วีดีโอ

These are an AMD Milan CPU, x86-64 like many desktops/laptops, but with a lot more cores. I admit I don't know what operating systems they ship or support, but my guess is that customers will expect to run the same Linux and windows server workloads their off-the-shelf servers do. Yes, in theory anything that can run unattended is a good candidate for running on a server.

Well, that's a socket SP-3, so potentially you're looking at 64 procs right there. Socket count doesn't really matter as much as it once did.

@@JohnVance Agreed, especially given the memory bandwidth available to one socket today.

Higher DC voltage means less current, so (generally) for power delivery/connectors you want higher voltage if you can. The upper bounds on that voltage are often between (1) regulatory definitions of what "low voltage" is (this legal limit varies between 48v to 60v depending on context), (2) what your DC-DC converters can convert cheaply/efficiently, and (3) amperage/heat limits of your power connectors

@@admalledd 48v seems like the standard value for low voltage that I’ve heard of, but i dont know much about DC in the data center. Is 54v a standard value there?

Disclaimer: I work for Oxide According to UL 62368, an ES1 circuit is the lowest hazard level covered by the standard and is generally equivalent to SELV (Safety Extra Low Voltage) and the voltage limit to be classified as ES1 is 60VDC. 54.5V is approximately 9% below this limit, which allows for voltage variation due to line/load/etc and still be able to guarantee that the voltage will still be considered "low voltage". While 48V is standard in telecom infrastructure (many times -48V), but it doesn't hurt anything to go as high as possible here since most standard 48V power supplies designed for telecom aren't dense enough for our applications (our input DC-DC is 1300W continuous rated) and the power circuitry (such as hot swap controllers) work just fine at 54V instead of 48V. The higher voltage you go, the lower your resistive losses are (which go up with the square of current). Some of the more recent Open Compute (ORV3) designs use 50.5V instead of 54V, but we haven't found anything that would make us go that route and 54.5V is still a "standard" voltage for the parts we would need.

@@EricAasen-Oxide OSHA requires additional safety requirements for anything about 50v so a lot of industries just go 48v to stay under that limit. Plus it is easily split into other common voltage levels (24v, 12v, 6v, 3v, and 1.5v) which is proballly useful for sourcing parts.

@@AndrewMorris-wz1vq As I understand the OSHA requirements, those are more around guarding to prevent accidental contact. Our system is touch-safe and nothing over 12V can be reached while it is powered. The busbar is also touch safe even when something is removed from the rack for service. The limit of 60V from the UL regulation has more to do with the minimum required isolation barriers you need between hazardous voltages (such as those above the ES1 limit of 60V) and lower voltages or accessible parts (things that can be touched by a user during normal operation) that might become energized with a hazardous voltage if there was a fault condition. This is the reason there are quite a few power supply designs now that are not isolated and go from 48 or 54V down to either an intermediate bus of 6-12V or direct from 54V to the very low (0.5-1.5V) voltages a large ASIC like a processor or GPU needs in one conversion step. This is done on designs like the AI accelerators from AMD and NVIDIA because of the extreme power levels they need in exceedingly small spaces. Of course there are tradeoffs there, but if those designs had to be isolated, they would not be as dense as they are.

Agree fully. The text-heavy approach is also a nice, passive protection against "glib talkers" that ruin companies by convincing the gullible majority of their hare-brained ideas and manage to practically silence employees with actual, domain specific knowledge. It's easy to get steam-rolled (as an introvert) by glib-talkers in a meeting but let these charlatans put their arguments to paper and be astounded how little substance there is after minimal amounts of scrutiny.

Thanks! I don’t expect us to add video. We’ve been told we have great faces for radio!

@@ahl0003 😂

the sleds are taller so they can use 80mm fans, which can move more air at slower speeds and with less noise!

The number one reason for cranking up the noise of these racks is not the amount of heat, but the static pressure required by the fan to push the air out. The more restrictive the chasis, the more static pressure needed. Because they use 80mm fans (and probably have better chasis openings), the fans have static pressure to spare, so they can run the fans at a lower rpm. That’s why the outflow of air is hotter, the airspeed inside the Oxide rack is slower, so they’re “soaking” more heat into the same volume of air. As long as they keep the components at the right temperature, not only that’s perfectly fine, it’s more efficient and quiet. There are also added benefits from going to bigger fans (it’s easier to make big quiet bearings than really small ones!) but those are second order effects.

Calling Bryan a "fool who understands" nothing just because you saw a short on yt out of context... This dude has forgotten more about computers that what you'll ever known lmao

Lots of telco racks still work the same way, but for standalone servers racks it's not common anymore because of the lack of proper standardization on DC bus bars and the constant increase in power demands.

DC does not travel very good on distance of wire, big cables, lots of heat. AC is very efficient over long runs.

It's not a matter of AC or DC it's a matter at voltage, their DC part is relatively low voltage, probably either 12V directly or 48V or something. The AC can be 230V mains which results in much lower currents. When comparing 230VAC and 230VDC the DC would be minimally more efficient but 230VAC is mains voltage so you work with what you get

Because it is bullshit. Question why we use AC and not DC for our powergrids around the world. Because it is not efficient to do so. In the rack it is maybe 1m up and 1m if the PSU is placed in the middle of the rack. For a large scale operation you might be talking about hundreds of meters.

@@seynoonrae2474 Which should still be within reasonable range, iirc Edison first got problems at 1-2 km distance.

@@surters You just don't want to have to deal with it. What advantages would the additional efficency los bring to the table. More proprietary systems in case of failure? High end servers nowadays come with PSUs that have Titanium rating. The issue the guy is talking about is a non issue. The internal components make much more sense in terms of efficiency. Like the new gen5 NVMe controller from Silicon motion uses like half the power than the current Phison one does. On a hundred SSDs that means about 300W less power draw for the rack and 300W less cooling. Wanna know what they are currently imagining, removing the step down stages in the PSUs and only have them deliver 12V and anything that needs less power either gets it from the motherboards step down stages or has their own internal step down. All to increase efficiency by reducing the distance from the DC source to the consuming part. Yet these guys and you proposing the opposite. Sure it might not be a lot of loss under normal circumstances but you are talking about installations that might be in the six to seven digit area of power draw. Let us saw you are looking at 250kW, at 48V you are looking at above 5200A of current, at 12V that is an insane 20800A of current. Compare that to the 1090A for 230V and the meager 625A for 400V. I dunno maybe you wanna try to weld like 2 tanks together to get some command and conquer type of monstrosity. At those currents you will end up with less efficiency and that is the opposite of what they want. They recently just decommissioned a super computer that was not even that old because with modern hardware you cold achieve just about the same performance at a fraction of the power draw.

It's a hardly a new way of doing things, blade servers have been like this for years You have multiple ac-dc sources on the bus

Saves me typing it ❤

They must be having redundant power supplies. Same as in AC. Which usually have two power supplies per server.

We offer 5+1 or 3+3 redundancy for the 6 rectifiers in the power shelf, and the option for two power shelves.