Supermicro breathes in, shimmies a PB of Intel flash into one rack unit Supermicro has crammed 1PB of Intel flash rulers into the slimmest possible 1U rack storage server. The two-socket server can hold up to 32 Intel EDSFF, NVMe-connected flash drives – giving a rack density of 1PB/U, the highest we have ever come across. EDSFF stands for Chipzilla's Enterprise and Datacenter Storage Form Factor …
Yeah, you can't.
I think you may want to try CDW but even there... not much.
I think this is something you have to go directly to your hardware vendor to get, and then maybe some of the vendor supply shops that usually don't deal directly with consumer.
Very pricey and most of these guys won't keep kit on hand and will have to special order it.
My guess... the 1U box would be over 500K (USD) to 1 million (USD) fully kitted out. (Probably closer to 600-700K range.
One of these in the center of the rack w 4 4U boxes below and 4 4U boxes above. (plus ToR switches)
And the 4 U boxes contain NVidia GPUs.
Now that's a killer platform for ML or 'Big Data' with the 4Us also having a bit of local disk too.
Of course you'd be looking at a cool 2 million+ USD per rack fully kitted out...
NVMe has no SAS stack and hence no RAID. You have a bunch of individual devices which are very very fast. Usual limitation is 10 per servers due to lack of PCIe lanes.
I need to take a closer look at this box and see how they upped the number of available lanes for a 2 socket system, if they didn’t it will be a world of pain.
This box needs a fast SDS layer to control all those individual devices, VSAN and Nutanix for example.
Well AMD Epyc has 128 PCIe lanes in either single or dual CPU configs.
32 Lanes for external comms (4 x 100GbE or 16 x 25GbE ports) and 96 lanes for 32 storage devices (3/Ruler) makes sense. That is nearly 3GB/Sec per device (2,955MB/Sec).
Of course Intel will probably not license the EDSFF form factor for AMD based systems so it might all be for naught.
Is it economically sensible to actually buy these compared to, say, four 256TB units and then spend the purchasing dollars saved on renting extra datacenter floorspace/power/cooling?
I'm sure in time they'll become economically viable, but is that 6 months from now, or 2 years?
I'm not a datacenter purchasing, commissioning or operations engineer (I write code), just curious to understand the economics here.
It's about density and you've got a workload that needs many petabytes then you'd go with something like this versus a number of smaller units. Also the incremental cost for each server, plus any licensing might be another 10k a unit though that's probably a drop in the bucket compared to the total cost of one of these loaded up with flash.
This post has been deleted by its author
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
