A showcase of the most powerful systems for deep learning, A.I. on the edge and more
Without fail, NVIDIA’s GTC 2018 has once again stirred the world of deep learning, inferencing and more with a slew of powerful hardware updates in the supercomputer, server and workstation segment, as well as more capable software stacks and APIs to set new records and capabilities such as real-time ray tracing. We round up some really interesting hardware we’ve come across from the show floor.
By Vijay Anand -
The company’s latest and greatest – the GV100 Volta GPU architecture based Tesla V100 32GB – is now available to the millions of professional creators in the form of the Quadro GV100 32GB.
Touted as NVIDIA’s largest and most powerful graphics card (Tesla cards don’t have display outputs), the company skipped the traditional green livery for a more classy gold finish that fits its pole position.
Coupled with company’s just announced NVIDIA RTX technology, Volta-class GPUs like the Quadro GV100 can now deliver real-time raytraced renders for professional design and content creation needs - a huge breakthrough in the realm of computer graphics.
Should your graphics workloads require an even larger memory footprint and more processing throughput, you can operate two Quadro GV100 GPUs together with NVLink connectors such as the setup shown here. On NVLink supported applications, this gives you over 10,000 CUDA cores and 64GB of unified memory at your disposal. Monstrous figures indeed, and so is its price – up to US$10,000 per card!
If that’s not crazy enough, within a span of six months, NVIDIA topped their DGX-1 supercomputer’s performance by a factor of almost 10x with NVIDIA’s DGX-2 supercomputer unveil at GTX 2018 that packs a whopping 16 Tesla V100 32GB NVLink GPU modules. NVIDIA CEO Jensen Huang is seen here posing in front of one of the company’s greatest solutions ever in pursuit of enabling the autonomous future.
NVIDIA had an exploded edition of the DGX-2 on display which gives a better sense to how it packs 2 petaflops of compute power, a world’s first for a single server machine.
Unlike the DGX-1 that made use of the Tesla GPU’s own NVLinks to mesh up to eight such GPUs, the DGX-2 employs a new connectivity topology that’s enabled by NVIDIA’s new NVSwitch interconnect switching chip (as seen here). Boasting 18 full-bandwidth ports, it is capable of addressing far more GPU connections than the six ports on a Tesla V100 GPU. The DGX-2 employs 12 of these NVSwitch chips for no-compromise performance between its 16 Tesla GV100 NVlink GPUs.
Here’s a side-by-side comparison of the Tesla V100 32GB unit next to the 2-billion transistor NVSwitch chip.
At the back of the DGX-2, you can see its several redundant hot-swappable PSUs, numerous Infiniband 100GbE connections to scale up the processing capabilities with more DGX supercomputers, and the connections to the NVLink plane card that interconnects the three planes within the server that hold the GPU array and the CPUs.
NVIDIA isn’t the only deep learning systems provider of course and there are many alternatives from traditional server vendors such as from OSS and SkyScale (pictured here). This particular system features 16 of the Tesla V100 PCIe edition GPUs.
If the US$69,000 NVIDIA DGX Station for personal deep learning development is out of your reach, BOXX systems solution provider has debuted a new compact deep learning workstation – the APEXX W3. Featuring a single Xeon W processor with up to 18 cores, it supports up to four graphics cards, half a terabyte of DDR4 ECC memory and up to five SSD drives.
Here’s a sneak peek at the BOXX APEXX W3 workstation that can house up to four graphics cards. It’s compatible with GeForce, Quadro or Tesla variants depending on your configuration needs. As such, the APEXX W3 can be priced from us ‘low’ as US$10k.
Looking for a powerful mobile creative professional workstation? While there are many such options in the market, the PNY PrevailPro workstation is one of the first few of its class to be endowed with a Max-Q design Quadro P4000 mobile GPU. As a result, this powerful mobile workstation with Core i7-7700HQ processor is only 18mm thick and weighs just 2.2kg.
Traditional mobile workstations would easily weigh twice as much, so the PNY PrevailPro is far more portable while still delivering uncompromised power and performance to run demanding professional applications that support GPU accelerated compute. And yes, it supports up to three external 4K HDR displays to comfortably expand your visual workspace.
When an NVIDIA Jetson based edge/embedded device solution isn’t quite enough to satisfy your computation needs, a Quadro MXM module riding atop a micro-STX motherboard can unleash desktop performance levels in a compact rugged form factor. Shown here is a PNY Quadro P5000 MXM module working alongside an industrial class ASRock MXM IPC-H110 motherboard that supports Intel 6th or 7th generation S-class Core processor.
Here’s a closer look at the PNY Quadro P5000 MXM module featuring 2048 Pascal CUDA cores, 16GB memory, 192GB/s peak memory bandwidth and a TDP of 100W.
Meet the HP Z VR backpack PC, the word’s first professional wearable VR PC. I gave it a go at NVIDIA’s numerous demo booths and was pleasantly surprised at how lightweight it is while partaking in the VR demos.
It’s no slouch either as it boasts specs like an Intel Core i7-7820HQ processor and NVIDIA Quadro P5200 GPU.
Lastly, this isn’t a system, but check out the futuristic self-driving transport vehicle E/NRIDE powered by NVIDIA’s Drive technology.
To be sure, yes, there’s no driver cabin in this all-electric transport vehicle and its entire length is reserved for freight stowage needs. Cool?
Our articles may contain affiliate links. If you buy through these links, we may earn a small commission.