When it comes to AI, Pure twists FlashBlade in NetApp's A700 guts

Pure Storage's AIRI FlashBlade is faster than NetApp's A700 all-flash array, according to two AI benchmark runs. The two suppliers have both designed tech using Nvidia DGX-1 GPU servers and their own storage arrays to provide storage-server systems for "AI" applications such as deep learning. Pure's deliverable hardware- …

[Mr.Nobody]

NetApp has an AFF A800 now, so?

The great news about this absurdity is that it is good for the consumer, i.e. businesses. Competition is good.

We already had very good pricing on our previous NetApp purchases, which for some reason has a lot to do with the business line one's company is in.

When we switched to AFF a few years ago, our pricing was even better, mostly because Pure was nipping at their heels. Pure has forced NetApp to be far more competitive in terms of pricing, our VAR at the time said it would be NetApp's undoing (his exact words were that they were dropping their pants).

Seeing NetApp's more recent financial statements seem to show otherwise.

[Secta_Protecta]

Re: NetApp has an AFF A800 now, so?

Unfortunately NetApp missed the boat with AFF, allowing Pure to get a foot in the door with their overpriced, substandard product. I'm Pure accredited btw due to working for a Pure partner so I'm familiar with the ins and outs of it. When NetApp did release AFF it proved to be a much better product at a much more reasonable price, which has been borne out by the sales figures. Pure is a marketing company rather than a technology company and FlashBlade is their last roll of the dice after years of disastrous results...

[Anonymous Coward]

Not enough Info

How many A700 nodes vs how many FlashBlade blades...

If they are comparing 15x Blades vs 2x A700 nodes you know who is winning this...

[RollTide14]

Hard Hitting Journalism

Great work here Chris! I’m expecting your next story to be about how a customer has saved 8 racks of equipment from moving from 8 year old 1TB drives SATA drives to 4u by moving to all flash.

[Arthur A.]

Wow. What a shame. Why do you use wrong NetApp's numbers? There are two groups of graphics in NetApp's white paper — with image distortion enabled and disabled. And you took results for test with image distortion enabled. Pure tested FlashBlade only with image distortion disabled.

So right numbers for NetApp are:

ResNet-152: 138, 274, 547, 1092

ResNet-50: 341, 631, 1259, 2521.

Please update the article.

[handogod]

Here is the NetApp related PDF ...

Here is the actual NetApp writeup - https://www.netapp.com/us/media/wp-7267.pdf

[Anonymous Coward]

Flasblade vs A700

That's an ass kicking exercise NetApp will win with its hands tied behind its back.

NO WAY.

[Terry P]

VENDOR A IS SO GOOD, LOOK!

GO TO HELL VENDOR B ROCKS, THEY DID XYZ ANS DIDN'T TURN OFF THE ROCKWELL RETRO ENCABULATOR

Meanwhile,

- Pick a company you like

- Pick an account team you like

- Do your due diligence and do right by your company, they pay your salary

- Be nice to people and ignore benchmarks.

[Secta_Protecta]

As others have stated, what are the specifications of FlashBlade against the A700? Not to mention that Pure would have been wayyyyyyy more expensive, even with their ludicrous Evergreen (or whatever they're calling it nowadays) promise.

[Anonymous Coward]

The numbers are very similar, maybe a 10% delta, and the data-set is merely 143GB

Netapp writes:

"TensorFlow benchmarks were run on a 1:1 configuration setup (one DGX-1 server and one A700 storage system) with ImageNet dataset (143GB) stored on a FlexGroup volume on the A700 system."

Considering how similar the Pure and Netapp results are; and considering the data-set is merely 143GB; I think it is safe to conclude that storage is not the bottleneck here, both arrays can easily cache the entire ImageNet in the storage array's RAM.

[Anonymous Coward]

Probably a good idea to run your own end to end test

At 250K for a dgx-1 you’d probably want to run your own Training workload and see which high performance NFS solution works out best for you. Small dataset

Image processing like Resent probably isn’t the best benchmarking tool for evaluating the relative merits of various storage layer technologies, and you’d probably want to evaluate other stuff like “how do I back this up”, but it is interesting that in both cases you’re looking at scale out NFS built on flash and not old school HPC like lustre or even HDFS

[WYSIWYG650]

How cute...now for apples to apples

This must be an old article or an uninformed author. NetApp has the A800 to deal with this competition and does so in ways they could only dream about. Both are real fast but only one has multiple protocol and cloud connectivity. Edge to core to cloud...FTW. https://www.netapp.com/us/media/wp-7267.pdf

[WYSIWYG650]

Update with current info...

A single NetApp A800 system supports throughput of 25GB/s for sequential reads and 1 million IOPS for

small random reads at sub-500s latencies3

. In addition, what sets A800 apart is its 100GbE4 network

support, which accelerates data movement and also fosters balance in the overall training system,

because the DGX-1 supports 100GbE RDMA for cluster interconnect. The NetApp A700s system

supports multiple 40GbE links to deliver a maximum throughput of 18GB/s.

[bbkudk]

AI/DL tests available for NetApp AFF A800 system with DGX-1

AI/DL tests available for NetApp AFF A800 system with DGX-1 in nva-1121 document at NetApp Library site.