At GTC Europe 2017 in Germany, NVIDIA today has announced the Drive PX Pegasus, a new entry to the Drive PX family of computing modules for self-driving cars. Building off of the previous Drive PX family members, the Driver PX Pegasus is intended to be the next step in self-driving hardware by being the company's first Level 5 system, meaning it's capable of supporting fully automated driving for fully autonomous vehicles. Put succinctly, this is the holy grail of what NVIDIA has been building towards over the last few years.

The Drive PX Pegasus is very much a forward-looking product. While NVIDIA is announcing it today, they won't even have dev kits available until later next year, and any kind of commercial release is farther off still. Consequently, the specifications for Drive PX Pegasus are equally forward looking: the board features two  unannounced post-Volta next-generation discrete GPUs, which will be doing most of the heavy lifting. To put this in context, NVIDIA has only just started shipping Big Volta (GV100) for compute products, and smaller scale Volta GPUs are not expected until 2018, so we're looking at something quite far into the future. Meanwhile, rounding out the package and serving as the hearts of the Pegasus will be a pair of NVIDIA's upcoming Xavier SoCs, which combine an integrated Volta GPU (complete with tensor cores) with an unnamed octa-core ARM CPU design.

With the Drive PX Pegasus, NVIDIA is targeting commercial applications in robo-taxis and driverless long-haul. It will, in turn, coexist with the to-be-launched Drive PX Xavier, NVIDIA's previously-announced small-scale self-driving hardware that essentially packs Drive PX 2's capabilities into a small 30W board and features a single Xavier SoC. Drive PX Pegasus, by contrast, is essentially a next-generation Drive PX 2, utilizing much more powerful SoCs and GPUs than before.

NVIDIA has stated that Drive PX Pegasus will be air-cooled, although it's likely to be right at the edge of what's reasonable. WIth a 500W TDP, NVIDIA is pushing the envelope on performance in part by packing so much hardware into a single board. Not that a car will have any kind of trouble delivering that kind of power, of course, but it's quite the interesting change in car design when we're talking about significant amounts of power being dedicated to a non-mechanical operation.

Getting back to the tech specs, of the few details given, the Drive PX Pegasus possesses 320 TOPS of AI inferencing and combined memory bandwidth of over 1 TB/s. Designed for ASIL D certification, PX Pegasus automotive input/outputs include up to 16 cameras (6 of which are lidar). In addition, next-generation GPUs include NVLink and will come in the SXM2 mezzanine form factor also seen with Tesla V100 (as opposed to MXM like the PX 2). While TSMC remains the foundry of choice, the exact process node was not disclosed, although for the Xavier SoC NVIDIA previously announced that they were using 16 FF+. Particularly for the next-generation GPUs, no other information was given, despite being the first official and public reference to any architecture past Volta; on that note, NVIDIA has not yet revealed any detail of consumer Volta.

At 320 TOPS for a dual GPU plus dual Xavier system, this works out at 130 TOPS per GPU (the Xavier SoCs are already quoted as 30 TOPS at 30W). Meanwhile on power consumption, with Xavier already speced for 30W each, this means we're looking at around 220W for each GPU. In other words, these are high-end Gx102/100-class GPU designs. Coincidentally, this happens to be very close to the TOPS performance of the current Volta V100, which is rated for 120 TOPS. However the V100 has a 300W TDP versus an estimated 220W TDP for the GPUs here, so you can see where NVIDIA wants to go with their next-generation design.

NVIDIA DRIVE PX Specification Comparison
  DRIVE PX Pegasus DRIVE PX Xavier DRIVE PX 2
SoCs 2x Xavier Xavier 2x Tegra X2 "Parker"
Discrete GPUs 2x Post-Volta N/A 2x Unknown Pascal
CPU Cores 16x NVIDIA Custom ARM 8x NVIDIA Custom ARM 4x NVIDIA Denver &
8x ARM Cortex-A57
GPU Cores 2x Xavier Volta iGPU
& 2x Post-Volta dGPUs
Xavier Volta iGPU
(512 CUDA Cores)
2x Parker Pascal iGPU & 2x GP104?
DL TOPS 320 TOPS 30 TOPS N/A
FP32 TFLOPS N/A N/A 8 TFLOPS
TDP 500W 30W 250W

Rounding out the specifications, NVIDIA has told us that Pegasus units will be air cooled, not liquid cooled like the 250W Drive PX2 systems. Meanwhile as part of the Pegasus system, NVIDIA has stated that these also need to support failover, but how the failover occurs (either by NVIDIA hardware or other) is up to separate customers in their designs.

In terms of availability, PX Pegasus development kits will be available for select automotive partners in the second half of 2018. However it was not clear if this sampling involved full dev kits with prototype-grade silicon, or an emulation-style system with current-generation hardware to kickstart development. Otherwise, given the long testing and validation cycles in the automotive industry, Drive PX Pegasus may not show up in vehicles for several more years. Last year NVIDIA announced their Level 4 autonomous hardware, but we are only now seeing Level 3 vehicles being showcased in early development for deployment, emphasizing that embedded time scales are much longer than consumer hardware. NVIDIA has stated that their automotive platforms are designed for typical embedded-level longevity, in this case their 'minimum' is a decade of use.

Finally, it's worth noting that this year’s conference is just the second annual GTC in Europe, and going forward NVIDIA has confirmed that GTC Europe will be NVIDIA’s primary automotive-related event. In that vein, NVIDIA has also announced a partnership with Deutsche Post DHL Group and ZF in testing autonomous delivery trucks, starting in 2018.

Source: NVIDIA

Comments Locked

31 Comments

View All Comments

  • XabanakFanatik - Tuesday, October 10, 2017 - link

    You also have to account for the fact that people can only work finite shifts and require rest. An automated vehicle can operate 24/7 without taking breaks, short of refueling and for maintenance or repairs.

    You could potentially get 2-3x the work from a single vehicle operating around the clock.
  • jwcalla - Tuesday, October 10, 2017 - link

    It'll be interesting to see how much they'll have to pay out in lawsuits each year.
  • Yojimbo - Wednesday, October 11, 2017 - link

    In the beginning I'm sure there will be an increase in lawsuits, before the precedents are set, as some people try to take advantage of any situation for financial gain. But, the regulators will look at the situation analytically, because they already do that professionally, and the court precedents will eventually fall in line with that analysis. Then assuming the technology is good, in the long run the amount of money spent on lawsuits, both pay outs and general legal costs, will be lower than it is now. That's because these autonomous vehicles will only be successful if they are safer than human drivers.
  • Qwertilot - Tuesday, October 10, 2017 - link

    Its a little bit interesting how quiet they've turned about post Volta stuff - have to imagine that the big post volta compute GPUs will come as soon as they're remotely practical in a technical sense so it won't be so far off.

    Presumably its because of the quite large disconnect developing between their consumer & compute lines.
  • Yojimbo - Tuesday, October 10, 2017 - link

    Yeah, NVIDIA seems to have recently stopped pushing their architecture roadmap and has started pushing a roadmap for various platforms, most notably automotive platforms.

    Well, since they have such a dominant position in their other markets it makes sense that they don't need to, or even want to, sell the future. On the other hand, self-driving vehicles have little to no market so far, and NVIDIA is trying to attract customers to their platform now for a future market, so it makes sense for them to be selling the future.
  • edzieba - Tuesday, October 10, 2017 - link

    I spy with my little eye, an array of discrete memory chips around the notional 'post-Volta' cores on the SXM2 cards rather than integral HBM as on V100.
  • BurntMyBacon - Tuesday, October 10, 2017 - link

    Good catch. GDDR6 perhaps?

    I have to wonder if a silicon interposer is durable enough to survive the shock and vibe stresses in a automotive vehicle.
  • thesenate - Tuesday, October 10, 2017 - link

    There are only six of them, which suggests that the GPUs are likely Gx106 (192-bit wide interface)
  • Yojimbo - Tuesday, October 10, 2017 - link

    Good catch. The Drive PX 2 also does not use the GP100, which is the Pascal-generation data center GPU that uses HBM 2. The power budget for the Drive PX Pegasus is twice that of the Drive PX 2, which makes it seem like they'll be using perhaps a Gx104-grade, or maybe even a Gx102-grade, GPU this time around.

    One would assume that the discrete GPUs used in Pegasus have Tensor Cores or some evolution or replacement of them. So the GPU not being the big data center iron implies that NVIDIA will be bringing Tensor Core-functionality to at least some GPU other than the Gx100 part in the post-Volta generation.
  • Yojimbo - Tuesday, October 10, 2017 - link

    "One would assume that the discrete GPUs used in Pegasus have Tensor Cores or some evolution or replacement of them. So the GPU not being the big data center iron implies that NVIDIA will be bringing Tensor Core-functionality to at least some GPU other than the Gx100 part in the post-Volta generation."

    Unless, of course, NVIDIA puts an NVDLA on the package with the discrete GPU.

Log in

Don't have an account? Sign up now