REGISTER | LOGIN
Breaking News
News & Analysis

HP Ships First ARM Servers

Applied X-Gene, TI Keystone are first SoCs
9/29/2014 07:45 AM EDT
27 comments
NO RATINGS
Page 1 / 2 Next >
More Related Links
View Comments: Newest First | Oldest First | Threaded View
Page 1 / 3   >   >>
krisi
User Rank
Author
Re: What additional benefit as compared to x86 Servers
krisi   10/1/2014 12:23:45 PM
NO RATINGS
Broadcom in servers? strange

eetex
User Rank
Author
Can we please stop getting excited about ARM-based servers!
eetex   10/1/2014 4:30:14 AM
NO RATINGS
Sorry about the length of this...

When I was architecting SoCs for a sizable chip company, we spent some time benchmarking a variety of CPU cores to see if, for a given level of performance, there was any benefit to one core over another.

No one should be surprised when I say there wasn't.  Ditto if the figure benchmarked was power.  The only real factor was cost, which essentially means in-house developed versus licensed.

I kind of crystalized this thought into a law of the semiconductor industry that I'm surprised no one has previously cooked up (maybe someone has, and I've just never seen it).  Simply stated, it is:

For (1) a given process technology with (2) a similar allocation of logic gates to ALU functionality and (3) SRAMs, performance of any given CPU architecture of identical bit width will be comparable on any chosen benchmark, e.g say within 5-10%.

(1), (2), and (3) essentially mean your CPU core complex occupies nearly the same die area regardless of ISA, which is what we saw regardless of the CPU we chose, and this results in comparable benchmark performance, e.g. CPU X might be 10% faster than CPU Y at benchmark A but the reverse might be true for benchmark B.

The same applies as far as power is concerned such that when averaged across an entire benchmark suite all CPUs were within single digit % of each other.  Consequently, the build (or use what you had) versus buy decision regarding the CPU core for a given SoC could never be based on performance or power.  It had to be based on other, usually business-related, reasons.

I believe this holds true because most companies are manufacturing devices in foundries, and, consequently, have access to the same process technology capabilities.  No one has any special tricks that they can use that their competitor doesn't also have.  Multi-Vt transistors, low-power SRAMs, clock-gating, asynchronous logic, whatever design technique you name, everyone uses them.

Save for Intel, of course, who as the process technology R&D behemoth, tends to be a generation ahead of everyone else.

What does this all have to do with ARM servers?  Simply put, no ARM server processor is ever going to offer anything better than a comparable Intel server processor.

Why?  Because in comparable process technologies, we should expect an ARM server CPU to allocate a similar percentage of the CPU core complex die area to logic and SRAMs as would an Intel server CPU intended to hit the same price point.  Both will offer similar levels of performance and power.

This means that the whole Intel vs. ARM server war exists as a construction of a compliant trade press aided and abetted by technology and stock analysts who command big bucks for their "reports."

There is, arguably, some reason for this to get exploited, too.  The current model for web-based development works rather nicely on lower-cost, lower-power processors with good single-threaded integer performance.  Until recently, Intel hadn't exactly been addressing this market, so this gave rise to ARM (or really, any other 32/64-bit embedded CPU architecture, e.g. SPARC or MIPS) being a possible alternative in order to reduce power without hurting performance.

And, it doesn't hurt that you never buy standalone ARM CPUs.  You always buy ARM-based SoCs, so something that looks like a server-on-a-chip (à la Marvell's Kirkwood products), scaled up to multi-core, makes perfect sense for these web workloads.

Intel, OTOH, hasn't been much of an SoC developer in recent years (their work in the mobile and, now, server space is changing that), so x86-based solutions that need a CPU, north and south bridges, and various peripheral ICs, couldn't exactly scale down to the power level that even a scaled-up ARM solution would achieve.

Naturally, both sides are racing towards the same middle ground: Intel with Avoton, and the ARM camp with X-Gene, Seattle, and the others (Broadcom, Cavium -- Calxeda's and Samsung's server programs are RIP).

Intel has occupied the server space for so long, has the validated SW stacks, the process technology lead, and now a true low-power server SoC offering with Avoton.  What exactly is it that the ARM camp has to offer?

TarraTarra!
User Rank
Author
Re: What additional benefit as compared to x86 Servers
TarraTarra!   9/30/2014 8:08:01 PM
NO RATINGS
Intel spec scores using ICC are really not used by anyone (other than Intel) and do not correspond to real-world performance. Intel's compiler team is incentivized to crack benchmarks and circumvent them. Remember antutu? Spec has also been compromized. Intel spec scores with GCC compiler are as much as 40% lower than the ICC numbers presented here.

 

Wilco1
User Rank
Author
Re: What additional benefit as compared to x86 Servers
Wilco1   9/30/2014 7:14:42 PM
NO RATINGS
You can't compare SPEC like that. Intel artificially inflates their scores by using compiler tricks in ICC, so those are just marketing claims. If you believe that these scores translate to real world performance then you will be very disappointed when you try both side by side using the same compiler and settings. Back in the real world Avoton finds it hard to keep up with old mobile phones but of course Intel won't tell you that...

vdara0
User Rank
Author
Re: What additional benefit as compared to x86 Servers
vdara0   9/30/2014 7:01:44 PM
NO RATINGS
AMD Opteron A1100
CPU Core Configuration: 8 x ARM Cortex A57
CPU Frequency: >= 2GHz
SPECint_rate Estimate: 80
SPECint per Core: 10
Estimated TDP: 25W

Intel Atom C2750
CPU MHz: 2400
CPU(s) enabled: 8 cores, 1 chip, 8 cores/chip
SPECint(R)_rate_base2006: 101
TDP: 20W

TarraTarra!
User Rank
Author
Re: What additional benefit as compared to x86 Servers
TarraTarra!   9/30/2014 5:19:49 PM
NO RATINGS
@Rick TCO is the bottom-line for the data-center purchasers. The whole reason to make a move to a different architecture is to save on capex and opex. That is where the rubber meets the road.

 

Wilco1
User Rank
Author
Re: What additional benefit as compared to x86 Servers
Wilco1   9/30/2014 2:58:38 PM
NO RATINGS
More FUD from Gondalf here. The A57 core used in AMD's first generation ARM servers is already 25% faster per core than Puma/Jaguar (AMD showed SPEC scores to prove it). The next generation K12 is claimed to be significantly faster by the lead architect Jim Keller - and he should know.

Wilco1
User Rank
Author
Re: What additional benefit as compared to x86 Servers
Wilco1   9/30/2014 2:44:16 PM
NO RATINGS
It's well known that even the old Cortex-A15 beats Silvermont by a good margin, so it's obvious that the much faster A57 beats it by a larger margin. X-Gene is significantly faster than A57. 

If you're not accepting those facts and trying to claim that a 2-way partially out-of-order core is going to beat big aggressive 3-way and 4-way OoO cores then you're doing exactly what you accuse others of - being a fan boy for one camp. Your choice of words suggests you're Intel marketing, already calling X-Gene crap and claiming Avoton is better/faster without any evidence whatsoever.

Let me remind you that Intel marketing not so long ago claimed that ARM could not be used for the internet and that you could only make smartphones using x86. That ended up a bit different...

Patrikb
User Rank
Author
Re: What additional benefit as compared to x86 Servers
Patrikb   9/30/2014 9:48:17 AM
NO RATINGS
Considering that Intel started shipping 8 core (Silvermont) Avotons with 4MB L2 cache with a a TDP of 12W in Q3 2013, and should start shipping another respin (new process node, same microarchitecture), not so far off in the future, I think we need to keep timing in mind here. Seems like the ARM suppliers as they bet on server CPUs, will face the risk of competing with Intel in the game Intel really knows well, which AMD experienced just too well in the PC client market, squeezing the competition between a high end offering at high ASPs that they can't compete with and an entry level offering where Intel can afford to reduce prices where they other guys don't make money. Mobile is different with Intel having to catch up. In server it's the ARM camp trying to catch up, with performance, ecosystem issues and OEM penetration. Looks to me that ARM will have a niche in servers, like Intel will most likey have one in tablets and smartphones, but if nothing dramatic changes, it's hard to see any major disruption any time soon. 

Gondalf
User Rank
Author
Re: What additional benefit as compared to x86 Servers
Gondalf   9/30/2014 6:49:09 AM
NO RATINGS
I don't know, looking at AMD new K12 ARM core, it will be in the Puma range and the IPC will not be the main goal. AMD claims more efficence, still this means more cores per die with a low monothread throughput. This is hated by datacenter officers, without a good 1-4 thread performance a multicore cpu is useless.

Broadcomm thinks different but the company has not enough experience and on street reliability  to do a dent in the server market, in this segment a new player has a more than decennial way to follow to be trusted.

Page 1 / 3   >   >>
Like Us on Facebook
EE Times on Twitter
EE Times Twitter Feed