Comment by smolder | Hacker News Reader

Comment by smolder | Hacker News Reader

Parent Full thread

Interestingly, computer chips can often be run at lower voltage and wattage for a given frequency if they are kept at a colder temperature. As a home user I can significantly reduce power draw for a CPU/GPU by improving the cooling solution and lowering voltages.

The reasons this doesn't work for datacenters is two-fold, I think: First, they won't see efficiency improvements just by keeping their CPUs and GPUs (or RAM) cooler because the power levels/tables for the chips are baked-in, and operators aren't going to the trouble of tweaking voltages themselves. Second, even if they did tweak voltages, the cost of sustaining lower temperatures with better cooling likely won't outweigh the savings resulting from lower power draw for the chips.

Still, this raises the question of whether designing hardware for higher operating temperatures is always the right move. At some point there's going to be a cost in performance and/or efficiency that outweighs the savings from allowing higher temperatures. Ideally these tradeoffs should be balanced as a whole.

I think you missed the biggest reason this doesn't do much for servers - they already run at low frequencies & voltages.

For example take the Epyc 7742, at 225W it sounds super power hungry. But the 64-core chip only boosts to 3.4ghz max (2.25ghz base). That's less than the base clock of almost any of the Ryzen consumer CPUs. And if you look at the lower frequency end of https://images.anandtech.com/doci/16214/PerCore-1-5950X.png there's not a whole heck of a lot of efficiency gains likely to be had below that ~3-3.4ghz mark. They're already basically sipping power at something like 3w per CPU core or less. 225w / 64c = 3.5w/c, but the IO uncore isn't exactly cheap to run and iirc sits more like in the 50-70w range. So subtract that out and you're at more like 2.5-2.7w/c. I don't think throwing cooling at this is really going to get you much of a gain.

This usually isn't true at scale though. The chip manufacturers do a ton of validation and qualification to set the operating parameters (voltage, etc).

You can undervolt (or overclock) one specific chip, individuals have been doing this at home for basically ever, but there's (almost) always a system-specific validation process you then do to make sure the system is stable for a specific workload with the new parameters.

And these parameters differ between batches of chips, or even between chips within a batch.

It's also significantly harder to drastically reduce the temperature of the chips inside of a single server, given the machine density of a typical data center.

> It's also significantly harder to drastically reduce the temperature of the chips inside of a single server, given the machine density of a typical data center.

It’d be fun, however, if we could dynamically adjust that according to workload. If workload is light, you could consolidate load into fewer sockets/memory sticks and power down everything in that socket.

It’s been a while since I was doing hardware in a DC, but even dynamic frequency changes in servers were noticeable in workloads.

We’d see serious latency spikes with high volume increases. And high volume would correlate to high I/O on shared storage exacerbating the problem. You’d get a bunch of transactional failures from the spontaneous latency hit.

Same was true with older virtualization/partitioning of hardware. You’d even see apps back up into memory and swap themselves to death.

Now, the apps were awful: inefficient and intolerant. But that’s the reality for the vast majority of business software. Churned out by devs with little understanding and tight timelines for features.

Reliability being an NFR kills me.

It’s an interesting thought. I think in the datacenter today it makes more sense to evacuate and shutdown servers vs. tune down the devices.

Reply on news.ycombinator.com