Naples AMD

It's over.

youtube.com/watch?v=PN93G6Rg2ek

arstechnica.com/information-technology/2017/03/amd-naples-server-processor-more-cores-bandwidth-memory-than-intel/


Per 2P server:

128 PICe Lane
32 DIMM Slots
8 channel memory
Chipset & CPU on the same PCB -- SoC.
*128 threads
Infinity Fabric

Arriving Q2 2017

Other urls found in this thread:

ocpussummit2017.sched.com/event/9TZ3/amd-amd-naples-the-future-is-open
ocpussummit2017.sched.com/event/9TZl/amd-amd-naples-open-innovation-for-the-software-defined-era
ark.intel.com/products/93790/Intel-Xeon-Processor-E7-8890-v4-60M-Cache-2_20-GHz
ark.intel.com/products/96899/Intel-Xeon-Processor-E5-2699A-v4-55M-Cache-2_40-GHz
twitter.com/amdserver?lang=en
youtube.com/watch?v=cLoytewvn0g
twitter.com/SFWRedditVideos

How can jewtel ever recover

Benchmark against top of the line

Intel® Xeon® Processor E5-2699A v4
2255M Cache, 2.40 GHz, 22 cores, 44 threads.

ocpussummit2017.sched.com/event/9TZ3/amd-amd-naples-the-future-is-open
ocpussummit2017.sched.com/event/9TZl/amd-amd-naples-open-innovation-for-the-software-defined-era

Amazing

>55M Cache, 2.40 GHz, 22 cores, 44 threads.

Fixed*

...

ark.intel.com/products/93790/Intel-Xeon-Processor-E7-8890-v4-60M-Cache-2_20-GHz

AYYMD as usual did not bench against the best

Rypoo Naploo is a fraud as usual

These arent games

...

come now we don't have to fight over this one, we'd have to just check q3 AMD finances to see who is more clever you or all the people responsible for datacenter hardware

This made me want to visit Naples. Real world things

Lol

But yeah it seems AMD has support for more memory and channels per socket while having more I/O AND more cores, if their TDP stays within that range then it's a win on all fronts.

>dem misaligned DIMM slots
Triggered

depends on clocks, zen can keep it 6w per core at 2.8 I think or somewhere near those numbers

someone didn't use a grid while modeling it

6.25W on 3.3GHz per core running AVX without cherrypicking dies like they would for naples.

35W TDP on 8 cores with 1.9GHz

It's looking pretty damn good for AMD.

No game benchmark, into trash it goes

Performance scaling at TDP

>7700k still has more fps

AMPOO BTFO

>being so much of a consumerist cuck that you have to root for "your side" and bash their competition in the enterprise market even though you'll never own enterprise hardware

Yeah but how fast is it in GTA V @ 1080p?

games

...

Some voltage scaling.

Thank fuck someone actually bothered going into low-level testing instead of just fucking benchmarks.

How do I read these graphs, WTF does fmax and vmin mean

>Intel still has microsttuter

INTEL BTFO

seriously will we ever agree that ryzen and kaby two absolutely different CPU for different needs, and you can choose either for benefits one of them provides?

laughed way to hard at this

I hope the 16 core Opterons don't cost more than $800, doesn't even have to be the highest clocked one, just needs to have a lot of memory and I/O

And hopefully 1P motherboards with IPMI don't cost more than a kidney.
Oh who am I fucking kidding I'm probably looking at $600 for mobos alone.

ark.intel.com/products/96899/Intel-Xeon-Processor-E5-2699A-v4-55M-Cache-2_40-GHz

>Max Memory Size (dependent on memory type) 1.54 TB

>Intel's comparable offering? Twenty-two cores and 44 threads, eight memory channels, and a maximum of only 384 GB RAM.

AYYMD intentionally crippled Intel again, what a scummy company

Those 2 core complexes instead of 1 monolithic one does WONDERS for low power pstates, effectively makes everything easier to gate and power consumption goes down.

Ars already called out AMD BS benchmarks:

Since Naples has twice the memory channels and therefore twice the bandwidth, of course Naples will out perform the Intel setup with a MEMORY CONSTRAINED workload.

Since Intel has literally twice as good AVX 256+ performance, a pure computationally constrained workload similar to their example should heavily favor Intel.

>intel
>only 22x2 cores in 2017

Both are valid, Intel doesn't have physically the memory and AMD doesn't have the AVX512 throughput, they're comparing what's there.

Lets see that stock drop now bois

They do offer quad channel memory per die... or 8 channel per socket.

So right there that's something more expensive than current Ryzen chips.

Plus they have WAY more I/O. Like 32 PCIe lanes per die, 64 lanes per socket... compared to only like 20 or 24 on Ryzen.

16snowy owl chips are rumoured to be embedded parts to compete with xeond. Think low power, clocks and expensive motherboards because of 10gbe.

I think they released a 24 one IIRC.

>seriously will we ever agree that ryzen and kaby two absolutely different CPU for different needs, and you can choose either for benefits one of them provides?

Did you forget what site you're on?

There are no quad channel dies.
There is only the singular Zeppelin die.
Naples is a 4 die MCM. It has up to 128 PCI-E lanes per socket.

BGA? I can deal with that, might make the setup cheaper if it's prebuilt.
Most of these large boards are.

>Naples
apparently Ryzen wasn't a bad enough name I guess

twitter.com/amdserver?lang=en

holy fuck that's some efficient shit right there
once the APUs come out I'm building myself a really efficient ryzen HTPC

fmax = max frequency
vmin = min voltage

for AMD's sake I hope they have their firmware and software support sorted when they launch this fucking thing this time. Data Centers have a lot less patience for broken buggy pieces of shit than your average amdrone.

>insufficient memory to load

voltage and frequency.

What are you doing in this thread, on this board.

you idiot the process crashed because the cpu couldnt feed fast enough the program it resulted on a halt->crash

there is a reason why those kinds of workloads are passing through the gpu on intel systems

Why do you think consumer 8 cores were released first(besides binning reasons), all these 32 core chips are made out of 8 core dies, fixing firmware and software on them fixes them for the Naples ones too ignoring edge cases.

Everyone currently owning a R7 Ryzen is doing validation testing for AMD's server platform.

Ah, I was thinking they were gluing two Ryzen7 dies together on a substrate.

You know what I meant though.

They're offering twice as much memory channels and over twice as much I/O per every MCM pair.

NOOOOOOOOOOOOOOO DELET THIS

BUT CAN IT REACH 300 FPS CAP IN DOTA?!

You're the idiot, it's about memory capacity, not speed

Intel's chip would certainly have sufficient memory to load that benchmark but AMD intentially equip it with 384GB only

Learn to READ, moron

...

The 32core won't be 3.3Ghz. That'd be like 180-200w TDP.

My guess is something like 1.9Ghz base, 2.6Ghz turbo for the 32core.
And maybe 3.3Ghz turbo for 16 core.

AMD didn't equip anything, these are prebuilt OEM 2p racks and no small number of them top out at 384GB of memory for the current Intel.

The top Naples SKU has a 180w TDP.
They have a significant clock advantage over Broadwell Xeons.

Max frequency minimum voltage

Cmon dude

Man that must be a massive amount of memory usage.

what is the gaming performance like

Shit, actually I'm wrong.

AMD has 8 MCM per socket... 32c/64t

Still won't cost less than $1000 bucks...

That's literally like 4 Ryzen 7 CPUs on a single socket.

this
run it at 480p pls

ryzen is beta test for servers
drop in the ocean compared to money they will make on data centers

lol oops. I retract what I said.
Yeah the benchmarks line up with it possibly being 3.3ghz all-core? On 32 cores?

Wow, that's crazy. Way way way beyond my expectations and any other expectations I've seen.

>criticize blatant consumerism and shilling
>HURRR U R Sup Forums

It runs games 5fps slower than a 7700k, so it's trash.

Jesus

>intel
>only 24 cores in 2017

...

I'll likely have a base clock somewhere in the range of 2.6ghz inside that 180w envelope, unless their hand picked enterprise dies are radically better than what we've seen in consumer Ryzen. Its possible for them to have better clocking parts, but I don't think we'll see that much of a leap. A 2.7ghz base clock would be tremendous.

That seismic analysis must use some obscene amount of memory, anyone knows exactly how much?

>Finally, AMD offered a workload that quadrupled the size of the data set. The Naples machine completed the task in 54 seconds, while the Intel system crashed with inadequate memory. This proved convincingly that, yes, 2TB of RAM is more than 1.5TB of RAM.

1.5TB

I don't know man, the 1700 has a configurable TDP at 35W and it runs 8 cores at 1.9GHz.
Server dies are always gonna be more cherrypicked on a tweaked process for higher efficiency, so I wouldn't be surprised to see server parts at that frequency at 2.1GHz.

8 cores at 2.1GHz, 35W

I don't really know what to say when you scale that up.

see

If you follow TheStilt's Cinebench scaling then at 45w they're hitting 2.5 or 2.6ghz.
Four 45w Zeppelin dies would come out to 180w.

2.0 base, 3.3 turbo, with a max of 2.5-2.6 all-core turbo sounds very likely to me.

LMAO

Will Linus update his shill boxes?

Jesus Christ

see
1000 score at 35W, 1150 at 45W

15% increase for 29% more power, efficiency loss.
Zeppelin is most efficient around 2GHz.

Only if they give him one, this shit must be expensive as hell

Intel is over. No gamer faggots, compilers, Microsoft feet-dragging will shit up this launch.

What is turbo, exactly? My only experience with turbo clocks is turning my unlocked chip's multiplier up to the turbo clocks or higher. Windows seems perfectly able to downclock everything during idle, even when the multiplier is overclocked, so I'm not clear on my manufacturers even make the distinction between "base clock" and "boost clock" when boost clock seems to kick in essentially whenever CPU usage is high, which is no different than when you don't have this turbo feature... right?

Thread theme
youtube.com/watch?v=cLoytewvn0g

>4U rack
>almost 300 lanes with full x16 lane support for 8 GPUs and every SAS/SATA/NVMe/NIC setup under the sky
>6600 watts Platinum with 3 PSU redundancy


MY
FUCKING
DICK

Also only $70000 starting

Turbo overclock certain cores and downclocks others while still respecting the combined TDP of the chip. It mainly improves performance in single threaded tasks on multicore chips. The TDP does not account for all cores running at the turbo frequency and the voltage might not be sufficient to run all cores at turbo frequency at least not in a stable manner.

Max Turbo is traditionally used for one core, there's an all turbo that boosts all cores by a few hundred MHz at best if it doesn't go over the TDP in certain conditions(non-SIMD load)

This all-turbo is very rarely marketed outside of datacenter spreadsheets.

Turbo is a temporary overclocked done by the chip itself when more single core performance is needed.

The turbo frequency is almost never what all cores can operate at, and the "all-core turbo" is always somewhere in between.

>128 PCIe Lane

PLX switch vendors on suicide watch.

couldn't they have atleast made a cooler design?

So if I have some random locked i5 CPU with a base clock of 3 GHz and a turbo of 4 GHz, and CPU-Z reports 4 GHz while I'm running Prime95, what is actually happening? Is just one core OC'd to 4? All 4?

It's a complete royal flush.

More memory, more bandwidth, more cores, lower power package power, lower platform power.

Is this the return of k8?
Zen seems to be called K17 internally, it's literally K8x2 and then some.

This is so sick

One.

Is there a way to view the clock speed of each core individually? And why does it stay at 4 GHz throughout the duration? I assume that depends on cooling and temps, correct?

...

16 channel memory? Not quad channel? Not octochannel?

What the fuck?

its K12, K8 and barely half more, sounds about right

hwinfo or openhardwaremonitor IIRC, all-core-turbo is SKU dependent.
see
For consumer i5-i7 I haven't found something similar.

I hope they have a 16 core 3.6ghz for workstations. If there's a 1P board it will be ideal.