>Compared 5 IBM Machine Type 3931 Max 125 model consists of three CPC drawers containing 125 configurable cores (CPs, zIIPs, or IFLs) and two I/O drawers to support both network and external storage versus 192 x86 systems with a total of 10364 cores. IBM Machine Type 3931 power consumption was based on inputs to the IBM Machine Type 3931 IBM Power Estimation Tool for a memo configuration. x86 power consumption was based on March 2022 IDC QPI power values for 7 Cascade Lake and 5 Ice Lake server models, with 32 to 112 cores per server. All compared x86 servers were 2 or 4 socket servers. IBM Z and x86 are running 24x7x365 with production and non-production workloads. Savings assumes a Power Usage Effectiveness (PUE) ratio of 1.57 to calculate additional power for data center cooling. PUE is based on Uptime Institute 2021 Global Data Center Survey (https://uptimeinstitute.com/about-ui/press-releases/uptime-i...). CO2e and other equivalencies that are based on the EPA GHG calculator (https://www.epa.gov/energy/greenhouse-gas-equivalencies-calc...) use U.S. National weighted averages. Results may vary based on client-specific usage and location.
Wondering how a dual socket AMD EPYC™ 9654 system with 2x96 cores would compare. It's not clear how many 14nm+ Cascade Lake machines were used and it's also not clear how many 32core machines were used. 10364cores/192servers averages about 53 cores/server.
It's apple and oranges anyway because I guess if you consider this you have different priorities but I'm skeptical that x86 couldn't come close in terms of performance/watt
It is certain that the new 5-nm AMD Genoa CPUs have a better energy efficiency than the 7-nm IBM CPUs, for the applications that are executed on the CPU.
Nevertheless, for certain workloads this IBM server might have a better energy efficiency, when they can exploit some of the many accelerators that exist in the IBM server, for offloading some tasks from the CPU.
Intel pursues now the same strategy, because the soon-to-be-introduced Sapphire Rapids cannot beat AMD Genoa in a CPU vs. CPU competition, Intel stresses now the advantages offered by the accelerators that they have included in Sapphire Rapids, following the traditional IBM model.
Nothing new here; domain-specific accelerators are a natural response to "dark silicon", which is increasingly relevant at leading-edge process nodes given the failure of Dennard scaling. I.e. you've got space for a huge amount of transistors on your silicon, but only a tiny fraction of them can be powered on at any given time.
My dad the data processor told a story decades ago about how when they swapped out the mini computers at work for new machines, they started getting condensation in the server rooms because the AC tonnage so outclassed the BTUs of the servers that the temperature would drop like a rock when the units cycled on.
"Reduce energy consumption by 75%" sounds like exactly that sort of, "Remind the customer to call their other vendors to schedule an appointment" situation.
Don't think I've ever thought too hard about it. Thinking 'out loud': If you heat or cool a whole room by changing out the air in it, then the humidity or dew point should be whatever was in the heat pump, so you might get condensation on glass walls between rooms but you shouldn't get any in the room.
But in general you're pumping really cold or really hot air into a room and doing only a partial air change and the blower is supposed to stir the room up to even out the temperature. I had to look it up because I wasn't sure, but the amount of humidity air can hold is in fact a curve, not a flat line. So if you mix one unit of very cold air (below the dew point) with four units of ambient air (also below the dew point), it does seem like the dew point of the resulting air might be higher than the temperature.
Some server rooms get too dry and moisture has to be added back in to prevent static discharge damaging electronics. That might be the sort of thing he wouldn't have been aware of but was part of the equation.
There's also the possibility that he was fucking with me.
You don't generally exchange air with the heatpump, you're blowing the conditioned air through a radiator and the heat transfers from the air to the coolant which evaporates; the coolant is piped (pumped) to the outside equipment where there's a condensor and another radiator where the heat is transfered to the outside air.
Anyway, if everything is working properly, you more or less get a fixed delta T between the inlet and the outlet of the air handler while the system is running.
If you eliminate a lot of the heat load, your inlet temperature might not get as hot, also even if it does get as hot when the A/C is off, when it turns on, it may cool down much quicker and then you may overcool; especially if the A/C has a minimum cycle time.
As you say, when you cool air, the humidity capacity drops. Most air handlers have some capacity to remove humidity, but it's no absolute. If you drop the temperature too much, relative humidity will be really high and it will condense on all sorts of surfaces.
The AC unit pulls the humidity down as it cools (actually it _has_ to remove the moisture first). Which means that _all_ the excess humidity/condensate goes down the AC drain from the evaporator which will be 20-40F (more or less, depending on system/etc, it might be more) cooler than the inlet. If you live say in the southern US and have ever had to use aux cooling in a datacenter you will be shocked at how much you have to empty the condensate collectors. Even on fairly small units you can fill up 55 gallon trash cans in a couple hours.
AKA even if its way overspeced you won't get a relative humidity > 100% which is what is required for everything in the room to start condensing.
So, if the outlet is really cold, and the humidity/room temp is high you can actually get a small amount of visible condensation as the cold air blows into a hot humid room, but I've never see that stuff manage to get more than a foot or two before it turns back into vapor. Basically a tiny cloud. Which would be how you _might_ get some condensate on something.. If your blowing the cold air against something (or its discharging up through a raised floor vent into a machine cabinet) and the humidity is high in the room it might be possible, but this is why you either mix the room/ac air first, or you seal the cold isle/machine cabinet/floor to avoid cold->hot leaks. If the AC is flowing from the cool/dry side through the machine the machine even if its not heating the air much will just raise the temp, not the humidity. Which lowers the RH in the room.
Also, the condensate on a machine room window, will be on the _outside_ where its warmer/wetter because the window is _cooler_ than the outside ambient.
Also, unless there is a huge air leak in the room, the machines will act like a heater, and the room will tend to have a fairly low relative humidity. AKA the AC will pull out the moisture for a given temp, and the machines/walls raise the temp of that air, so you have air with low relative humidity.
So, the only way the room starts to condense is if a bunch of external warm+humid air is added to a room, but that can happen independent of how much over/under cooling is happening. Running a perfectly setup room at say 70F, and throwing a the doors open to 100F 95% RH air and replacing the air faster than everything in the room can warm up will condense.
I'm on the west coast. Now in the Seattle area with A/C, and the air handler in the basement (which conditions the basement and the first floor) will form a fair amount of condensate where it's supposed to, but it can also form on the outside of the duct work too... Although the basement isn't super air sealed. So I'm just saying most of the condensation happens where it should, but not all of it. Of course, my basement isn't a machine room, either. I've only got a couple computers down there, and they're low power.
This stuff is run through attics and the like all over the southern US where the attics are at ambient humidity levels, and I can't remember the last time I saw condensate forming on one that was properly sealed with mastic. In fact it would probably be a problem if it was because it would be dripping through the attics into the drywall ceilings. R8 doesn't sound like much, but its apparently enough to keep the outside of the duct from feeling cool at all.
Basement is conditioned, and the air handler and the duct with condensation are in the conditioned space. Most of the duct work is out of sight so I can't say if it's got insulation, but right by the air handler, it's just sheet metal, no insulation. It's not properly designed, or built... but it's there. Mostly it works ok, just gotta have a separate dehumidifier near the air handler to keep things under control... fixing the zoning helped a bit too.
And it may have been seen when people walk into the room. IE open the door, let in warmer air. Similar to a walk in refrigerator.
> There's also the possibility that he was fucking with me.
That too. I've seen parents do this over random things. I hope not to (I don't want to betray trust) but I may be tempted one day if my kid is being willfully ignorant over something that matters.
> Starting at $135,000 USD*, the single-configuration LinuxONE III Express is designed as a new cost-effective offering for the most popular workloads, including data serving and Red Hat® OpenShift®. IBM internal tests show that a 12-core IBM LinuxONE III Express server running WebSphere and Db2 workloads requires 18.7 times fewer cores when compared to x86 servers and delivers a 57% lower TCO over three years.
They go one to explain their benchmark in the footnotes. I don't know how to translate from IBM units to something that an organization with a mortal-sized checkbook would already be using, though.
Its worse because as I read it the base system isn't the config they are actually comparing. So based on personal experience with low end z/series machines in the past, basically the base price quoted ends up doubling by the time disk/software/etc all get added. Then what you have is a machine which is intended for DR situations, so the CPU "capacity" is limited to the point where its enough to boot the machine and little else. Then in order to match the named perf, what happens is that you cycle to a higher capacity tier (for $$$$$$$ of course).
These marketing materials basically need an IBM decoder ring, and its on purpose because if they just published the TPC-C/$ numbers everyone would be able to compare the machine vs the competition, and it likely wouldn't look good. So the benchmarks are contrived the same way apple was contriving altivec/photoshop benchmarks in the early 2000's. Sure with this one benchmark the machine is 10x faster than the competition but for these other 99 benchmarks its 1/2 the speed kinds of thing.
I'm sure buried in the IBM license there remains a restriction on publishing benchmarks.
Which maybe at some point I will post my own personal results running on an older machine when I don't think IBM would be able to trace me... but that said, lets say, in the past these numbers were cooked in such a way that the absolute top of the line machines were roughly equivilant to a mid-range x86 server available for about $40k. Now that said, if you can use all 40TB of redundant ram, and your workload needs 100 PCIe slots that it can only drive at about 1/100th of the total bandwidth (aka you need them for connectivity with the ability to burst at line rate) then sure, the zseries can come out ahead. But frankly your probably doing it wrong if you plug in 20 100Gbit adapters that average 10Gbit each.
I'm guessing a large part of the posted efficiency of these machines comes from the ML accelerator they now have, which gives them the ability to optimize some ML+DB type of workload that requires a big fat nvidia GPU under normal circumstances but doesn't make good use of it (or they are just comparing it to a x86 without a GPU...).
If you run Oracle, you can use the IBM partitioning to limit your per CPU core assignment. That’s typically the business case for this sort of thing or buying whatever hollowed out husk of Sun that Oracle is selling.
If you’re the CIO of an insurance company or whatever, you roll something like this into your mainframe renewal.
It's not immediately clear either way, but I'm guessing this is based on composable / software defined infrastructure components, like e.g. CXL memory.
If its from IBM Telum processors, this is the weird processor where a processor's L2 cache serves as the L3 and L4 caches of other processors. The processors network with each other, notice when they're not using all of their caches, and shares that cache with things far away (ie: in side the box, or even outside the box).
This way, IBM can overprovision L2 cache and "not worry" about it going unused. Remote L2 cache is slower than local L2 cache, but its still SRAM and should still be faster than DRAM to access.
If you buy an IBM mainframe you'll have a very attentive full-time account manager for life. I did a ride-along with such an account manager once when I was working at IBM and it was eye-opening.
When the cloud was first really taking off, I had a friend who swore by his mainframes, touting high reliability, great performance, excellent customer service. It would be interesting to do a cost comparison.
I'm going to channel the efficient market hypothesis (I apologize) here and say that people have done that cost comparison and not one of them bought mainframes. Heck, Google ported their entire monorepo to Power but not to z.
Lots and lots of banks (and non banks) rely on IBM mainframe still and will for years to come. They have done this calculation on whether to replatform and migrate several times. It’s very hard to stand up a positive npv for such an idea unfortunately. I guess greenfield is very different.
The thing is, there aren't really any successful greenfield developments in these types of business spaces. So either these are industries that are exceptionally difficult to break into (which I think is at least partially true), or the type of computing they do actually benefits from being done on a mainframe (which is also true given that there is a lot of record processing for these industries, and not too much serving lots of web requests to users).
Of course, all of these companies also have x86 servers that run the customer facing web apps and stuff. You really wouldn't want to waste your MSUs running web servers, although I'll admit I don't know if these new Linux mainframes are charged by rolling avg MSUs or other MSU metrics. Traditionally, Linux on Z has had more favorable pricing due to the ability to use IFLs and basically just declare that a cores MSU usage doesn't count towards your licensing since it's been locked to only running Linux.
Yah, given my experience, you buy the mainframe for one of the native zos/etc flavors, and then the IBM sales people basically throw in the IFL's which will never get used for the zos/etc workloads effectively for such a nominal price, they might as well be free.
Then you have half a mainframe of capacity lying about, and why not run some linux workloads on it.
I find it _really_ hard to believe that outside of a few really narrow niches anyone is really buying mainframes to run linux workloads. Its just not cost effective.
PS: and then of course IBM gets to add it to their total "sold capacity" for the quarter to make it look like the business is growing. So they _want_ you to use all that idle hardware.
> So either these are industries that are exceptionally difficult to break into
AIUI, Oxide Computer is targeting the "private cloud" space which is sort of adjacent to this stuff and could easily intrude on it. Of course they're based on AMD EPYC hardware, but the rest of the system is custom designed for ease of management and reliability.
I havent heard of a successfull[1] greeenfield for an enterprise ERP system replacement, especially not one done using off the shelf technologies.
The thing that recent grads(weather engineering or business) tend to miss here is that for a lot of those organisation the central mainframe system and the organization's policies and culture is pretty entangled to the point where you cannot just replace either.
[1] on time, on budget and without leaving some aspects of the old system running under some sort of virtualization.
> I'm going to channel the efficient market hypothesis (I apologize) here and say that people have done that cost comparison and not one of them bought mainframes. Heck, Google ported their entire monorepo to Power but not to z.
Nah, there are a ton of businesses of all sizes across the world with many man-years of effort put into writing software that runs on various layers of the IBM stack from Cobol on up.
For one regional grocery store I know of they started in the 70s. Let's average 10 developers on staff churning out 4GL, Cobol, and whatever else since 1979. I hear it runs webservices now to serve the desktop and mobile client apps.
10 engineers * 43 years = 430 man-years put into a pile of software highly customized to exactly how they do business, from corporate & accounting to the warehouses to the ice cream factory to the industrial-scale bakery to the retail stores. Every single process at the company driven by mountains of custom software. The thing that knows more people buy Texas Toast at Store #473 so it prompts the manager to order extra and the bakery to put an extra Texas Toast run into the already tight schedule? Yeah that runs on the IBM mainframe... because that's where the cash register sales data goes and where the managers handle orders and where warehouse tickets are printed and where the bakery solves the traveling salesman problem to create schedules around quickly expiring products (did I mention the delivery truck schedules also run on the mainframe?).
If we assume it would "only" take 10% of the effort to replace, mythical man-month is not a problem, and you can handle dual-running the integrated systems for the decade it will take to perform the swap... you saved a year of your on-staff developer's salaries while putting every part of your business at risk of exploding (and ceasing all improvements to the existing stack lest you chase a moving target)... and you end up with something more or less equivalent except on commodity hardware.
$1m for IBM hardware? $5m? $10m? Pfffffffttttt... Chump change. Where do I sign the check? Sold and sold. Don't even bother sending a rep out.
Pff it's IBM you buy the Consultant with it, he sits in the lower rack and has spacefood rations for 5 years, then you have to do maintenance and the consultant get exchanged and retired.
Nonsense, there would be at least one more redundant consultant if not two.
Seriously though, these machines call home when there's a hardware failure, they'll fail over and someone will show up to replace what failed. Enterprise tech is sometimes cool.
>Nonsense, there would be at least one more redundant consultant if not two.
Gosh you are right i always forget him.
>Seriously though, these machines call home when there's a hardware failure,
HP's EVA Device did that too, so sometimes a Engineer calls you and ask's if "we" have time to update the firmware, not even HP engineers are allowed to do EVA updates...just specialized EVA ones.
This is pretty standard across much of the enterprise world. The sales team just has to earn their commission.
Indeed, even when they have publicly listed prices, it's usually possible to negotiate a substantial discount by dealing directly with the sales team. Like 50% is often possible.
Yeah, I know. Just poking fun of their level of hype being such that they make it sound like you are one of a privileged few if they even take your call and allow you to buy this apparent masterpiece.
The general rule of thumb is if the product page says "contact us" instead of a price, you can't afford it.
It is a shame though. IBM has no idea how to sell to anyone other that gigantic megacorps - which, fair enough, but they could probably pick up a lot of sales to medium-sized businesses if they weren't so old-fashioned about everything
Medium-sized businesses can easily buy IBM stuff through VARs. (If your VAR does not recommend any IBM products, that's a different situation.) If you're smart enough to buy IT directly then you're probably also smart enough to realize that very few IBM products are good.
Our IBM was the ONLY thing that didn't get vaporized when we got ransomwared. Tens of linux boxes, MS servers, 3rd party who knows what the f they are, all gone. Poof.
To be honest you should not be buying this level of equipment "off the peg" without a serious pre-sales process of evaluation and sizing unless you have significant prior experience, and the people with that experience will already have some kind partner agreement in place.
It's much more like placing an order for an container full of severs based on an reference design from an OEM then going to HPE or Dell for a commodity x64 system.
>Compared 5 IBM Machine Type 3931 Max 125 model consists of three CPC drawers containing 125 configurable cores (CPs, zIIPs, or IFLs) and two I/O drawers to support both network and external storage versus 192 x86 systems with a total of 10364 cores. IBM Machine Type 3931 power consumption was based on inputs to the IBM Machine Type 3931 IBM Power Estimation Tool for a memo configuration. x86 power consumption was based on March 2022 IDC QPI power values for 7 Cascade Lake and 5 Ice Lake server models, with 32 to 112 cores per server. All compared x86 servers were 2 or 4 socket servers. IBM Z and x86 are running 24x7x365 with production and non-production workloads. Savings assumes a Power Usage Effectiveness (PUE) ratio of 1.57 to calculate additional power for data center cooling. PUE is based on Uptime Institute 2021 Global Data Center Survey (https://uptimeinstitute.com/about-ui/press-releases/uptime-i...). CO2e and other equivalencies that are based on the EPA GHG calculator (https://www.epa.gov/energy/greenhouse-gas-equivalencies-calc...) use U.S. National weighted averages. Results may vary based on client-specific usage and location.