Nvidia's Monstrous Volta GPU appears

Nvidia's monstrous Volta GPU appears, packed with 21 billion transistors and 5,120 cores
Taking the wind out of Radeon Vega's sails.
Brad Chacos SENIOR EDITOR, PCWORLD | MAY 10, 2017 11:09 AM PT
http://www.techconnect.com/article/...lc=&cid=tcon_nlt_techconnect_daily_2017-05-11

Spoiler

"“We need to find a path forward for life after Moore’s Law,” Nvidia CEO Jen-Hsun Huang said at the beginning of his annual GPU Technology Conference keynote. But Nvidia isn’t hesitant to throw around more iron to make its ferocious graphics processors even more so, as evidenced by the reveal of the first product based on Nvidia’s badass next-gen Volta GPU.

Nvidia’s high-end “Pascal” processors still rule the graphics roost, though AMD’s rival Radeon Vega GPUs are scheduled to launch before the end of June. Volta helps Nvidia take some of the wind out of AMD’s sails before Vega even hits the streets, even if the Tesla V100 GPU is focused on data centers.

This beastly GPU—both in size and capabilities—boasts a whopping 21 billion transistors and 5,120 CUDA cores humming along at 1,455MHz boost clock speeds, all built using a 12-nanometer manufacturing process more advanced than that of Nvidia’s current GPUs. By comparison, today’s Pascal GPU flagship, the 14nm Tesla P100, offers 3,840 CUDA cores and 15 billion transistors. The GeForce GTX 1060 has a quarter as many CUDA cores as the Tesla V100, at 1,280.

To fit all that tech, the Volta GPU in the Tesla V100 measures a borderline ridiculous 815mm square, compared to the Tesla P100’s 600mm GPU. Monstrous.

Volta is “at the limits of photolithography,” Huang said with a smirk, created using an R&D budget of over $3 billion.

Nvidia says it’s redesigned Volta’s streaming microprocessor architecture to be 50 percent more efficient than Pascal’s, which is damned impressive if it proves true. That enables “major boosts in FP32 and FP64 performance in the same power envelope,” Nvidia says. The Tesla V100 also includes new “tensor cores” built specifically for deep learning, providing 12 times the teraflops throughput of the Pascal-based Tesla P100, Huang said. (Google’s also invested in tensor processing hardware.)

The Tesla V100 hits a peak of:

• 7.5 TFLOP/s of double precision floating-point (FP64) performance;
• 15 TFLOP/s of single precision (FP32) performance;
• 120 Tensor TFLOP/s of mixed-precision matrix-multiply-and-accumulate.

The Tesla V100 utilizes 16GB of ultra-fast, 4096-bit high-bandwidth memory to process data quickly. It’s unknown whether Volta-based consumer graphics cards will feature HBM2, however. Radeon Vega does, but the tech is still relatively new and pricey. The GeForce GTX 10-series graphics cards debuted with new GDDR5X technology based on classic memory designs, and SK Hynix recently said that it’s “planning to mass produce the product for a client to release high-end graphics card by early 2018 equipped with high performance GDDR6 DRAMs.”

That HBM2 memory hits 900GB/s speeds, Nvidia says, and the Tesla V100 features a second-gen version of Nvidia’s NVLink technology. At 300GB/s transfer speeds, Huang claims NVLink is now 10 times faster than standard PCIe connections.

If you want to know more about Volta’s datacenter and architecture details, be sure to check out Nvidia’s in-depth Tesla V100 explainer. Look for the Tesla V100 to launch in a revamped version of Nvidia’s pricey DGX-1 system in the third quarter, and more widely in the fourth quarter.

The impact on you at home: None, immediately. But this first glimpse at Volta gives us an idea of what Nvidia’s next-gen GeForce graphics cards will be capable of. Remember, Nvidia revealed its Pascal GPU at GTC 2016 in the form of the Tesla P100, and that full-fat version eventually trickled down into the Titan Xp, with the GeForce GTX 1080 Ti coming damned close. The Tesla V100’s extreme GPU size and strong machine learning focus skews that somewhat when it comes to Volta, though.

One more thing: The GeForce GTX 10-series launched a mere month after Pascal’s GTC reveal, about one year ago. Don’t necessarily expect Nvidia’s Volta-based GeForce cards to launch in the near future—especially if SK Hynix’s GDDR6 memory is indeed headed for Nvidia cards rather than next-gen Radeons.

This story, "Nvidia's monstrous Volta GPU appears, packed with 21 billion transistors and 5,120 cores" was originally published by PCWorld."

Spoiler: Article images

 

GTC 2017: Nvidia gpu technology conference Tesla V100 Volta
Starts at 3:17, Volta comes in at 52:00

 

Nvidia's new Volta-based DGX-1 supercomputer puts 400 servers in a box
The Nvidia supercomputer will ship for $149,000 in the third quarter
http://www.techconnect.com/article/...-supercomputer-puts-400-servers-in-a-box.html

"You won't need to buy a rack of 400 servers if you have one high-powered Nvidia DGX-1 supercomputer with a Volta GPU sitting on your desktop.

The DGX-1 supercomputer -- which looks like a regular rack server -- gets most of its computing power from eight Tesla V100 GPUs.

The GPU, the first one based on the brand-new Volta architecture, was introduced at the company's GPU Technology Conference in San Jose, California, on Wednesday.

"It comes out of the box, plug it in and go to work," said Nvidia's CEO Jen-Hsun Huang during a keynote speech.

But the DGX-1 with Tesla V100 computer is expensive. At US$149,000, it's worth some people's life savings. But Huang encouraged people to order it, saying the box will ship in the third quarter.

The new supercomputer 40,960 CUDA cores, which Nvidia says equals the computing power of 800 CPUs. It replaces the previous DGX-1 based on the current Pascal architecture, which has the power of 250 two-socket servers, according to Nvidia.

Nvidia says the system delivers about 960 teraflops of half-precision -- 16-bit floating point -- performance, which means lower single-precision and double-precision performance. The numbers weren't available, but half-precision performance is considered valuable for machine-learning tasks.

Spoiler

Accompanying the GPUs are two 20-core Intel Xeon E5-2698 v4s running at clock speeds of 2.2GHz. The system has four 1.92TB SSDs and runs on Ubuntu Linux.

The system draws 3,200 watts of power, so don't keep it running all day long, or it'll run up your electricity bill.

Gamers shouldn't get excited about the machine. The DGX-1 with Tesla V100 is perhaps too expensive to be a huge gaming rig; it is instead designed more for machine learning.

GPUs already power machine-learning tasks in data centers, and the Nvidia supercomputer is an example of how GPUs are making applications like image recognition and natural language processing a reality.

Huang said CPUs do not provide enough power for computing, especially for artificial intelligence, which is where GPU fit in.

The Tesla V100 in the DGX-1 is five times faster than the current Pascal architecture, Huang said. It will have new technologies like NVLink 2.0, a new interconnect with bandwidth of up to 300Gbps (bits per second). The GPU has more than 21 billion transistors and 5,120 cores. It also has 900GBps (bytes per second) of HBM2 memory bandwidth.

Nvidia has also included a cube-like Tensor Core, which will work with the regular processing cores to improve deep learning. Nvidia focused on structuring cores to speed up matrix multiplications, which are the heart of effective deep-learning systems. The structure will help align low-level floating-point calculations, which should speed up deep learning.

Huang boasted the GPU offers 120 teraflops of deep-learning performance, though that'll be hard to verify. Standard benchmarking tools don't exist for machine- or deep-learning applications, though development is underway at companies like Google.

The supercomputer works with many high-performance computing and deep-learning frameworks like CUDA, Tensor, and Caffe2.

The graphics company also introduced the DGX Station, which is a smaller version of the new DGX-1. It looks more like a workstation and has four Tesla V100 GPUs, half that of the DGX-1. It is priced at $69,000 and will ship in the third quarter.

Nvidia didn't immediately say if the products will be shipped worldwide."

DGX Station
https://www.nvidia.com/en-us/data-center/dgx-station/

DGX Systems
https://www.nvidia.com/en-us/data-center/dgx-systems/

GPU Cloud
https://www.nvidia.com/en-us/gpu-cloud/home/

Spoiler: Images

 

Nvidia's bleeding-edge Volta GPU: 5 things PC gamers need to know
Reading the tea leaves the and spec sheets.
Brad Chacos, PCWORLD | MAY 11, 2017 3:00 AM PT
http://www.techconnect.com/article/...olta-gpu-5-things-pc-gamers-need-to-know.html

"Nvidia just revealed the most powerful graphics processor in the world yet again. The next-generation “Volta” graphics architecture made its debut at the annual GPU Technology Conference on Wednesday, jumping off Nvidia’s roadmap and into the real world.

But what does that mean for you, a humble PC gamer? Here are five key things you need to know about Nvidia’s Volta GPU.

Spoiler

1. It’s not for you, yet
If you’re looking for hot details about the future of GeForce graphics cards, well, keep waiting. Like the Pascal GPU architecture before it, Volta’s grand reveal comes in the form of a monstrous Tesla V100 GPU built for data centers. CEO Jen-Hsun Huang focused on the hardware’s AI chops and announced a new $149,000 DGX-1 deep learning system with eight of the Volta processors inside. Big spenders get first crack, it seems.

That said, the details Nvidia released about Volta hint at what we might expect from the inevitable next-gen GeForce cards.

2. Volta beats the pants off Pascal
The Tesla V100 rocks 21 billion transistors and 5,120 CUDA cores running at 1,455MHz. That downright dwarfs Pascal’s flagship data center GPU, the Tesla P100, which packs 15 billion transistors and 3,840 CUDA cores running at a slightly faster 1,480MHz maximum clock speed.

It’s impossible to know what GeForce graphics cards based on Volta will look like, but the recently released Titan Xp wound up with a full-blown implementation of the Tesla P100’s Pascal GPU, and it’s the fastest gaming GPU in the world. Heck, the GeForce GTX 1080 Ti sports a slightly nerfed version with 3,584 CUDA cores and it can still game at 4K resolution without breaking a sweat.

It’ll be fun to see how many CUDA cores fit into consumer graphics cards, especially since Nvidia says the Tesla V100 offers 1.5 times the teraflops performance of its Pascal predecessor. For yet another point of reference, this Volta GPU packs four times as many CUDA cores as the GeForce GTX 1060, twice as many as the beastly GTX 1080, and 320 texture units compared to the GTX 1080 Ti’s 224.

3. Cutting-edge tech showcase
Nvidia spared no expense preparing Volta’s flagship. The Tesla V100—and presumably all eventual Volta-based graphics cards—use new 12nm “FFN” manufacturing process technology, an upgrade over the 16nm tech that GTX 10-series GPUs rely on. Smaller transistors mean better performance and power efficiency.

The Tesla V100 also includes 640 “tensor cores” designed to greatly accelerate machine learning, and a second-gen version of Nvidia’s NVLink technology, which the company says surpasses PCIe transfer speeds by a whopping tenfold. None of that matters to PC gamers though.

What could matter is the Tesla V100’s memory. Like the Radeon Fury series and AMD’s imminent Radeon Vega graphics cards, this data center GPU includes high-bandwidth memory technology—16GB of second-gen HBM2, in fact, with peak speeds of 900GB/s. That’s damned fast. But there’s no guarantee it’ll wind up in GeForce cards based on the Volta architecture. The GTX 10-series relies on traditional GDDR5 and GDDR5X RAM despite launching after AMD’s Fury, and it’s no worse for wear because of it. Nvidia will likely ditch HBM2 and stick with the memory we know for at least some of the inevitable Volta-based GeForce cards given the high price of HBM.

4. Volta is massive
Cramming all that technology into a graphics processor is no easy task. The Volta GPU found in the Tesla V100 is massive, measuring a whopping 815mm squared—literally the limits of the GPU fabrication process being used. By comparison, plus-sized GPUs found in Radeon’s Fury cards and recent high-end GeForce chips clock in at roughly 600mm.

In any case, there’s never been a GPU this large in a consumer graphics card before. Whether the this gigantic version of Volta appears in the GeForce lineup of tomorrow depends on Nvidia’s willingness to break new ground. Don’t bet on it. Yields must be abysmal with such a huge chip on a new manufacturing process, and consumers won’t pay anywhere near the premium that businesses will for something this radical.

5. The timing’s no coincidence
The Tesla V100 won’t appear in data centers until the third and fourth quarter, depending on how you’re buying it. Business buyers need time to prepare for big cutting-edge expenditures. But there’s probably another reason why Nvidia’s already showing off Volta: Vega.

AMD’s enthusiast-focused Vega GPUs (and the data center-targeting Radeon Instinct) are scheduled to launch some time before the end of June, with Computex and E3 being two likely launch venues. The Radeon hardware offers some intriguing advances of its own, including a programmable geometry pipeline that casts aside unnecessary work and HBM2 paired with fascinating new “high-bandwidth cache controller” technology. How will Vega compare to the GTX 10-series? Nobody knows—but launching the GeForce GTX 1080 Ti at a comparatively affordable $700 and revealing products based on the Volta GPU now leaves Nvidia well-positioned to weather the potential storm.

The real question is when we can expect to see Volta-based GeForce graphics cards. Volta has appeared on Nvidia roadmaps for 2018 in the past, and memory supplier SK Hynix recently said that it’s prepping cutting-edge GDDR6 RAM to ship in high volume on a graphics card in early 2018—though the company didn’t say whether the graphics card is from AMD or Nvidia.

It seems unlikely that Nvidia would launch a new generation of Volta GeForce cards so soon after the GTX 1080 Ti and Titan Xp’s release. But hey, the GTX 10-series launched a mere month after Pascal’s release at GTC 2016. We’ll hear more about the next-gen GeForce line-up someday, but whether that’ll be sooner or later is very much up in the air.

This story, "Nvidia's bleeding-edge Volta GPU: 5 things PC gamers need to know" was originally published by PCWorld."

 

The interesting take away I see from both NVidia and Intel from the AMD onslaught is that neither company wants to fight on price.

They just bring 'bigger' and 'badder' tech to the playing field with ever increasing price points.

Whether these high prices will be sustainable will be for corp's/businesses and consumers to decide on; and the fate of all these players will be decided accordingly.

 

From a compute point of view... its 20% above Vega's touted 12.5 TFlops (and we don't know if this will be the final stage for Vega).
Granted, Nvidia is using a more advanced manuf. process, but as we know, AMD can pack quite a lot of transistors on the same die space as well (high dense library design anyone?).

This should be interesting... if AMD could take this into account for Vega and further improve on it in terms of efficiency (though at this stage it's all rather murky yet)... is there any possibility that if AMD releases another Vega version at the same time Volta comes out, it might just go up against it?

 

So it was just about 1 year and 1 month ago that Nvidia Announced the Pascal DGX-1 appliance, and we had the GTX 1080 about a month after that.

So if timing holds the same, we could have a GTX 1180 announcement in June, or September if based off of V100 shipping date.

I'll find a better post, but this is one from me about the Pascal DGX-1 from April 5th 2016.
http://forum.notebookreview.com/thr...00m-series-gpus.763032/page-174#post-10233968

Here is a link to an announcement of the GTX 1080 May 6th 2016:
https://blogs.nvidia.com/blog/2016/05/06/geforce-gtx-1080/

Disclaimer, it could be *long* time before we see a GTX 1180 / Volta GPU consumer release, the Volta release schedule might be totally different than the Pascal release, and also depends on the shipping date for V100 hardware.

 

NVidia Volta V100 GPU Detailed: Completely New Architecture

 

SK Hynix is mass producing GDDR6 for a high end GPU release in early 2018, if that qualifies as "long".

 

That's too bad, I don't think Nvidia is going to want to wait that long to release Volta consumer cards. Probably going with DDR5x again

 

You think they'll do 11_0 and first release Titans with 5x and then go with 6 for the Ti and Titan re-release? Might be plausible. Depends on how Vega does I guess.

 

There's usually some ramp up and teething problems with a new memory format, I doubt Nvidia's gonna risk it with Volta, the DDR5x should have some more headroom - maybe even OC it a bit.

Vega's a mystery

 

If that's how it plays out the bigger mystery would be where all that GDDR6 is going. AMD's committed to HBM2 from what I've read.

 

Probably Nvidia, after manufacturing has a chance to ramp up, maybe mid 2018?

Volta Refresh

 

I give it Q1 2019... Nvidia has no need to hurry. They'll milk another Pascal revision or two.

 

Wow, that's way out there, 1.5+ years away... is that for desktop Volta, or Laptop Volta?

 

Desktop. Nvidia will not move on this unless they absolutely need to, if past performance is any indication. I mean, hell, they got three years out of Maxwell.

 

That's a fair point, though I still say they've got something happening Q1 2018, whatever that is.

 

"Volta for gaming, we haven't announced anything. And all I can say is that our pipeline is filled with some exciting new toys for the gamers, and we have some really exciting new technology to offer them in the pipeline. But for the holiday season for the foreseeable future, I think Pascal is just unbeatable. It's just the best thing out there. And everybody who's looking forward to playing Call of Duty or Destiny 2, if they don't already have one, should run out and get themselves a Pascal."
~Jen-Hsun Huang, Nvidia's CEO.

 

This is called, the 1080 TI has no real competition, so we are in no hurry to compete with ourselves. It is too bad Vega was not more of a challenge to nVidia.

 

Agreed, it's too bad the 64 didn't' compete with the 1080 TI. Sure it competes with the 1080 in some games, but the price has sky rocketed anyways.

 

Price is an issue and power draw and heat yet another. 17 months too late and way more than a dollar short.

 

Still it shows they're at least attempting to get back in the game. I'm trying to remain optimistic for a competitive future regardless of current news.

 

Yeah, pretty sure we all are. It's just a shame their competition didn't hit the gaming market. AMD has always had good chips for mining, unfortunately that doesn't do me much good

If Vega 64 reaches about $500 I may snag one up. But considering the 1080 is about ~$530 (for a similar cooler)... I'll just stick with my 5 year old 290x. Might have to go Volta unfortunately.

What does intrigue me is the fact Eurocom is working to bring vega to standard MXM. If they can keep standard MXM alive one more generation... that would be nice

 

I've heard of this, if they can keep the power in line, that would be awesome.

 

Just found this thread, sure there was another Volta thread, but couldn't find it just now. This might be interesting: https://www.notebookcheck.net/Nvidia-reveals-more-details-on-the-upcoming-Volta-GPUs.244332.0.html

Ha, they're wrong about the 32 square inch, what are these notebookcheck.net guys on nowadays! ;-)

 

"The massive 32 square-inch GPU die is flanked by four stacks of HBM2 memory chips cumulating 16 GB. (Souce: Nvidia)"

Wow, imagine all the thermal paste it would take to cover that GPU

 

Ha, it would be obscene! I mean a 32 square inch chip would barely fit in most cases just all by itself (forgetting the rest of the card)! Ha!

 

Looks like Volta might give us in the region of 1.5x the performance of Pascal then - fp64/fp32 is showing as 1.5x Pascal in that table in that notebookcheck.net article above.

 

Wonder how much difference that will mean when it hits mainstream video cards. Wasn't Pascal projected to have more of an advantage over Maxwell than it ended up having too?

 

Isn't that how it's always been lol?

 

True, more wondering how big the hype/reality gap will be.

 

I might have been too far into AMD's camp at the time to notice but I didn't realize the pascal improvements over maxwell was considered a disappointment. The 1070 notebook card is more powerful then every maxwell reference card possibly up to the 980TI at half the TDP. by comparison the 970m is not more powerful than the reference desktop 770

 

That's because Pascal has same IPC that Maxwell had.
Nvidia used a smaller manuf. process for Pascal GPU's which was optimized for higher clock speeds... so, essentially, they shrunk Maxwell, and overclocked as much as it would go while keeping with the same or reducing power requirements.

 

Which makes you wonder what Nvidia could be doing in that market segment if they actually had a competitor pushing them to devote more engineering resources to GeForce.

 

Well, it doesn't help that they are paying developers to optimize for their hardware on the PC and push AMD out of the picture - directly affecting AMD's ability to compete.

They do the same by pushing for proprietary features even though open source solutions existed for a while and are arguably better.
Take TressFX for example. It's open source, simulates hair much better than Gameworks tesselation and it's also easier on the hardware.
Then of course there's FreeSync and OpenCL (dunno if Vulkan can be counted, though it pretty much takes advantage of many DX12 features anyway).

 

Probably focusing their development on premium cards (XX80/XX80Ti) so that people mentally associate the performance difference with all the lower cards, even if those have little cost/performance advantage.

 

It is curious that the 1080 TI was so late to the Pascal party. Perhaps the thought process at Nvidia was was to move the bar just a little bit higher and that would make it easy for them to hold Volta back until GDDR6 and TSMC's latest manufacturing process are ready. Which for now has worked, looking at Vega 64 prices, it'd only make sense to buy one if you can use its compute power, otherwise the 1080 TI is a better buy, based on newegg.com prices. 1070s are overpriced due to miners still, so Vega 56 has a chance at least. Unless the same pricing fiasco happens with it, too. It'll be interesting to see where Vega prices settle over the next month.

 

Hopefully the ether bubble will deflate and give some relief to the market briefly until the next big thing.

 

Volta looking promising, let´s see how this translates to the consumer products.
http://wccftech.com/nvidia-volta-tesla-v100-gpu-compute-benchmarks-revealed/

"The score puts the Tesla V100 in an impressive lead over its predecessor which is something we are excited to see. It also shows that we can be looking at a generational leap in the gaming GPU segment if the performance numbers from the chip architecture carry over to the mainstream markets. "