Marvell Technology to buy chipmaker Cavium for about $6B | Hacker News Reader

Marvell Technology to buy chipmaker Cavium for about $6B

arstechnica.com

59 pointsrbanffy20 comments

The deployment of ARM in the data center has been more words/vapor than reality for years. And if you try one of Scaleway’s baremetal ARM offerings, you realize that using Intel CPUs is similar to using an SSD. Once you use it, it’s difficult to use something far slower. I’m not holding my breath on ARM in the Server, and I think we will see ARM desktops become widespread much sooner than ARM servers.

All that being said, I really want ARM servers to happen. One because a lifetime of x86 monoculture makes me curious about other platforms, and two because like AMD when they’re pushing hard, it’ll drive Intel out of their slumber. Intel’s major advancements come when they’re at least challenged if not threatened. Otherwise they do the equivalent of the old +25Mhz SKUs.

Still if I could have one wish, and it had to be CPU related, I would prefer to see an open ISA like RISC-V delivered by multiple vendors competing become widespread.

Scaleway uses the ARM cores of 2011. The cutting edge ARM server cores from Qualcomm and Cavium are a little faster than that.

https://blog.cloudflare.com/arm-takes-wing/

These were impressive results from Qualcomm and Cloudflare. It seems like a serious threat to Intel.

Cavium not so much for some reason, not sure why their numbers are so much worse that Qualcomm.

That was the (somewhat old now) Cavium ThunderX. Cavium are prototyping ThunderX2 which has better performance.

One great hurdle a new platform (or an old and now obscure) is credibility. Would you bet your success on POWER or SPARC? It's been years since I used either, to the point I'm much more comfortable predicting performance on EC2 instances than on actual metal for these.

If I were trying to pull this off, I'd try to market units with failed cores deactivated at lower price points and help a workstation maker to the market with low priced units. If a high end fully functional chip performs like a Xeon E7, maybe they could have a couple units priced at Core i3 prices with similar performance. The idea would put these machines in the hands of key developers so they would help legitimize the platform.

Qualcomm Centriq / Amberwing absolutely screams. Totally comparable to Intel Xeons.

Keep an eye on power. ARM, in general, has a useful-work-per-watt number that is very hard for x86 to hit. If useful-work-per-rack-slot is equal, and the power is lower, ARM has a good way to attack the rack.

I think ARM is too broad to make such a claim at this point. The scaleway ARM servers are slow, but because they are using an older implementation. The A11 is proof ARM can be fast... fast enough for a server, at least.

Don't desktops make less sense than servers for ARM in terms of per core performance? Over my dead body would I use a 2.1-2.4ghz desktop chip but that seems to be what most cloud VPSs are around with ~3 for "high performance".

Really curious to see what happens in the next 24-36 months with 7nm nodes from non-intel fabs coming online. Combined with huge investment from China into nand/dram I think were in for some pretty exciting price/performance options in the next couple years.

It depends on what you need the server for. For many applications, including web, it's all about efficiency, not speed. And ARM can be quite efficient in both performance per dollar and per watt.

Maybe if you only look at the core cpu. Amazon bought Annapurna labs, an ARM chip maker, just for their data centers so they definitely are making progress.

The performance has nothing to do with ISA and everything to do with caching - both size of caches and algorithms used. [1]

[1]: https://www.nextplatform.com/2016/05/31/intel-lines-thunderx...

Am I getting downvoted for stating a fact. Is there someone here who believes Intel's ISA to be superior? Please cite the science behind such a claim

I wonder how this will affect Cavium's ThunderX line. When Applied Micro was bought by Macom, their X-Gene line was more or less trashed. If this keeps happening to companies trying to develop ARM servers, you'll never see an ARM server in a datacenter.

ThunderX 2 seems to be quite competitive already, even though it's more than a year old at this point:

https://www.nextplatform.com/2017/11/13/cray-arms-highest-en...

I would imagine ThunderX was the primary reason why Marvell bought Cavium, but I'm not sure how committed and competent Marvell is with ARM chips these days, so hopefully this acquisition isn't for the worst.

Remember though that X-Gene wasn't very good when examined. And the current ThunderX is actually the continuation of the very interesting (and canceled by) Broadcom Vulcan, not the mediocre original.

Oh I'm well aware -- the lab I work in bought a couple of X-Gene 1s for testing, and even an X-Gene 2. They were complete and utter crap performance- and power-wise -- but they were (at the time) the only thing available.

Didn't know that about the ThunderX -- fascinating! Thanks!

I wouldn't expect much from ThunderX(1). Qualcomm Amberwing or Cavium ThunderX2 are where it's at for Xeon-like performance.

Remarkable the consolidation going on in general, but silicon is really seeing a wave of (possible) mergers and acquisitions.

Why is the image a picture of a D-Wave chip?

Reply on news.ycombinator.com