If you've got $1m+ to blow on AI, meet Pure, Nvidia's AIRI fairy: A hyperconverged beast

Pure Storage and Nvidia have produced a converged machine-learning system to train AI models using millions of data points. It’s called AIRI – AI-Ready Infrastructure – and combines a Pure FlashBlade all-flash array with four Nvidia DGX-1 GPU-accelerated boxes and a pair of 100GbitE switches from Arista. The system has been …

[Bitsminer]

Ooops

$3,000 for an Arista 100G switch? Think again.

[returnofthemus]

Again, it’s in a different league from the IBM system....

LOL.... Really!

Which one?

[Anonymous Coward]

Bullshit Bingo Mixup

Hyperconverged? This is quite obviously a converged system.

[Nate Amsden]

similar to HP?

Each DX-1 system appears to be a single Xeon server with 8 V100 GPUs in a 3U enclosure consuming 3500W of power

HPE's offering is a single xeon server with 8 V100 GPUs in a 4U enclosure, a bit bigger but the HP offering supports a lot more memory and local storage, power capacity is 8800W or 4400W if you were operating on one pair of active/standby power circuits(4 power supplies). So power usage is likely to be about the same as well (not surprising).

Comparing 4 servers each with 8 GPUs to a single server with 8 doesn't compute well.

[baspax]

Re: similar to HP?

The HP box has PCIe which is a huge bottleneck whereas the DXG1 has NVlink

https://www.nvidia.com/en-us/data-center/nvlink/

[MiguelC]

A very fine machine indeed, but can it run Crysis?

(sorry, but someone had to...)

[Brenda McViking]

Re: A very fine machine indeed, but can it run Crysis?

Showing your age there...

The kids these days are only interested in mining Ethereum with this sort of hardware; as in - how many mega-hashes can it smoke through when it's on fire?

[JohnMartin]

Already out of date ...

Why buy DGX-1 when the DGX-2 just announced is 10X faster ?

[Anonymous Coward]

Re: Already out of date ...

I agree. I would bet they are already working on the same thing with DGX-2s. Also, you are wrong, DGX-2 is 2x (not 10x) faster - see the flops comparisons: https://www.nextplatform.com/2018/03/28/nvidia-dgx-2-system-packs-an-ai-performance-punch/

[Anonymous Coward]

Just another vendor lockin..

Why would anyone with wallets this deep go for a vendor lockin on AI when the market's in such flux? Pure stupidity.

[Anonymous Coward]

I am curious if anyone could comment on the data reduction potential of the kind of workloads a machine like this would run? High, low, medium? In the article I think it implies a 5:1.