News Categories

NVIDIA supercharges DGX-1 supercomputer with Volta, bumps it to US$149,000 and debuts DGX Station at half the price

By Vijay Anand - on 11 May 2017, 7:59pm

NVIDIA supercharges DGX-1 supercomputer with Volta, bumps it to US$149,000 and debuts DGX Station at half the price

NVIDIA just debuted their first Volta architecture based GPU and it’s a phenomenal chip. Rising from the US$3 billion investment that went into creating the very best compute unite money can buy, NVIDIA is firmly committed to providing the most productive platform for deep learning.

In fact, the DGX-1 supercomputer is touted to be the fastest path to deep learning and it’s apparent from customers who’ve invested in them. Configuring and producing the hardware isn’t really NVIDIA’s key differentiator in offering these supercomputers. But, offering the right software stack in the DGX-1 to get it going immediately from day one and support to navigate the realms of deep learning is as important as the hardware that’s powering it.

At GTC 2017, when NVIDIA debuted the crazy fast Volta-based Tesla V100 compute unit that can churn out 120 Tensor TFLOPS, they made sure to reserve the first batch for their DGX-1 supercomputers. Interested customers can make a payment out to NVIDIA for US$149,000 and expect a DGX-1 equipped with 8 x Tesla V100 compute units sometime in Q3 this year. OEM vendors will get their hands on the Tesla V100 for their implementation from Q4 2017. Here's a quick breakdown of the updated DGX-1:-

In case you didn’t realize, that’s US$20,000 more expensive than the Tesla P100 Pascal based DGX-1 supercomputer, but judging by the performance upswing that the Tesla V100 variant boasts, NVIDIA’s asking price is probably justified. Not to mention, they’ve invested dearly in creating the GV100.


Enter the personal AI supercomputer, the DGX Station for US$69,000

Now this new entrant is a supercomputer in a familiar tower form factor that’s desk friendly and super quiet thanks to its water cooled internals. It’s put together with the same ideals as the DGX-1, only that this is more accessible, both in cost and in deployment thanks to its form factor.

It's a sized down DGX-1 in every way – half the CPU(s), half the GPUs, half the system memory, half the power draw, you get the picture. Here’s how the new DGX Station decks up against the first version of DGX-1 and the updated Tesla V100 equipped DGX-1:-

GPUs 4 x Tesla V100 8 x Tesla P100 8 x Tesla V100
TFLOPS (GPU FP16) 480 170 960
GPU Memory 64GB 128GB 128GB
CPU Intel Xeon E5-2698 v4  (20-core) 2 x Intel Xeon E5-2698 v4  (20-core) 2 x Intel Xeon E5-2698 v4  (20-core)
NVIDIA CUDA Cores 20,480 28,672 40,960
NVIDIA Tensor Cores 2,560 -- 5,120
System Memory 256GB 512GB 512GB
Price US$69,000 US$129,000 US$149,000

Now the really interesting bit is that the DGX Station actually has more compute power than the original DGX-1 supercomputer – 480 TFLOPS vs. 170 TFLOPs – and it costs a lot less as well. This is all thanks to the new Volta architecture based Tesla V100 compute units that’s used on the DGX Station.

If that piqued your interest and have US$69,000 to spare, you can pre-order the DGX Station from NVIDIA and it will be ready to ship from Q3 this year, just like the updated DGX-1.

A peek inside this personal AI supercomputer.

Note the motherboard supplier and model used.

You can better see the water cooling kit's radiators and coolers from this view.

Quite a nice looker indeed. It better be for US$69,000.

Source: NVIDIA (1), (2), (3)

Join HWZ's Telegram channel here and catch all the latest tech news!
Our articles may contain affiliate links. If you buy through these links, we may earn a small commission.