Yours if you can afford it… and wait long for the fabs to make the chips
Here’s a quick summary of what went down, minus the fluff.
- Huang acknowledged in a Q&A session that there is indeed a shortage right now of top-end Nv GPUs, mainly due to people grabbing them for cryptocurrencies and blockchain ledgers. Nvidia simply can’t make enough Tesla chips to go around, partly due to demand and partly due to yield, we reckon. Thus far, the chip designer’s only solution is: make and ship as many GPUs as possible. The biz wants to concentrate on giving hardware to cloud and supercomputer builders, gamers, graphics artists, scientists, engineers, enterprises… anyone but those annoying crypto-kids.
- If you’ve got roughly a million dollars to blow on deep-learning research, Pure Storage and Nvidia have produced a hyperconverged stack of flash memory and flagship Tesla Volta GV100 GPUs, plus some extra bits and pieces, called AIRI. Assuming the GPUs are available, natch.
- Programmers, engineers, and other techies dreaming of creating robots that sport some sort of machine intelligence can ask nicely to check out Isaac: this is a forthcoming software development kit and simulator, with libraries, drivers and other tools, for designing, testing and building machine-learning-based robotic gizmos.
- Speaking of programmers, TensorRT 4 – a GPU-accelerated deep-learning inference framework – has landed. Nvidia and Google boffins have integrated TensorRT into TensorFlow 1.7, if you prefer to use that engine for AI coding.
- Nvidia is stepping into the world of networking with the NVSwitch, a switch for its high-speed NVLink interconnect. Meanwhile, Tesla V100 GPUs – Nvidia’s top of the line chips for data centers – are now available with 32GB of HBM2 memory rather than the usual 16. Tying it all together is the new DGX-2 workstation, which has 16 32GB V100s connected via 12 NVSwitches for 2.4TB/s of bisection bandwidth and 512GB total HBM2 memory. The box has 1.5TB of system memory, two Intel Xeon CPUs, 30TB of flash storage, and Infiniband, 100GbE, and 10/25GbE interfaces. Nvidia claims it can hit as high as 2 PFLOPS with mixed-precision floating-point math. Yours for $400,000. This gear builds upon the $150,000 DGX-1 and DGX Station previously launched.
- Nvidia has jumped aboard Project Trillium, Arm’s effort to cram AI inference processing into chips powering wearables, gadgets, and Internet-of-Things devices. This is aimed at silicon designers: Nvidia is offering NVDLA as a free and open architecture for building deep-learning accelerators into hardware.
- If you’re working on self-driving cars – and who isn’t – you may or may not be tempted by Nvidia’s new Drive Constellation, a stack of boxes for simulating autonomous vehicle control software without crashing robo-rides or killing people (right, Uber?) Don’t hold your breath – these GPU-accelerated machines won’t arrive until the third quarter of 2018 at the earliest, and that’s only for Nv’s favorite customers.
- Scientists, engineers and artists needing some serious fire power for simulations and rendering are offered the new Quadro GV100 with 32GB of HBM2 memory. Each offers up to 7.4 TFLOPS of double-precision floating-point math performance, or 14.8 TFLOPS with single-precision, and can be linked via an NVLink interconnect to form one GPU, doubling the maximum potential performance and HBM2 capacity. These should be available now direct from Nvidia, or from suppliers next month.
- For those interested in virtualized GPUs in the cloud, there’s a bunch of announcements here. And if you’re a virtual reality designer or developer, there’s stuff to tease you here.
Of course, this is all subject to GPU availability. ®