Nvidia Announces Pascal-Based Tesla P100 GPU With HBM2

[N.Broekhuijsen]

Nvidia's new Pascal GPU is aimed at super computing, packing 150-billion transistors, 16-nm FinFET technology and HBM 2.0 memory. Meet the Tesla P100.

Nvidia Announces Pascal-Based Tesla P100 GPU With HBM2 : Read more

 

[gadgety]

Mindblowing numbers there. Finally MadVR will have a hardware platform to match it's capabilities?

 

[turkey3_scratch]

Is this like a Quadro replacement? I'm not so up-to-info on workstation/data center cards.

 

[Snipergod87]

Tesla P100.
All they need to do is add a "D" on the end, then you will have presumably the next Model S variant, the P100D.
I doubt this is a coincidence.

 

[Matt_550]

No HBM 2 until Fall. No units until Q1 2017 (March?). Is nVidia really paper launching a card that is a year off?! Are they THAT desperate!?

Also:
FirePro S9300 X2 = 13.9 TFLOPS Single Precision @ 28nm + HBM 1 & 300w TDP
NVIDIA Tesla Pascal = 10.6TFLOPS Single Precision @ 14nm + HBM 2 & 300w TDP

I guess Pascal really is Maxwell v1.1.

That's a dual GPU card, I'd really hope it would beat a single GPU card. Though not by much at all and after it's single precision the S9300 X2 is a joke. Double precision of 868GFLOPS, embarrassing.

 

[ragenalien]

No HBM 2 until Fall. No units until Q1 2017 (March?). Is nVidia really paper launching a card that is a year off?! Are they THAT desperate!?

Also:
FirePro S9300 X2 = 13.9 TFLOPS Single Precision @ 28nm + HBM 1 & 300w TDP
NVIDIA Tesla Pascal = 10.6TFLOPS Single Precision @ 14nm + HBM 2 & 300w TDP

I guess Pascal really is Maxwell v1.1.

For one please note that the TDP for the new pascal card was not revealed and pascal is on 16nm FinFets not 14nm. Secondly, and this is probably more of something you should try and pay attention to in the beginning, the FirePro S9300 X2 is two Fiji GPU's on one board. Since their clock speed is roughly the same as their singular counterparts that gives a single one 6.95 TFLOPS of single precision performance. Given that 10.6 is SIGNIFICANTLY faster than that I think this should be a wonderful card.

 

[viewtyjoe]

That's a dual GPU card, I'd really hope it would beat a single GPU card. Though not by much at all and after it's single precision the S9300 X2 is a joke. Double precision of 868GFLOPS, embarrassing.

I doubt that the S9300 X2 was built with double precision loads in mind. Why waste design time/space on FP64 performance when your target audience only cares about FP32 performance?

This is a shot across the bow at Intel's compute cards more than anything else, since Intel's products were more aimed at a balanced FP32/FP64 performance.

 

[SpAwNtoHell]

So no pascal in my pc this year but maybe next year Q3 if am optimistic? Where is that credit card to order myself a 980ti as is the best i am going to get for another year or 2... sad

 

[FritzEiv]

So no pascal in my pc this year but maybe next year Q3 if am optimistic? Where is that credit card to order myself a 980ti as is the best i am going to get for another year or 2... sad

I don't think that's what you should take away from this. GTC, and Nvidia's push around GTC, is about more complex workloads, happening at scale, and so this (Tesla P100) is just the first glimpse of Pascal that we are seeing for now.

 

[danbfree]

What about using current, decent efficiency architecture in the new 16nm process as a solid stop gap? I'm assuming the delay is due to new architecture and interface required. At 16nm, current architecture could see significant clock boosts, right?

 

[Guest]

This is a platform to get and play games on DX11 under Windows 7 with more visual details faster than Microsoft AMD/Windows 10/DX12 exclusivity bullshit.

 

[InvalidError]

What about using current, decent efficiency architecture in the new 16nm process as a solid stop gap? I'm assuming the delay is due to new architecture and interface required.

HBM is not really a new interface, it still works fundamentally the same way as DDR except for the much wider interface width.

The reason why Nvidia is not relaunching Maxwell on 16nm is likely because Pascal will launch too soon for Nvidia to recover the millions of dollars it would cost to shrink Maxwell and tying up resources in Maxwell die shrinks could also delay Pascal.

 

[ern88]

Where is the Pascal gaming GPU'S at?

 

[manleysteele]

You can order one today as long as you want eight, already in an enclosure with an included SDK and have $129,000 lying around, gathering dust.

 

[stevenrix]

I'm ready for those cards, i have a multi GPU workstation that can accommodate up to 24 GPUs.

 

[manleysteele]

No way to tell what the individual P100's cost as the enclosure includes a couple of unspecified Xeon processors, system memory and 32 SSD's with controllers.

 

[manleysteele]

Is this like a Quadro replacement? I'm not so up-to-info on workstation/data center cards.

No this is a Tesla card. Think K80 and K40.

Where is the Pascal gaming GPU'S at?

Sometime in Q1, 2017. Of course that's a long time from now. Plenty of things could happen to slip it further or, less likely, advance the schedule.

 

[TJ Hooker]

how about the premiums you will have to pay for new motherboards that have this proprietary nvlink. When PCIe 4.0 specs are already out with of course backwards compatibility. I can't stand companies that block other companies by strong holding developers like Intel did Nvidia is guilty of this as well I guess it is a Oregon way of thinking as their HQ's are both located in Oregon.

NVLink probably isn't relevant for 99% of people. Not when today's graphics cards barely utilize the full bandwidth of a PCIe 2.0 x16 connection, let alone PCIe 3.0 (for normal use like gaming anyway). Pretty sure it's intended use is for high performance computing situations, like people running hardware-accelerated analyses and simulations. Hence the reason it's being promoted in conjunction with a Tesla card, which is meant for workstation PCs.

 

[blazorthon]

how about the premiums you will have to pay for new motherboards that have this proprietary nvlink. When PCIe 4.0 specs are already out with of course backwards compatibility. I can't stand companies that block other companies by strong holding developers like Intel did Nvidia is guilty of this as well I guess it is a Oregon way of thinking as their HQ's are both located in Oregon.

What about the premiums you will have to pay for motherboards that support PCIe 4.0? Besides, PCIe 4.0 is far slower than NVLINK and in workloads that care about the interface's performance, that's a big deal. Also, in the high performance sector, you tend to have to replace the board (and often many other things, if not the entire system) every time it's time for an upgrade.

AMD and Intel have advanced interface technologies they can use to compete with NVLINK if they want to.

 

[InvalidError]

What about the premiums you will have to pay for motherboards that support PCIe 4.0?

I doubt PCIe 4.0 will carry any more of a premium than 3.0 did on the motherboard side of things, which is $0 since many PCIe 2.0 boards supported 3.0 with a BIOS update to support Ivy Bridge CPUs and CPU upgrade to Ivy i5/i7. All it requires from the motherboard is adequate signal integrity and length matching from the CPU to the PCIe slots. Most motherboards likely already meet the PCIe 4.0 specs in a similar manner and would be able to support it on CPU-based PCIe lanes if Intel released CPUs with PCIe 4.0 ports compatible with 170/99 series chipsets the same way Ivy added PCIe 3.0 support to many Sandy-generation motherboards.

 

[manleysteele]

Yeah, what kind of breakthroughs in the world of science, we may see by end of year. Nothing or something game changing!!

No way to tell, but at "only" $129,000, the DGX1 puts a lot of power into the hands of a lot of very smart people who were having to schedule compute time. Now they can just set up, start the program, and come back in a few hours for first results, refine the parameters and run again on the same or the next day. Really a very interesting problem to have.

 

[Elrabin]

how about the premiums you will have to pay for new motherboards that have this proprietary nvlink. When PCIe 4.0 specs are already out with of course backwards compatibility. I can't stand companies that block other companies by strong holding developers like Intel did Nvidia is guilty of this as well I guess it is a Oregon way of thinking as their HQ's are both located in Oregon.

NVLINK is for HPC workloads only. It is extremely doubtful we'll see NVLINK on anything less than a workstation or server class motherboard.

The reason NVLINK exists is for many GPU configurations.

A 2 socket E5 2600v3 or v4 server has 80 total PCIe lanes, 40 per socket.

Something like the Dell C4130 has 4x GPUs and the ability to add two additional PCIe cards for high speed connectvity on top of a RAID card and 2 integrated 1gbe ports.

RAID Card is a x16 slot, 4 GPUs is another 64 lanes and the two additional PCIe cards are x16 each. That's a total of 112 PCI lanes.

Dell had to use a PCIe 3.0 switch to make it work.

NVLINK is a separate interconnect with none of the latency issues of a switch and is a vasty superior option for HPC.

 

[photonboy]

It says "datacenter GPU" right in the picture.

People need to quit assuming every post related to GPU has to do with PC gaming. Not everybody does, but so many don't even seem to read the articles before posting.

 

[turkey3_scratch]

It says "datacenter GPU" right in the picture.

People need to quit assuming every post related to GPU has to do with PC gaming. Not everybody does, but so many don't even seem to read the articles before posting.

I know. Most people don't realize that as gamers they are ants. The server industry is what is actually important in this world. Intel has 99% of the server industry currently, which is an unexplainable advantage it has over AMD.

 

[InvalidError]

Intel has 99% of the server industry currently, which is an unexplainable advantage it has over AMD.

Unexplainable? Datacenters are far more sensitive to performance per watt and performance per rack. If Intel's CPUs have 30% better performance per watt than the next best option, that allows the datacenter to cram 30% more compute in a given floor space, power and cooling budget.