Hacker News Reader

Comment by sgillen | Hacker News Reader

I'm having a hard time figuring out who would be buying this. Only thing I can think of is that the chip has some extra features not really advertised that the customer needed.

View on HN

The finance industry have some really whacky requirements. HFT firms often use heavily overclocked servers with HEDT processors, because microseconds of latency on a server request have direct monetary implications. I can't think of an application that would justify paying $5000 for an extra 300MHz of base clock, but I can imagine that someone needs it.

If your DC had limited rack space it might start to make sense. Buying a cheaper processor means you would need more of them to accomplish the same amount of work, but also you need more space.

>If your DC had limited rack space it might start to make sense

Not just rack space. If workloads have to be on the same machine or socket (not NUMA/RDMA friendly), your only choice is to get bigger and bigger processors.

$5k is barely a 1 week fully loaded cost for an experienced SSE.

If I can spend $5k to save a week of optimization work, it's a good deal. Also buying clockspeed can't create new bugs either, which is a risk with any code change.

In my experience, most companies don't think this way, but definitely should think more like this.

That said, companies that need more speed often don't need that extra speed on just one node. So that's really $5K x N nodes. When you consider that just the CPU costs $15K, and not the complete cost of a system, I think it's a lot harder to justify. You better have that next lower tier $10K CPU and know that you need just that little bit more. And then... you probably ought to be investing in scaling out horizontally while that $5K x N buys you time.

> Also buying clockspeed can't create new bugs

Famous last words

> Also clockspeed buying can't new create bugs.

> buying clockspeed can't create new bugs either

Some software does not like high clockspeeds. For example, game engine in GTA V starts acting up when framerate gets too high.

It’s similar to why you used buy RISC vs Intel. Sometimes decisions like this are driven by by software licensing. $10-20k on a server may offset $200k in Oracle.

If you look at a longer time horizon, there are other dynamics. Your contract may not allow you to reduce license spend in three years, but you can shrink your hardware spend.

There's some stuff that isn't horizontally scalable either intrinsically, or the cost benefit hasn't gotten there yet. Back in the day I heard rumors.that Borg scheduler nodes needed a lot of cores on the local box _and_ a high per thread IPC. I could see Google happily spending $5k for 300 more MHz.

There have got to be customers out there for whom 10%+ base frequency performance is worth well north of $5K.

No reason not to capture margin for a handful of top-binned chips.

There's plenty of times that extra compute capacity saves money, but not spending $5K to get a relatively small bump in clock on one chip. $5K will get you way more compute power if you just spend it another machine.

If you pay licences per machine/cpu socket/cpu core, those $5k might be your cheapest option.

Sure, if what you’re doing is easily scalable to another machine.

There’s a lot of software out there, legacy and otherwise, where the throughput ceiling is what you can do with one box.

Just the mere fact that such a processor exists,is not in the official catalog and costs substantially more, tells me that there's adequate demand for these or at least the particular customer has very deep pockets if they can ask for custom build.

Reminds me of the mystery largest buyer of glitter: https://www.nytimes.com/2018/12/21/style/glitter-factory.htm...

Here are some good comments and discussion about the possible industries.

https://www.reddit.com/r/UnresolvedMysteries/comments/aop378...

tldr; Most likely food, makeup, toothpaste etc. Especially food safe and biodegradable glitters. They are also located close geographically (New Jersey) to those industries.

> I can't think of an application that would justify paying $5000 for an extra 300MHz of base clock, but I can imagine that someone needs it.

Here's one: EVE Online servers. Back in 2013, CCP had at least one 4.4GHz Xeon they used to host fleet fights with thousands of players:

https://forums-archive.eveonline.com/message/3917593/#post39...

Why? Because EVE is mostly written in Python, which is both very slow and, thanks to the GIL, effectively single-threaded. It's a fundamentally limiting architecture, but that extra single-core performance let them push it just a little further – which for them was (is?) well worth it.

maybe someone who runs rendering farm with CPU utilization is always in the 99%? baffles me too.

Render farms scale horizontally, you can just buy more CPUs.

Reply on news.ycombinator.com