beating them to the punch, for sierra to be interesting it needed to be out years ago. They've done the delay tactics of scrapping avx512 and making a mess of redefining 256 bit with avx10.256, lets hope the gen after sierra has avx10.512 (I'm expecting more delays with them using avx10.256 and using misleading marketing, but who knows). If intel messes up clearwater then we are in a bad situation as AMD will be the new intel, I don't have faith in ARM competing more than in specific niches. I know you're getting rolled in the comments, I'm not trying to pile on I just gave honest feedback about the content.
Still dreaming that Intel continue making something like Haswell/Broadwell E Low power SKUs (Like E5-2650Lv3). Since atom C series back in denverton days was a thing... Welp we never gonna see a true low power enterprise CPU SKUs that still offer upgrade path for more powerful one within the same socket/platform.
The increased power efficiency at idle is normal, but I think it's worth clarifying a few things for folks about why it's surprising/exciting here. 1. Generational leaps are usually pretty small nowadays. Bigger jumps in efficiency are getting more and more rare. An example on the desktop side would be AMD's socket AM3, AM4, and AM5. AM3 doesn't idle very well. Their Ryzen CPUs on AM4 began to see some really good idle optimization, with some of the 5000 series able to idle at
Just so you know, I was one of like 40 people in the room in Oregon during that presentation in 2017 so so remember it well. These days, chiplets are the way forward
I'm more of a user so I have a different approach to server management. Compared to an IT person. I do a cost benefit analysis for my workload. Does the cpu count / gigahertz / memory equal what performance metric for my workload. Then based on that, cost. Weirdly some cpus you expect to be faster are technically faster in the performance metric. But the cost per server is not proportional to the improved performance. So I usually go a little lower and am able to buy 2 or 3 more servers. If I'm not focused on cost. I care about density. So whatever configuration gives me the most memory and the most compute cores. Currently I'm running dual socket xeons. Do I care that it's a dual socket xeon? not really. But reliability is better among intel cpus verse amd. It might be another component the memory or the motherboard. But in an array of 20 the intels have no crashes while the amds have 3 crashes a week. This is a hardware level crash where the system reboots. So it's hardware related and random enough that I don't want to track it down.
I enjoy the way that you explain the benefits of assorted server choices. BUT when the graphics on your shirt match the background of your set & it's decorative lighting - it's just weird. I can't hear you cause it's so odd - like costumes for kindergarten or something.
Home lab. $10,000 CPU? You and I live in very different homes! Oh, and I often chat with Intel about the CPU set I'm about to use ... not! This channel has gone so far upmarket I see less and less relevance.
Hi, so others seems to have notice what ARM is doing since a long time… I think it is the future, more and more calculation/speed with less power, which will be awesome for everybody.
We have been doing Arm servers since 2016 (see the main site Cavium ThunderX pieces.) Remember that the PPA gains are for the execution cores, not the SRAM caches, PCIe controllers, DRAM controllers, and so forth. Arm has been ahead in higher core count but lower performance cores optimized for cloud native workloads.
Assuming a server draws 500w constantly 24/7 (likely not the case if there are some idle moments - watch the video), that's about 12 kWh per day. Lets assume the worst and you live in CA, where power is around .50c per kWh. That's 6 bucks per day ($183ish per month). You're off by a zero, possibly 2 if your power is cheap.
single core performance was self referencing among the 3 cpu and not comparable with the rest of the world CPUs. geekbench or similar real value would have been more useful. at the end you can have 48 cores but the single core is slow you need more core for the same vm on a server compared to the other. the comparison of the video was more focused on the consumption than the real cpu performance, which was fine, but impossible to compare with rest of the world servers.
Geekbench 6 is borderline useless for servers. Or to be a bit more clear, it is useless for this core count of server and higher. Over maybe 12-16 cores GB5 is usually more indicative of server performance than GB6. The vast majority of workloads, outside of things like HPC codes, on modern servers do not run on full chips. Usually they are run in containers and VMs to increase overall system load.
@@ServeTheHomeVideo I mentioned GB as example you compared the core speed among the 3 chips saying the slowest 100%, the others 120% and so on, that means nothing in real world of performance. I would like to see how the single core speed is compared to my server chip that runs esxi... I might upgrade in future...
Why not compare to amd equivalent parts with low core count and high frequency? Is this a fluff piece for intel? The intel pillows in every shot surely suggest that it is.
I think that Patrick has to first evaluate the parts for what they are advertised for and suggested value to give a comprehensive picture of it's purpose. While I understand that the lack of comparison at every step is unsatisfactory for you, it does not mean it is a fluff piece. To be honest, I don't know how bad it would feel for the creator to have to deal with this much vitriol! And also he said in the video that AMD CPUs are better when he was giving a sponsorship disclaimer (or rather the lack of sponsorship) only for you to jump to the conclusion that he must be sponsored!!!
@@ServeTheHomeVideo I have posted my thoughts of the comment you are responding to, in your defense, but I would like to say that not having the manufacturers trinkets on display would help avoid this quagmire, however unwarranted it may be!
The idle power draw is a useless measurement for this setup. You paid a lot of money to get all these cores, which means you have a need for a lot of cores. You will find a way to use them to their full extent, otherwise, it was a waste of money.
I'm glad to see intel finally becoming more energy efficient. Thank you for all the testing you did to show this all off!
Thanks. It takes a lot to even get to numbers like these. Have a great day!
@@ServeTheHomeVideo I'm glad that you and your team are dedicating a ton of time to stuff like this. Having this information in an easy-to-digest video is difficult by itself, let alone the logistics & testing. Thank you for all you do!
v@@wittywilla
I think single socket will be the best way forward for most businesses. The core count and performance of a single socket is already fantastic, and like you said the licensing costs more than the hardware. Everyone I've talked to about Azure Stack HCI I'm saying to go with 16c32t nodes until they hit 6-8 nodes in a cluster. 16c per node is the minimum to license with DC on an EA, and you have to account for N+1 node redundancy. Even if you were pushing it a bit more and went with 6x 32c64t nodes, one socket can handle that just fine. Obviously if you're looking at 4 racks full of servers, then going dual socket and consolidating the number of nodes could save you some money. But most businesses I've talked to don't need more than 8-16 nodes in a single location. Really my wish would be for Intel to release a platform in-between LGA4677 and consumer that's a single socket with 16c32t chips and 6-8ch ECC that's "cheap" and low power.
Stay tuned for next week on STH. :-)
Something like AMD Siena?
Very impressed to see Intel coming back into the game with their lowered idle power consumption. Interesting to see where Granite Rapids and eventually Diamond Rapids lands down the line in about 2-3 years.
Personally I am excited for Sierra and then Clearwater
Great Video!! I know that you are highlighting the single socket SPR and EMR processor platforms but I must say that the 2 socket systems are also very phenomenal systems. In our SAP HANA testing for vSphere support we have observed user throughput levels that match the 8 socket Cascade and Cooper Lake platforms. These systems consume way less power too. The only TCO bummer is the use of DDR5. In sizing these systems for HANA, the memory is a big part of the cost. However, I for one am very excited with the direction that Intel is going. I can't to get my hands on the new Sierra Forest line and see how that performs.
Great comments. We have been running the cloud-native series comparing modern CPUs to Cascade Lake, albeit using 2P systems.
Ddr5 give you the option to go 96gb rdimm route
@@zbigniewmalec4816 Indeed. It also gives you the option to go with 256gb rdimm too. But these are very expensive. We tested HANA using 128 gb rdimms, so 2TB per socket (60 core). These awesome little systems handled almost the same amount of current users as the older 8 socket cascade lake system with similar response time for OLTP and OLAP processes. I am very amazed at these processors.
Totally agree with you Pat. Most folks should really save on core count and invest the savings in more RAM or storage. I run my own private cloud on a dual socket 12 system with 24 cores total. Despite having some 20 VMs (including AI inferencing and in-memory databases), the average CPU utilisation tends to be under 10%. Also, higher GHz tend to be better overall.
Great point
As someone whose homelab is still using 2690 v4s and 772gb of ram. This only made me see how ahead AMD still is on a pure value proposition.
In what way? Emerald rapids showed Intel with slightly higher core counts with slightly more performance, but less power consumption across the board compared to the AMD CPU tested here.
@@AlexSchendel Completely untrue, but also new vs old parts.
Interesting! A couple of suggestions: since the BMC is running even when the server is off, you could have powered down the server, measured the wattage used, then subtracted that from the total. Also, there are USB cables you can get with a built in LED that shows how many watts of power is being used that you could have used to find the power draw of the USB NIC. And given the different core counts, I think Watts/Core probably makes for more comparable numbers rather than just total system power.
When I say 5-7W for the BMC that is inclusive of the fact that when the server is running the BMC is usually doing a bit more logging.
Like others have said, I find myself hunting for pcie lanes and ram channels for homelab, more so than higher core count. At decent prices.
I look forward to getting one of these EMR cpus maybe in 3 years when cheaper and mated to a cheapo board.
Great video, highly educational 👍 Thank you.
Glad you enjoyed it. Have a super day
5:55 -- thanks, was wondering about power for UPI btw sockets.
I'd be more impressed if the idle power was below 20W. I have a 12-core Haswell-EP that idles at under 2W without any special setup. Sounds like modern Xeons could be a lot more aggressive with power saving.
True, got E5-2650Lv3 as my offsite backup server. Barely sips power, but man, server boards at those days really doesn't have much wiggle room for standby like S3/4. Either on or full on system halt 😅. Need those pricy workstation board for better idles power
I’ve got a 10c broadwell-ep system and cpu idle power is way over 20W. Maybe it’s a server board thing?
The CPU's tested have about 3-4x as many cores... and additional components such as additional accelerators for "AI" more PCI-e lanes just more everything. That's going to add up in terms of idle power draw.
@@jolness1 yup, like I pointed out in my first comment, in Broadwell/Haswell-E era, motherboard is also an contributing factor. Even with dual socket Asus Z10PE-D16 WS has proper idle and standby mode implementation (note that the motherboard still using server grade C612 chipset).
2W sounds like BS figures that some internal metrics report but are always far off from actual power consumption..
I'd be surprised to see any S2011 system draw less than 80W from the wall.
The lowest I've seen is 16W from the wall on C236 and that's rare AF.
I'm sad to see sub 100w server chips gone.
Stay tuned to STH next week and we will have something for you.
Inhell has to heat your home somehow ;p
@@ServeTheHomeVideo Hopefully it's something with plenty of lanes. Which is the only thing missing from consumer machines.
Efficiency. Yes, we're in a period of pizza oven levels of wattage, but with process improvements, we could swing back to sub 100w units.
@@konnorj6442hurr durr intel bad! Amirite?!?!
I can't speak for all HomeLabbers, but for me I needed PCIe Lanes more than raw compute for my HomeLab. So in January I went with an Epyc 7313P on a RomeD8-2T board with 256GB of DDR4 3200 RAM. I found the Milan chip for $470 shipped, bought the mobo new ($666), and the memory was $381. I couldn't find a better value for Milan or newer platform (Intel or AMD) so looking back I'm still pretty satisfied with my choices.
Totally. In 3 or so years when EMR is the same age as Milan is now, I would expect more value options for home labs. When that happens, a point of this was showing the difference in power consumption and performance when picking SPR or EMR
This is interesting but I think AMD is still winning here. You can get a lot more cores and a lot more cache with them and they are clocked a lot higher than what Intel is providing here. I would be interested in seeing a similar comparison on idle power draw with them as well but they do have a lower TDP and if their consumer chips are anything to go by they will likely draw less power. That being said it is also interesting that Intel is actually the "budget" option now. The closed competitor to the 8558U would probably be the 9454P and that thing cost 4,598 USD.
The Intel Xeon Platinum 8558U is 260MB of L3 cache versus 256MB on the 9454P. Emerald Rapids XCC changes the cache side a bit.
@@ServeTheHomeVideo Yes but it is only 4MB different which is negligible. Again I tried to find the closest part. If I wanted the fastest part I would have compared to the 9474F which has a base clock of 3.6Ghz or almost double that of the xeon. Comparing top to top we could have compared to the 9684X which has 2x the cores at 96 and even with all those cores the base clock of each core is still 550Mhz higher. It also has a whopping 1.1GB of L3 cache. AMDs chip even gets higher single core performance as far as I can tell thanks to the massive cache and higher clock. So in my mind the real story here isn't so much that this thing is beating AMD because it probably doesn't even manage that in the power consumption you are measuring. The story is that they are actually the budget option compared to AMD now being like a full 30% cheaper.
They are better than their last gen and I do think it is worth evaluating these processors. I just don't know if your power efficiency or even software argument hits it here because if software licensing is done on a per core basis then I probably would go with that 9474F that is quite a bit faster. If power consumption is the thing here then that 9684X is going to be the better option because it still has faster single core and still has 2x the cores. Yes the TDP is quite a bit higher at 400W but as you said that is only part of the power draw of the whole server and additionally is only part of the power draw of the whole rack. A single server with 1 9684X or the 9754 is going to consume a whole lot less power than the 2 or 3 servers you would need to build around the 8558U.
@@chaosfenixwell it is different in that the L3 on Intel is ‘more’ continuous than on that AMD processor. In the Intel processor that’s 260MB spread across two chips, or 130MB per chip, versus 256MB spread across 8? Chiplets at 32? MB each (not exactly 100% on AMD’s product line). And so you could have 128MB of data entirely in L3 cache on each chip for Intel, whereas that would be spread across 4 chips on AMD.
And I’m pretty confident that Chiplets on AMD don’t benefit as much from cache on other chiplets, otherwise AMD would would offer CPUs where they replace CPU chiplets for just the 3D Vcache chiplets (double the l3 of a cpu tile), or even double vcache chiplets (4x the L3 of a CPU tile and 32MB more than even the X3D chips have).
@@levygaming3133 Yes my point isn't that there are differences. They are different companies. I just tried to find the closest competitor in the stack and at similar as possible specs AMD is going to win overall in most areas but price. The number of workloads where that 4MB of cache or smaller cache latency is going to win in performance vs a chip that has almost 2x the clock speed has to be negligible at best. Again, it isn't going to win on performance or power here. But it is winning on price which is where I think this thing is the most interesting.
That's a great improvement from Intel for a single socket solution (while still great for AI because the included accelerators and such)!
Why didn't more people talk about this??
I do not think many have tested it
Here in Europe these efficiency gains really do matter a lot. I wonder how efficient these are compared to Milan if you're able to utilize the onboard accelerators
On the single socket SKUs I wish there was more acceleration active. Still, when those are used it is a big delta
why do they matter in Europe ? Because you live in Europe ?
@@xl000large swathes of Europe abandoned nuclear power in favor of tying their power grid to Russian oil, natural gas, and (German?) coal in the vein hope that this would reduce Russian aggression and not, as actually happened, emboldened future aggressions. As a result of recent trade embargoes on Russian energy sources, and some members of the EU starting to finally realize coal is bad, power in Europe has increased in price dramatically. I’m pretty sure I saw reports out of the UK that people’s power bills were eclipsing rents. And even before this, power in Europe was already more expensive, for one reason or another.
For example, in the US, 6 cents per kilowatt aren’t uncommon, with 13 cents being relatively high. In the EU, I’ve seen figures ranging from 21 cents to 38 cents per kilowatt.
Guys will you able to make some compatibility tests on tinyminimicro machines with 96GB RAM modules? Youve tested it with HP 600 9gen and mentioned that you also tested it with other 1L pc's but you did not mention which ones. Such compatibility list in some article would invaluable for all folks who want to buy miniPCs for their first homelab. And as we know RAM is the most important limitation for them.
We are talking about when we test with 96GB in each of the system reviews
Another great video!
Every time you hear the "S" sound, take a drink.
I think Patrick is too focused on the latest technology that he's forgotten about the very important technology called a de-esser lol
some people are physically incapable of making the S sound
I wonder how this works in other languages, like Arabic Japanese or Spanish, with a very different set of of base sounds. People in those cultures are made fun of for not being able to produce other arbitrary sounds I guess
How much of the power consumption is coming from the 12x 40mm fans, + the PSU fans though? How does this stack against 2u dual socket systems when you factor in the watts wasted on blowing air? (Also hearing damage)
A decent amount is cooling and so forth at the system level. As CPUs generate heat, fans spin up
If you don't need a lot of PCIE Ryzen cpus have been the go to for a homelab for several years now. Im using a 5900x with 128gb RAM as my homelab and it is working great with low power consumption because of slightly undervolting.
Also memory capacity and bandwidth. We pulled a ton of memory out and are using smaller 32GB DIMM but this is still a 256GB configuration.
@@ServeTheHomeVideo That is true, but if you need more than 128GB of memory you probably need more PCIe and have a larger budget also.
Been following your website and channel for many years now and I love the content. Greetings from Norway :)
Lower IDLE power makes perfect sense. It's a newer node, more modern package, and probably more advanced power management. It's maybe not something you should expect with every generational upgrade, but occasionally it does happen, especially after several generations.
SPR and EMR are on the same Intel 7 process, and the core is effectively identical. The power savings are despite more than tripling the cache size.
"Occasionally it does happen." Well . . .
Intel 10nm Enhanced SuperFin n.k.a. *Intel 7* per July 2021 company press rel. Loved that moment in Intel history ❤ I think it was called "Accelerating Process Innovation." Exactly, change name 😂
Licensing software on a per core basis is the STUPIDEST decision I have ever seen.
It gets the automatic response of, well somebody else provides something better.
When I saw it in windows server, I was like. guess I'll make it work on linux.
i think it's important to think about the average load you plan on running on your servers when considering these sorts of numbers.
obviously idle and fully loaded both need to be included in this sort of review comparison, but when you're making a purchasing decision one or the other of those numbers might have very little relevance to your future power bills
Fair, but also remember that the 8558U is beating the 8458P at 300W v 350W. So from idle to 100% load, EMR is winning.
@@ServeTheHomeVideo yea it's definitely impressive and unexpected. good scoop!
some real good info. thanks
Any time!
Which mini-itx board can I use these on?
My 4th gen Xeon is both fast and power efficient (E3-1285L v3). I could buy a Z97 board and a 5th gen Broadwell Xeon (faster and more power efficient than Haswell and with L4 cache), but that would be too expensive at this point.
i have a quick question about that idle power thing. do cpus not in idle neary use 0W. i mean the mainbord it self uses power too. i have a r630 it uses 150W idle. i know it can go low as 60W but i dont have only 1 cpu and one stick of ram and running of a usb stick ok. my system has 4 sata ssds in it and 4 2tb sas12g hdd, all connected to then perc h330 mini in two virtual drives, then i have 2 xeaon E5 2699v4 in it and 768gb of ram with all 24 slots populated, also i have the dvd drive and the display also i have the idrac8 enterprise and a 2x2 10gbit nic. so year this adds many power hungry components. 20W alone for the idrac, each ram module uses 2w to 3w each so this is another 48w to 72w, the h330 mini uses also 20w so this adds up fast.
so i think cpu idle power is nearly 0w what is shown by the system monitor. i guess this is correct. i dont kow for sure what happen if you enable dynamic power but much deeper you cant go and i thinks this impacts the performance and would lead to bad performance response and at some point it would not mater because the system wont go deeper because of no need for c states accordingly to the system utilization. if you optimize the components around and throw away everything you dont need this is possible. ass i sad if you go with the minimum requrements you can see impressive numbers but at some point this does not longer mater.
When the system is on, the CPU is providing the PCIe root, memory channels and so forth, so the idle is above 0W.
Desktop Intel CPUs can dip down into the single digits because they are fully monolithic, and have relatively few memory channels and I/O connections to keep active. These chips all have to power a lot more cores (48P instead of 8P+16E), 4x the memory channels, many time the I/O connections, and they spend some power on the connection between the CPU tiles.
You can see some of those idle power losses looking at Zen4 on desktop, where the 7950X often sits above 20W where an otherwise more hungry i9 will happy sit at 10W.
It would probably beneficial to know what C/P states were enabled and if they were the same accross testing. And how influencial they are int saving idle Power.
0:51 ha! We test in *production* they way it should(n't) be!!
"Our servers burn less power when not doing anything" isn't the win this ad wants to think it is. Maybe my business is just doing it wrong. We tend to deploy servers to do things. 😅
Huh? We showed the 300W TDP CPU out-performing the 350W TDP CPU and over 80W less at the wall. So this is idle up to 100% load. Also, we have to say sponsored even if the chips are marked as non-retail confidential. Very odd comment.
@@ServeTheHomeVideo You're right, the generational gains at load are noteworthy. I just found the heavy focus on idle usage very strange for a server bound product.
As for the rest, sponsored is sponsored. For the duration of this video you are an extension of the sponsors marketing department. It doesn't make the video "wrong" or not worthwhile, but it is what it is. It should be watched in that context.
If only I could put one of those Xeons on an affordable desktop motherboard in the future. I got addicted to cheap used Xeons. :)
Give it 2 years
Give it a few years and I'm sure there will be a plethora of cheap ail-express boards and a flood of these on the used market as Sierra Forrest and the like take over the high-performance segments.
So, I'm looking for a replacement for the E5-2630L v4 (a 2016 10c/20t processor) for my homelab. It has 40 PCIe lanes, and as far as I'm aware there's nothing like it left on the market. Neither AMD or Intel seems to make these types of processors any more, where you get low wattage (55W), but a decent amount of PCIe lanes. Anyone know upgrades?
You've pretty much fallen out of the enterprise range at that TDP unfortunately.
Something like a Ryzen 9 7900 65w, ECC support, 24 lanes of PCIe gen5, combined with an X670 board that provides an additional 12 lanes of PCIe gen4 would likely be the way to go.
Supermicro, Asrock rack etc for boards with a BMC for IPMI.
If you can't find a suitable board with IPMI, a Ryzen Pro 7945 could be ok in a consumer board combined with the AMD Pro tech for Out of Band management.
Perhaps wait for STH next week as well.
My production system has three hosts with 10 cores each (Scaleable V2), but then again I only have 5 windows servers and a few debian thrown in. Cost was the factor at the time or I would have gone with 20 cores each. I didnt think about memory channels, only 2 modules installed in each host, probably should have filled all the sockets for a little faster performance. Used a 2x64 config. so I had room to grow.
Great video! It’s a bummer (in this case) that there is no eps12v for the cpu to just put an amp clamp on. I think this is the best way of isolating given the constraints. I’ve collected data like this before and it is very time consuming. Appreciate the work you guys do,
Thanks. We had the package power in this one but I generally do not love using it. The real number that matters is at the PDU since that is what drives rack sizing and overall consumption
Single socket systems are the only ones that really make sense now that a single socket can cram all of the things into it. Sod non-uniformity, it complicates things.
Maybe the gen after sierra forest intel might be competitive again. It is amusing having to talk idle power consumption just to give them a positive spin.
Why just idle? We also showed the Platinum 8558U being faster than the 8458P at 50W lower TDP and lower total server power consumption.
@@ServeTheHomeVideo I didn't say just idle, it was just the most prominent example. The 8558U being better than the 8458P is good, but as the 8458P is from intel's "struggle to keep up by pumping power" generation, it doesn't really say much.
The e core gen after sierra forest is the one I'm looking at for intel getting back in the game, sierra forest is already a bust as the main benefit of moar cores has already been handled by the competition beating them to the punch, for sierra to be interesting it needed to be out years ago. They've done the delay tactics of scrapping avx512 and making a mess of redefining 256 bit with avx10.256, lets hope the gen after sierra has avx10.512 (I'm expecting more delays with them using avx10.256 and using misleading marketing, but who knows). If intel messes up clearwater then we are in a bad situation as AMD will be the new intel, I don't have faith in ARM competing more than in specific niches.
I know you're getting rolled in the comments, I'm not trying to pile on I just gave honest feedback about the content.
That's not what I saw in this video. The EMR CPU was faster than the SPR chip and consistently pulled less power.
Still dreaming that Intel continue making something like Haswell/Broadwell E Low power SKUs (Like E5-2650Lv3). Since atom C series back in denverton days was a thing... Welp we never gonna see a true low power enterprise CPU SKUs that still offer upgrade path for more powerful one within the same socket/platform.
Stay tuned for something next week
The increased power efficiency at idle is normal, but I think it's worth clarifying a few things for folks about why it's surprising/exciting here.
1. Generational leaps are usually pretty small nowadays. Bigger jumps in efficiency are getting more and more rare. An example on the desktop side would be AMD's socket AM3, AM4, and AM5. AM3 doesn't idle very well. Their Ryzen CPUs on AM4 began to see some really good idle optimization, with some of the 5000 series able to idle at
I thought Intel humiliated AMD over their "glued" cores
Just so you know, I was one of like 40 people in the room in Oregon during that presentation in 2017 so so remember it well. These days, chiplets are the way forward
@@ServeTheHomeVideo Chiplets are definitely the way, but we have to remember to poke fun at Intel ;)
I'm more of a user so I have a different approach to server management. Compared to an IT person. I do a cost benefit analysis for my workload. Does the cpu count / gigahertz / memory equal what performance metric for my workload. Then based on that, cost. Weirdly some cpus you expect to be faster are technically faster in the performance metric. But the cost per server is not proportional to the improved performance. So I usually go a little lower and am able to buy 2 or 3 more servers.
If I'm not focused on cost. I care about density. So whatever configuration gives me the most memory and the most compute cores.
Currently I'm running dual socket xeons. Do I care that it's a dual socket xeon? not really. But reliability is better among intel cpus verse amd. It might be another component the memory or the motherboard. But in an array of 20 the intels have no crashes while the amds have 3 crashes a week. This is a hardware level crash where the system reboots. So it's hardware related and random enough that I don't want to track it down.
What's the name of the software used for testing?
We used stress-ng
I enjoy the way that you explain the benefits of assorted server choices.
BUT when the graphics on your shirt match the background of your set & it's decorative lighting - it's just weird. I can't hear you cause it's so odd - like costumes for kindergarten or something.
Now compare that to an Epyc chip
We did the 5th Gen Xeon to AMD EPYC like 6 months ago.
I got here a 13600T can idle 50w total computer... these new xeon is crazy! more efficient than a ryzen CPU combo
Intel is completely underrated. They are a major player into the future!
soooo... 4 chiplets consume more power than 2? I can't say I'm all that surprised...
It is also four smaller consume less than two bigger. The implication is that the interconnect is using a lot of power.
@@ServeTheHomeVideo true
Home lab. $10,000 CPU? You and I live in very different homes! Oh, and I often chat with Intel about the CPU set I'm about to use ... not! This channel has gone so far upmarket I see less and less relevance.
Huh? Very odd comment. We have been doing 8x GPU servers since 2016 and most folks work in IT. This is probably down market from our average
Hi, so others seems to have notice what ARM is doing since a long time… I think it is the future, more and more calculation/speed with less power, which will be awesome for everybody.
We have been doing Arm servers since 2016 (see the main site Cavium ThunderX pieces.) Remember that the PPA gains are for the execution cores, not the SRAM caches, PCIe controllers, DRAM controllers, and so forth. Arm has been ahead in higher core count but lower performance cores optimized for cloud native workloads.
@@ServeTheHomeVideo I understand that, may be they'll take care of all peripheral functions once cores will be optimized to the limit.
ARM isn't magically lower power, if it was then AMD and Intel would be using already.
1 server electric bill must be like 1k a month to way too much energy
Not even close in the US, even in expensive power parts like California.
Assuming a server draws 500w constantly 24/7 (likely not the case if there are some idle moments - watch the video), that's about 12 kWh per day. Lets assume the worst and you live in CA, where power is around .50c per kWh. That's 6 bucks per day ($183ish per month).
You're off by a zero, possibly 2 if your power is cheap.
My server draws 650-800 24/7 😬
single core performance was self referencing among the 3 cpu and not comparable with the rest of the world CPUs. geekbench or similar real value would have been more useful. at the end you can have 48 cores but the single core is slow you need more core for the same vm on a server compared to the other. the comparison of the video was more focused on the consumption than the real cpu performance, which was fine, but impossible to compare with rest of the world servers.
Geekbench 6 is borderline useless for servers. Or to be a bit more clear, it is useless for this core count of server and higher. Over maybe 12-16 cores GB5 is usually more indicative of server performance than GB6. The vast majority of workloads, outside of things like HPC codes, on modern servers do not run on full chips. Usually they are run in containers and VMs to increase overall system load.
@@ServeTheHomeVideo I mentioned GB as example you compared the core speed among the 3 chips saying the slowest 100%, the others 120% and so on, that means nothing in real world of performance. I would like to see how the single core speed is compared to my server chip that runs esxi... I might upgrade in future...
Why not compare to amd equivalent parts with low core count and high frequency? Is this a fluff piece for intel? The intel pillows in every shot surely suggest that it is.
Huh? We explained that. There are AMD EPYC, Ampere Altra in shots as well. Strange comment.
I think that Patrick has to first evaluate the parts for what they are advertised for and suggested value to give a comprehensive picture of it's purpose. While I understand that the lack of comparison at every step is unsatisfactory for you, it does not mean it is a fluff piece. To be honest, I don't know how bad it would feel for the creator to have to deal with this much vitriol!
And also he said in the video that AMD CPUs are better when he was giving a sponsorship disclaimer (or rather the lack of sponsorship) only for you to jump to the conclusion that he must be sponsored!!!
@@ServeTheHomeVideo I have posted my thoughts of the comment you are responding to, in your defense, but I would like to say that not having the manufacturers trinkets on display would help avoid this quagmire, however unwarranted it may be!
@@ServeTheHomeVideo Strange indeed!
Found out they're junk? Well, duh!
Huh? We found big perf/ watt gains and even much lower idle power
PU series lol AMD P Intel U PU
Ha
The idle power draw is a useless measurement for this setup.
You paid a lot of money to get all these cores, which means you have a need for a lot of cores.
You will find a way to use them to their full extent, otherwise, it was a waste of money.
So that is one thought and probably more relevant for the HPC space. Many servers run at sub 10% utilization quite often is the real world.
Xeon bundle 2034 for 50$
*sigh* 2 seconds in the video, and already paused and will stop watching it.
Sss I have found far more uses for than dss and for used as well that applies 4x
Now for an amd rundown