Hacker News Reader

Comment by techrat | Hacker News Reader

Intel's hubris is still getting in the way.

When companies like Cloudflare are openly announcing that none of their upcoming servers will use Intel, you have to admit something's wrong.

https://www.theregister.com/2021/09/01/cloudflare_picks_amd_...

> "We evaluated Intel's latest generation of 'Ice Lake' Xeon processors," Howells wrote. "Although Intel's chips were able to compete with AMD in terms of raw performance, the power consumption was several hundred watts higher per server – that's enormous."

Lake has turned out to be Intel's decline into Bulldozer chip style design issues. They might perform fine but they're too hot, too outdated and too expensive.

View on HN

> When companies like Cloudflare

I agree. Although I think CloudFlare's is a bit more unique than most data centre usage. CloudFlare say it's all about efficiency here: https://blog.cloudflare.com/designing-edge-servers-with-arm-...

Also note that Intel's revenue is only about 1/3 data centre (DCG in https://www.intc.com/segments - I didn't find 2021 numbers).

Every data centre is all about efficiency. Power usage is the #1 expense.

Really? Power costs vary greatly. Does your statement apply only where you live perhaps?

I'm sure regionally there are differences, but by far the most important considerations are power and cooling, both of which are a function of efficiency.

Island is quite attractive and have a few data centers (Google for sure and IIRC some crypto companies).

Servers are the #1 expense according to James Hamilton of AWS.

But power is important for sure.

> Servers are the #1 expense

If you buy top line Intel processors they are. But reliability, higher speed and lower power consumption used to make up for it.

Errr... yes and no. The Equinixes, Digital Realty’s of the world still operate DCs that have a lot of Low hanging fruit in terms of efficiency. Running your DC at 21C in a tropical or sub tropical climate is not efficient but it’s part of the SLA. Owner operators like FB and Goog run their DCs warmer. The equipment can handle higher operating temps (to a point) and the savings in IT load make up for the uptick in failure.

It’s also important to distinguish CAPEX vs OPEX.

For CloudFlare, their workload is bandwidth intensive. So it's bad battle field for Intel because AMD offers 128 lanes PCIe on single socket. It seems that even Sapphire Rapids isn't looks good. But the weakness isn't applied to all use cases.

Cloudflare is using a 25G NIC and two NVMes, so maybe 16 lanes total. I think the cores are what they want.

What they want is high throughput per watt, just as they say, hundreds of Watts more per unit that stand up is a lot at scale. Infrastructure providers are eating up business purchases, and they do care about what used to be a blip in accounting with the electricity and estate bill, now becoming a significant part of their costs which must be optimised to compete with margin. Energy efficiency is more and more important, Intel got Apple's love with their core architecture, killing Ibm/Motorola in the mobile space. I think the same fate will happen to Intel in the server space if they don't deliver an energy efficient architecture. With the headways the apple silicon has made recently, it seems improbable that Intel will catch up.

I think people know Ice Lake is going to be a flop; they are more excited about Sapphire Rapids

On paper, saphire rapids is a large step forward. Compute tiles to offer larger core counts (while maintaining a monolithic topology), 512 reorder buffer entries (up from 354), 6 decode (up from 4), 2 more execution units, chips with HBM memory...

AMX might be interesting, too, although I currently stick to `Float32`+ precision, so I don't have use at the moment for <=16-bit FLOPS.

Overall, saphire rapids looks very interesting to me. Of course, being built on an upgrade of the same node as Ice Lake means heat/energy use aren't getting better, so perf/watt gains will probably mostly have to happen on the perf side of things.

Tiger Lake clocks much higher than Ice Lake, but I'm not sure how their energy use compares at the same clock speeds. Saphire Rapids is another node upgrade on top of Tiger Lake.

Ice Lake Server is still made on the 2nd variant of the 10-nm Intel process (the 1st being that used in the pathetic Cannon Lake).

Because of that, it is much less efficient than Tiger Lake, which is made on the 3rd variant of 10-nm Intel process. Tiger Lake is still less efficient than the CPUs made with the TSMC 7-nm process, like Zen 3.

Sapphire Rapids will be made with the 4th variant of the 10-nm Intel process, the same as Alder Lake, now renamed as "Intel 7".

So Sapphire Rapids is 2 generations above the current Ice Lake Server and it should be much more efficient. We still have to see how Alder Lake will behave, but there is a chance that "Intel 7" might be more efficient than the now old 7-nm TSMC process.

However, when Sapphire Rapids will become available next year, AMD will probably soon launch Zen 4, using the 5-nm TSMC process, so it might become again better than Intel.

Reply on news.ycombinator.com