Nvidia Volta.

[cube]

I already have the InfiniBand switches - of course I would order it.

OK, but if P100 is way too expensive for an organization, and the application does not need IB, and they can manage with Linux, being able to use consumer cards can make the difference, even if they have to pay for some level of support. That's the point.
[doublepost=1494383635][/doublepost]

https://forums.macrumors.com/thread...rence-second-day.2043059/page-4#post-24564759"

"No. The point is if the design also makes sense for desktops for the cards to be cheap."

So, I did not mention any cheap desktop, but that a reverse card design be also interesting for consumers so that organizations are not forced to pay professional-level prices to fill dense servers.

 

[AidenShaw]

OK, but if P100 is way too expensive for an organization, and the application does not need IB, and they can manage with Linux, being able to use consumer cards can make the difference, even if they have to pay for some level of support. That's the point.

P100 is not "too expensive", but it adds no value for our application. We don't need ECC, we don't need FP64. Why pay $7000 for a card that is no better for your application than the $700 card.

The simplest answer is that this expensive blade chassis that packs 20 dual slot cards just isn't a good choice if you don't need high density and high bandwidth fabrics.

There are lots of other systems that will hold two to four consumer cards. There are eGPU boxes that can hold up to 16 consumer cards.

 

[SoyCapitanSoyCapitan]

New category device to be revealed today. Could be:

- pocketable eGPU based on a mobile GPU package

- cryptocurrency miner

- AI/machine learning upgradable modules

 

[AidenShaw]

Nvidia's Success Defies The Pundits

Nvidia (NASDAQ:NVDA) showed no signs of slowing down as it posted 48% y/y revenue growth to $1.937 billion. Gross margin was an admirably rich 59.4%. GAAP net income was $507 million, which was up an incredible 144% y/y. Diluted EPS was $0.79.

Nvidia's GPU business was up 45% y/y to $1.562 billion in revenue, but this was eclipsed by the even more impressive growth of the Tegra Processor Business, which was up 108% y/y and 29% sequentially
...

https://seekingalpha.com/article/4071422-nvidias-success-defies-pundits

http://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-first-quarter-fiscal-2018

 

[cube]

http://www.pcworld.com/article/3195...upport-to-make-you-a-minecraft-superstar.html

 

[koyoot]

15 TF FP32
835 mm2(!)
5120 CUDA cores.

Tesla V100

Registry File size is actually revolutionary in this chip.

 

[cube]

835 mm2(!)

I guess there will not be a $700 version.

$1200 and we are talking about a real Titan.

 

[koyoot]

1200? Nope. Expect breaking 1999$ price tag, if ever this GPU will land in consumer space.

It will not.

 

[Asgorath]

1200? Nope. Expect breaking 1999$ price tag, if ever this GPU will land in consumer space.

It will not.

Source?

 

[koyoot]

Source?

The GPU is slated for HPC and AI markets. Why would it land in consumer market?

 

[Asgorath]

The GPU is slated for HPC and AI markets. Why would it land in consumer market?

Oh, you spoke with such authority that I thought that you had sources to back up your definitive statements on the matter. If this is just your opinion or speculation, perhaps you could phrase it as such?

Personally, I wouldn't be surprised to see Quadro and Titan cards based on GV100 at some point, probably later in the year after they've filled all their supercomputer orders.

 

[koyoot]

Oh, you spoke with such authority that I thought that you had sources to back up your definitive statements on the matter. If this is just your opinion or speculation, perhaps you could phrase it as such?

Personally, I wouldn't be surprised to see Quadro and Titan cards based on GV100 at some point, probably later in the year after they've filled all their supercomputer orders.

Is Titan brand available anywhere else than consumer?

It will not appear in consumer space.

https://devblogs.nvidia.com/parallelforall/inside-volta/

 

[Stacc]

15 TF FP32
835 mm2(!)
5120 CUDA cores.

Tesla V100

Registry File size is actually revolutionary in this chip.

Something interesting is that transistor density on TSMC's 12nm is almost the same as their current 16 nm. If you were to scale up P100 to that size it would have a similar amount of transistors. There are probably gains in efficiency though since clock rates on the two chips are similar but they both stay within the 300 W TDP.

Also they are still using HBM2, so whatever hypercube memory was that was on previous roadmaps must be dead.

 

[AidenShaw]

https://finance.yahoo.com/news/nvidia-launches-revolutionary-volta-gpu-170325537.html

Is Titan brand available anywhere else than consumer?

It will not appear in consumer space.

Volta will appear in the consumer space, but not the GV100 chip. Wait for GV104 and others.

Same as with Pascal - introduced for HPC as the GP100, later in consumer cards as GP104 and GP102

 

[koyoot]

Something interesting is that transistor density on TSMC's 12nm is almost the same as their current 16 nm. If you were to scale up P100 to that size it would have a similar amount of transistors. There are probably gains in efficiency though since clock rates on the two chips are similar but they both stay within the 300 W TDP.

Also they are still using HBM2, so whatever hypercube memory was that was on previous roadmaps must be dead.

12 nm is just marketing name for slightly improved 16 nm process from TSMC. It is in line with low-power process, on which there will be second Volta product coming this year.

https://finance.yahoo.com/news/nvidia-launches-revolutionary-volta-gpu-170325537.html


Volta will appear in the consumer space, but not the GV100 chip. Wait for GV104 and others.

Same as with Pascal - introduced for HPC as the GP100, later in consumer cards as GP104 and GP102

That is exactly what I am pointing at. As for 5120 CUDA core chip, there will be one from Nvidia this round, for consumer space called GV102. But without FP64, and much smaller. GDDR6 also for that GPU.

 

[cube]

Little Crimson

 

[SoyCapitanSoyCapitan]

Can't wait for the buggy beta Mac drivers and run it on 8 year old CPUs )

 

[cube]

Is this the biggest chip ever made?

 

[koyoot]

Yes. Previously biggest one was GP100.

 

[cube]

Yes. Previously biggest one was GP100.

I was thinking about Xeon Phi.

 

[Stacc]

https://finance.yahoo.com/news/nvidia-launches-revolutionary-volta-gpu-170325537.html


Volta will appear in the consumer space, but not the GV100 chip. Wait for GV104 and others.

Same as with Pascal - introduced for HPC as the GP100, later in consumer cards as GP104 and GP102

There were pictures of a PCIe based add in board for GV100 that would be used in the 4 card workstation that was announced.

 

[cube]

There were pictures of a PCIe based add in board for GV100 that would be used in the 4 card workstation that was announced.

But maybe they don't have ports.

 

[Stacc]

But maybe they don't have ports.

What ports?

 

[cube]

What ports?

Video ports. This was not announced as a graphics or general purpose workstation.

 

[Stacc]

Video ports. This was not announced as a graphics or general purpose workstation.

Its got 3x DisplayPort.