Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

2022 Cloud Report

cockroachlabs.com

Post reply on HN
Print view View original

81–90 of 123 posts

Re: 2022 Cloud Report

#81

Post by coder543 »

guepe wrote: post #56
canucklady wrote: post #17

Earlier quoted context omitted.

Graviton3 on AWS is better for most applications on a cost/performance basis, and in some cases also on straight latency. Honeycomb has some great blogs about it: https://www.honeycomb.io/blog/present-future-arm-aws-gravito...

I stand corrected ! Note that is on cost/performance, not raw performance yet. Impressive.

And on raw performance: https://www.phoronix.com/scan.php?page=article&item=graviton... (direct link to the geometric mean performance of the whole test suite)

ARM isn't the optimal solution for every application at this time, but anyone who isn't seriously considering it probably needs to update their information.

Re: 2022 Cloud Report

#82

Post by erik »

babypuncher wrote: post #60
jeffbee wrote: post #55

Earlier quoted context omitted.

The memory in an M1 is not "on die" it is plain old DRAM that they buy from the Koreans and solder to the board just like Intel and AMD and everyone else. DRAM is made on a fundamentally different semiconductor process and there will never, ever be a CPU with on-die DRAM. DRAM that can be made on a CPU logic process is called eDRAM. A huge eDRAM is a few tens of megabytes, while a huge DRAM is gigabytes. The bit cell…

I think the person you are responding too used the wrong vocabulary. Apple mounts the SoC and DRAM together in a system-in-a-package design, which is pretty different from how thin & light x86 manufacturers solder DRAM chips to the mainboard. The proximity between the SoC and DRAM is part of what makes the M1s bandwidth possible.

The M1's bandwidth is possible because Apple uses high end LPDDR ram and a memory controller with a lot of channels. They aren't doing anything exotic.

Consumer PCs don't match this bandwidth because DDR DIMMs generally aren't as fast as LPDDR. Plus AMD & Intel limit their mainstream consumer CPUs to two memory channels, both for cost savings and to segment the market.

Re: 2022 Cloud Report

#83

Post by viktorcode »

DeathArrow wrote: post #54
hesdeadjim wrote: post #14

As a AMD fanboy who loves seeing them back on top, I’m just happy we have a competitive CPU market now. I’ll include the M-series from Apple as well, despite being platform locked, because it also forces the other players to up their game. I would love to see an AMD chip as fully integrated as an M1, moving the RAM fully on die and part of the Infinity fabric directly. The insane memory bandwidth of the M1 is what ke…

>As a AMD fanboy who loves seeing them back on top I am not a fanboy, but a realistic dude. AMD ruled the last years but Alder Lake overtook Vermeer on both performance and price/performance. And that is with a process node difference, Intel using 10nm vs AMD using 7nm. And the future looks like Intel will enhance the distance between its performance and AMD's.

Take a note, that node sizes of different factories aren't comparable, because they are measured quite differently. You can only compare TSMC v TSMC, Intel v Intel etc.

Re: 2022 Cloud Report

#84

Post by soco »

Queue29 wrote: post #23

This blog post is literally unreadable with animations flying around covering up content, is that by design?

This feels like back in the 90s with those GIF-laden scrolling text webpages. I can't fathom what goes into the mind of a designer creating all this nonsense for what basically is a table report. Does anyone in their right mind expect people would be more attracted by banners frolicking around?

Re: 2022 Cloud Report

#85

Post by madeofpalk »

soco wrote: post #84
Queue29 wrote: post #23

This blog post is literally unreadable with animations flying around covering up content, is that by design?

This feels like back in the 90s with those GIF-laden scrolling text webpages. I can't fathom what goes into the mind of a designer creating all this nonsense for what basically is a table report. Does anyone in their right mind expect people would be more attracted by banners frolicking around?

> I can't fathom what goes into the mind of a designer

Don't underestimate the damage an overly eager and naive developer can do.

Or, (maybe more likely in this case?) a CMS author who's been given a little bit too much power?

Re: 2022 Cloud Report

#86

Post by smarx007 »

From p. 74:

> Overall, the gap for most AMD-based processors closed almost immediately when we controlled for NUMA nodes – in other words, when we only considered runs where each instance showed all vCPUs running across a single NUMA node. When we did this, the performance gap dropped from 22% to 1%, which is smaller than our margin of error.

How does one avoid machines with vCPUs across multiple NUMA nodes? Do you just spin the machine up, run `lscpu | grep 'NUMA node(s)'` and kill the machine if the value reported is anything but 1 and try to spin a VM again?

Re: 2022 Cloud Report

#87

Post by vitus »

danuker wrote: post #47
Geonode wrote: post #37

Earlier quoted context omitted.

I don't know about competition, but I think we'll see prices bottom out below MSRP before the 4000 level cards come out in the next few months.

Do you think so, because the crypto craze is back in the crypt?

We're already close to that point. Low to mid-end AMD cards are already available at / below MSRP (e.g. 6600XT), and nVidia cards no longer command a > 2x premium across the board.

At the height of the craze, I saw a five-year-old card, the GTX 1050 (not even Ti!), going for over 200.

https://pcpartpicker.com/trends/price/video-card/

From February: https://www.techradar.com/deals/looking-for-a-gpu-dont-buy-y...

Re: 2022 Cloud Report

#88

Post by estambar »

smarx007 wrote: post #86

From p. 74: > Overall, the gap for most AMD-based processors closed almost immediately when we controlled for NUMA nodes – in other words, when we only considered runs where each instance showed all vCPUs running across a single NUMA node. When we did this, the performance gap dropped from 22% to 1%, which is smaller than our margin of error. How does one avoid machines with vCPUs across multiple NUMA nodes? Do you j…

That's what we did, yes

Re: 2022 Cloud Report

#89

Post by estambar »

_joel wrote: post #43
Thaxll wrote: post #38

Earlier quoted context omitted.

It does supports ARM. It's a Go project so it should except if they do exotic things.

Explained why they don't produce official, supported builds. https://github.com/cockroachdb/cockroach/issues/62903#issuec...

We're planning on having official ARM binaries in the fall, so I expect ARM benchmarks will return in next year's report

Re: 2022 Cloud Report

#90

Post by DeathArrow »

coder543 wrote: post #80
DeathArrow wrote: post #54

Earlier quoted context omitted.

>As a AMD fanboy who loves seeing them back on top I am not a fanboy, but a realistic dude. AMD ruled the last years but Alder Lake overtook Vermeer on both performance and price/performance. And that is with a process node difference, Intel using 10nm vs AMD using 7nm. And the future looks like Intel will enhance the distance between its performance and AMD's.

I think Intel stands to regain some lost ground over the next year or two, but Alder Lake isn't a compelling argument in a datacenter-focused discussion. Alder Lake relies on brute force, inefficient power consumption to regain the performance crown. AMD's chips are much more efficient, and efficiency matters in datacenters. There is only so much power and cooling available to each rack unit. I think Sapphire Rapids…

You are right. But parent comment was about desktops, not data centers.

And in desktops performance is what matters. And price/performance ratio, and both are in Intel's favor.

Post reply on HN