Cloudflare switching to ARM Server, Intel "free" by Q4.

Page 2 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.
Toggle sidebar Toggle sidebar
D

Dolan

Junior Member
Dec 25, 2017
14
10
51
For cache latency, well it is what you might expect from this organization. For details you will have to look for official sources or some leaks...
ksec said:
Qualcomm already has improved Core and IPC for 7nm Centriq in the pipeline likely coming in 2019. Intel will have 10nm ( A node that should be slightly better then Samsung's 7nm ) at around the same time. So it really isn't a advantage.
Click to expand...
Well, this is pretty important. Node is already better and 7nm will be well practically well ahead of 10nm. There is no "mess in node names".
 
I

itsmydamnation

Platinum Member
Feb 6, 2011
2,764
3,131
136
Dolan said:
Well, this is pretty important. Node is already better and 7nm will be well practically well ahead of 10nm. There is no "mess in node names".
Click to expand...
What does this even mean?
Intel 10nm will beat everyone else's 7nm, by how much and in what metrics, will be an interesting argument. But it looks like its that aggressiveness in the 10nm stack ( specifically their wires ) that has created the nightmare for intel.
 
  • Like
Reactions: pcp7
D

Dolan

Junior Member
Dec 25, 2017
14
10
51
itsmydamnation said:
What does this even mean?
Intel 10nm will beat everyone else's 7nm, by how much and in what metrics, will be an interesting argument. But it looks like its that aggressiveness in the 10nm stack ( specifically their wires ) that has created the nightmare for intel.
Click to expand...
It will not beat it even in metrics they prefer. They are using metrics which is favorable for them but they never really talk about routed gated density. And there is good reason for this (same reason why every their foundry customer ran away).

To get back to topic: It is important for die sizes and efficiency... but well, Centriq already proves that 14nm is no match for 10nm so another explaining would be probably pointless.
 
I

itsmydamnation

Platinum Member
Feb 6, 2011
2,764
3,131
136
Dolan said:
It will not beat it even in metrics they prefer. They are using metrics which is favorable for them but they never really talk about routed gated density. And there is good reason for this (same reason why every their foundry customer ran away).

To get back to topic: It is important for die sizes and efficiency... but well, Centriq already proves that 14nm is no match for 10nm so another explaining would be probably pointless.
Click to expand...

Im sure you have a wide variety of workloads to backup your facts. Taking a high count of mid performance cores and then comparing to the very wide skylake-x cores and determining process performance on what is a latency tolerant, high concurrency scalar workloads* and going see SS 10nm is better then intel 14nm is ignorant at best, down right deceptive at worst (ignoring the like 3 years 14nm was out before SS 10nm). How about we take the performance of high throughput but latency sensitive very wide vector code and use that to determine which process is best, yeah sounds just as stupid as your position.

Also since when was density king, pro tip, it was only the foundries failing at the 20/16/14nm stage that saw intel take a real lead here, all the previous nodes 130,90,65,45,32 foundries had better density, they also had worse transistor performance. Also trying to link intels failing as a foundry to just one metric is again, ignorant at best , deceptive at worst. Intel have a foundry chicken and egg problem across design rules and 3rd party IP which means its hard for them to attracted clients because they just dont have the eco-system to support clients. Thats an order of magnitude bigger problem then only just losing a transistor density e-pen contest.

* falkor presentation is full of spec rate, cloudflair workload is obvious.
 
Last edited:
N

Nothingness

Platinum Member
Jul 3, 2013
2,402
733
136
Cloudflare posted another blog entry: NEON is the new black: fast JPEG optimization on ARM server.

Nice to see Centriq more than 6 times more power efficient than a Bi-Xeon, while being faster at the same time.


jpegtran-power-1.png
 
  • Like
Reactions: coercitiv
S

simboss

Member
Jan 4, 2013
47
0
66
itsmydamnation said:
* falkor presentation is full of spec rate, cloudflair workload is obvious.
Click to expand...

It was probably obvious, but until now there were not many widely publicized and detailled analysis were an Arm server was that competitive, even on workloads that were supposed to be Arm friendly.
 
K

ksec

Senior member
Mar 5, 2010
420
117
116
simboss said:
It was probably obvious, but until now there were not many widely publicized and detailled analysis were an Arm server was that competitive, even on workloads that were supposed to be Arm friendly.
Click to expand...

Well there was, just not from Qualcomm Centriq. All previous ones wont really any good at all.

teejee said:
Vlad is comparing with his own hand written assembly for x64.
Click to expand...

My bad, I thought he was using the normal libjepgtran.
 
N

Nothingness

Platinum Member
Jul 3, 2013
2,402
733
136
teejee said:
Vlad is comparing with his own hand written assembly for x64.
Click to expand...
He's comparing against his old intrinsic based x86 code.

From the comments on the blog:
That is not an apples to apples optimization. The goal was to write an optimization that would equal the *current* Intel performance, for us to make the switch to ARM. Actually the Intel performance improved slightly too during this exercise, and could probably improve further.

When the code was written we were still running Sandy Bridges (and I think Westmeres too).
There was an AVX2 optimization though: https://github.com/cloudfla..., still slower.

As for AVX512, it is extremely powerful, but as I mention in the text can't be used for secondary tasks, without hurting platform performance: https://blog.cloudflare.com...
Click to expand...
 
Midwayman

Midwayman

Diamond Member
Jan 28, 2000
5,723
325
126
senseamp said:
Intel is not having a good year:
AMD Ryzen is competitive
ARM servers at the edge
Power9 is getting traction in datacenters for AI
Apple is ditching them
RISC-V is getting interest
Start of transition of AI inference off Intel CPUs to accelerators.
Click to expand...

Maybe it means we get to see if Intel has really been sandbagging it for the last decade and dribbling out releases due to no competition?
 
moinmoin

moinmoin

Diamond Member
Jun 1, 2017
4,944
7,656
136
Thanks @Nothingness for the quote, that part's pretty damning even outside the topic:
As for AVX512, it is extremely powerful, but as I mention in the text can't be used for secondary tasks, without hurting platform performance: https://blog.cloudflare.com...
Click to expand...
Some insightful quotes behind that link. Maybe one should start a separate thread for that particular topic?
 
dave_the_nerd

dave_the_nerd

Lifer
Feb 25, 2011
16,788
1,468
126
coercitiv said:
Ask yourself this: what sane company CEO publicly burns bridges with his lifelong CPU supplier by touting their competitor's supremacy? It's one thing to (entirely) switch suppliers, another thing to post kill-a-watt pics and suggest Intel's products are more expensive even if given for free. These are 100% PR moves and I doubt Cloudflare needs the publicity.
Click to expand...

Any CEO has to "cheerlead" any decision they make; convincing employees, customers, and investors that it was a BRILLIANT idea, regardless of whether it is or not. Commit to everything 110%, at least publicly.
 
urvile

urvile

Golden Member
Aug 3, 2017
1,575
474
96
Do you understand how much money it would actually cost. For a multi national with say 100k employees to switch to a different architecture? These data centres are tiny in comparison and there has to be an ROI for them because these data centres are still a business. So there will be technical reasons and also business reasons for the switch.

That's the issue. In the datacentre/server space. Why not? Everywhere else why? What I am mainly thinking about is all of that bespoke software in enterprise and government that has been written for x86.

Also I don't have a US perspective.
 
urvile

urvile

Golden Member
Aug 3, 2017
1,575
474
96
dave_the_nerd said:
Any CEO has to "cheerlead" any decision they make; convincing employees, customers, and investors that it was a BRILLIANT idea, regardless of whether it is or not. Commit to everything 110%, at least publicly.
Click to expand...

Yep. I have seen that sort of thing in goverment. Where the "CEO" will institute a huge IT program. Collect his kudos and then leave. The one I am thinking off went over budget by 800 million bucks. Don't worry it's just tax payers money.

EDIT: Did anyone ever admit that if your budget blows out by 800 million the project was severely mismanaged and if it wasn't federal government (the money never runs out) it would have failed dismally?

Of course not it was actually a huge success!
 
Last edited:
amd6502

amd6502

Senior member
Apr 21, 2017
971
360
136
I think AMD had some acorn server parts released (A1100 series fileservers or such). Any chance they'll get some business out of this transition?
 
Yotsugi

Yotsugi

Golden Member
Oct 16, 2017
1,029
487
106
amd6502 said:
I think AMD had some acorn server parts released (A1100 series fileservers or such). Any chance they'll get some business out of this transition?
Click to expand...
No.
Why would they touch ARM, it smells bad and poking at Intel in x86 is easier and more fun.
 
  • Like
Reactions: Arachnotronic
N

Nothingness

Platinum Member
Jul 3, 2013
2,402
733
136
Arachnotronic said:
Qualcomm figured out what smarter companies like Broadcom discovered long ago.
Click to expand...
You mean that attacking the big gorilla requires piles of money and accepting losing a lot hoping to get some place under the sun?

At least it took them less time than Intel to understand they were ridiculous in the SoC businnes :D
 
J

jpiniero

Lifer
Oct 1, 2010
14,584
5,206
136
Seems hasty for QC to do this so quickly, esp now that Intel's foundries is now falling behind.

Maybe Cloudflare will stick with Centriq for now and go with ThunderX2 later; or perhaps go back to x86 with Epyc 2.
 
You must log in or register to reply here.

TRENDING THREADS

COMPANY
RESOURCES
FOLLOW
Top Bottom