Naples

>>59690856
That's a fucking LOT of memory.

inb4 180W lelelel typical AMD xD

>>59690856
8 DDR4 channels per socket is nice, I just wish there was a fucking workstation platform AMD was selling now that supported >64GB total capacity.

16c MCMs or just buffered ECC support, I don't even give a fuck at this point.

>>59690881
16 per socket, the new Skylake-EX have 12 per socket.

>>59690856
>all that unused pcb space
disgusting

>>59690890
Wat. It's 180W for 32/64 while Intel has 28/56 with 200W.
>>59690898
>he
>doesn't
>know

File: Screenshot from 27-03-2017 18:07:18.png (208KB, 458x572px) Image search: [Google]

208KB, 458x572px

>>59690890

>>59690899
Is is still 4 channels with 3DPC, or is it 6 channels * 2DPC?

>>59690856
>m-muh optane

>>59690915

>>59690898
>platform AMD was selling now
>selling now
>now

please enlighten me.

>>59690919
The latter.

>>59690910
>has IPMI
>has Infiniband
>has ethernet
>has serial or VGA

Nothing else needed, enjoy your needles board power from your 15W chipset

>>59690940
if so, could they squeeze in 6ch*3DPC for 36 slot configs then?

>>59690933
Just wait™.

>>59690955
On 2 socket? Doubt it.

>>59690856
How much memory does it support?

512gb or 1tb?

>>59690856
Would there be any gains to filling all the slots vs just 1 or 2 with the same total memory? Assuming same statistics on each mem card and all that.

>>59690987
4tb for 2P system AFAIK.

>>59690942
what the fuck are you talking about, idiot.
I'm talking about UNUSED PCB SPACE. For what fucking purpose they made this PCB so big if it's 1/3 unused?

Are you retarded?

>>59690987
Max memory support is 5TB IIRC.

Not that you'll ever get that even with 128GB DIMMs.

>>59690998
Have you considered that PCIe cards extend beyond their slot length, and computers need air flow to cool them?

>>59690998
For starters it's an engineering board, secondly the entire width of the PCB is taken by sockets and DIMMs.

>>59690991
Yeah, Ryzen ASICs each have 2 DDR4 channels, so you get 8 channels on a 4 die Naples chip.

You won't get full bandwidth without at least half the slots populated.

>>59690856
I counted 32 DIMMs but I could be wrong. Fuck, that would hold 512 GB of RAM using 16 GB sticks.

GAME OVER
A
M
E
O
V
E
R

>>59690918
Oy housefires.

>>59690998
"I've never seen a server and am also retarded" the post

File: single socket holocaust.jpg (28KB, 550x550px) Image search: [Google]

28KB, 550x550px

>>59691055
Like the good old times.

>>59690918
>these are turbo frequencies

>>59691052
Might as well use 128gb sticks for 4TB of RAM

>>59691082
Literally nobody buys 128GB DIMMs, at least not anybody outside of Amazon/Google/MS

Most enterprise costumers are satisfied with 32GB DIMMs

>>59691066
>Single Socket Holocaust
LMAO
Hyper-pipeline technology

>>59691082
Imagine having an entire large production database loaded in RAM.
The future is now!

File: bibeline.png (973KB, 801x1500px) Image search: [Google]

973KB, 801x1500px

>>59691115

>>59690856
>that AMD SMT performance on 32 cores

Holy Molly

>>59690856

AMD needs to fix it's enterprise level support and warranty department first before they grab hold of the server market again.

Those huge cloud providers like AWS, Google Cloud, Azure, etc. Build out their own servers and buy components in bulk, because it's cheaper and they can also easily design them around their facilities custom cooling systems. Dealing with AMD for RMA is a nightmare at the present moment, Intel will literally just ship out another CPU in a few hours after you report it.

Even conventional server builders like Dell, Supermicro, HP, etc. will probably be hesitant to sell AMD servers on a large-scale because of the same issue in terms of delay in RMA.

It's more then just performance. Intel's enterprise/datacenter level account managers make stuff happen for these providers on massive scales.

File: rip.jpg (167KB, 990x557px) Image search: [Google]

167KB, 990x557px

>>59691122
That's the one!

>>59691052
It's not about raw capacity, it's about capacity/$ and bandwidth.

32GB ECC DIMMs ~= $300, 64GB ~= $900, 128GB = you can't afford it

So you can now get 1TB RAM in one box for ~$10k vs. ~$14k in a Broadwell Xeon, plus twice the bandwidth. New Xeons with 6 channels each will sit somewhere in between.

>>59691138
I think AMD is aware of that more than you.
If they're selling a competitive HPC chip, they won't sell it without good support, hopefully something as smart as Lisa Su understands that, since she did kinda work in that field in IBM

>>59691157
You're not wrong, but I wouldn't expect a private citizen to purchase that much RAM. I was thinking enterprise.

>>59690899
From what I've been hearing, not all skylake xeons are 6 channel(I assume the 18-28c ones only) while it's 100% guaranteed that all 4 die MCMs Naples will be 8 channel.

>>59691138
Nice pasta Schlomo.
I've seen it several times already.

>>59690856
>180w
Mom basement right

>>59690856
128 threads, I could serve a decent running VM to 32 users with this.
I'd pay out of the ass for switches though.

>>59691236
Why is high grade networking hardware so fucking expensive compared to 6 billion+ transistor ICs?

>>59691270
1. Because jews
2. Because networking hardware runs ultra low latency ASICs on real time OSes
3. Lower demand
4. See jews

>>59691270
because consumers don't need 10+ GbE so volumes won't be high, and enterprises both need it and can pay out the nose for it.

there does appear to be some sort of nigger-tier 2.5 and 5 GbE over UTP standard emerging though. presumably doesn't need the LDPC coding with the 3-5W decoders.

>>59690856
Sparse amount of CPU VRMs for 180W chips, even at stock.

HOW CAN RAMLETS EVEN COMPETE?

>>59691576
You're just looking at quantity, some amps are rated at 20A and higher which is more than enough for a 180W IC even in low number.

And from what I know this is a validation board, so there's not gonna be strict power circuitry on it

>>59691066
>stuck one of these in a G40
>runs for 5 minutes before the power draw and heat forces a shutdown
GOD TIER

>>59691032
>For starters it's an engineering board
only valid point.
kys

>>59691181
32GB DIMMs are the current size/$ sweet spot for anybody above baseline consumer level.

I'm looking forward to 16c32t workstation Ryzen I can throw 4*32GB in and leave room for an upgrade to 256 GB total in the future if I wanted.

>>59691138

While correct, AMD doesn't need to take over the market, they just need to crack it open to start becoming profitable again.

When they start becoming profitable, they can hire all sorts of people that were culled just for the enterprise service alone.

If AMD takes over just 5% of the data center market, that's several billion dollars for them to reinvest and keep going forward.

>>59691236
buy an lb6m bro.
buy mellanox connect x-2's off ebay for 15 a pop, and 5meter sfp+ cables for 15 as well.
enjoy 10gig networking for up to 26 appliances for 200 + 30 bucks per appliance.

its amazing.

>>59692646
God bless ebay prices on used server hardware

>>59692646
Can you connect an infiniband port to a 10gbit sfp port?

>>59692801
A lot of Mellanox equipment (most of their card, not sure how many of the switches) can talk with either ethernet or IB.

In general, no, though.
Also not sure how many Mellanox cards are SFP+ (not SFP) vs. QSFP. Infiniband has been at least "40Gb" (actually 32 you filthy desert kikes, you're not allowed to count coding 8b10b overhead) for quite a while now. Older 10Gb (8Gb) IB might be over the old CX connectors only.

>>59690856
Can I play gayms on It?

File: 000001.jpg (290KB, 1741x975px) Image search: [Google]

290KB, 1741x975px

>>59691141

>>59691161
>I think AMD is aware of that more than you.

Clearly not considering their market share.