Netgear Nighthawk AX12 Wi-Fi 6 router review
Corsair HS60 PRO Surround Headset Review
MSI Radeon RX 5700 Gaming X review
Team Group Cardea II NVMe 1 TB SSD Review
ASRock X570 Phantom Gaming ITX/TB3 review
Promo: URCDKey Autumn Sale Windows 10 Pro for $11
Patriot P200 1TB SATA3 SSD Review
ASRock Radeon RX 5700 XT Taichi OC+ review
Guru3D Rig of the Month - September 2019
KFA2 GeForce RTX 2070 Super EX Gamer Black review
Nvidia Announces PCIe version Tesla P100
Nvidia released the PCI-Express version of the P100 in the form of the Tesla P100. The hpc-card was revealed back in April already however was based on a Mezzanine-connector.
Nvidia is to release two versions of the PCI-Express models with 16GB HBM2 memory, and one with 12 GB HBMs (one stack disabled). The cards can communicate over the NVLINk interconnect with each other. The cards will get a slightly lower boost frequency compared to the Mezzanine NVLINK version, the TDP is lower as well at 250 Watt.
| Nvidia Tesla | ||||||
|---|---|---|---|---|---|---|
| Tesla P100 (Mezzanine) |
Tesla P100 (16GB) |
Tesla P100 (12GB) |
Tesla M40 | |||
| Gpu | GP100 (610mm2) |
GP100 (610mm2) |
GP100 (610mm2) |
GM200 | ||
| Architecture | Pascal | Pascal | Pascal | Maxwell 2 | ||
| Core | 1328MHz | - | - | 948MHz | ||
| Streamp processors | 3584 | 3584 | 3584 | 3072 | ||
| Boost | 1480MHz | 1300MHz | 1300MHz | 1114MHz | ||
| Mem. | 1,4Gbit/s HBM2 | 1,4Gbit/s HBM2 | 1,4Gbit/s HBM2 | 6Gbit/s gddr5 | ||
| Mem bus | 4096-bit | 4096-bit | 3072-bit | 384-bit | ||
| Mem bandwidth | 720GB/sec | 720GB/sec | 540GB/sec | 288GB/sec | ||
| Mem MGB | 16GB | 16GB | 12GB | 12GB | ||
| Half Precision | 21,2 tflops | 18,7 tflops | 18,7 tflops | 6,8 tflops | ||
| Single Precision | 10,6 tflops | 9,3 tflops | 9,3 tflops | 6,8 tflops | ||
| Double Precision | 5,3 tflops (1/2 rate) |
4,7 tflops (1/2 rate) |
4,7 tflops (1/2 rate) |
213 gflops (1/32 rate) |
||
| Transistors | 15,3 Billion | 15,3 Billion | 15,3 Billion | 8 Billion | ||
| Tdp | 300W | 250W | 250W | 250W | ||
| Formfactor | Mezzanine | pci-e | pci-e | pci-e | ||
| Cooler | N/A | passive | passive | passive | ||
| Fab | tsmc 16nm finfet | tsmc 16nm finfet | tsmc 16nm finfet | tsmc 28nm | ||
The 12GB version thus will have a 3072 bits wide memory bus, the 16GB version has the full 4096 bit wide memory bus. These cards are intended for high performance computing of course.
Nvidia Announces the GeForce GTX 1070 and 1080 - 05/07/2016 10:11 PM
Right then, time for a proper news post about it. Nvidia hours ago announced the GeForce GTX 1070 and 1080. Both cards are obviously Pascal architecture based and made on a 16nm FinFet design. The ru...
NVIDIA and Samsung Agree to Settle All Outstanding IP Litigations - 05/03/2016 05:59 PM
NVIDIA and Samsung have agreed to settle all pending intellectual property litigation between the two companies, NVIDIA announced today....
Nvidia announces Tesla P100 data-center GPU - 04/05/2016 07:37 PM
Nvidia just announced a HPC data-center module based on Pascal GPU architecture. The Tesla P100 is a daughter add-in board that holds 150B transistors, the GPU itself however will hold 15B transistors...
NVIDIA announces new bundle with Tom Clancy The Division - 02/17/2016 03:59 PM
Mentioned a few days ago already, now confirmed. From today, 17 February, gamers who purchase a GeForce GTX 970 or above GPU or a notebook with a GeForce GTX 970M or above will receive a free copy of...
Nvidia and AMD Cross Multi GPU Tested In DirectX 12 - 10/26/2015 08:14 PM
Though I’m not quite sure if I’d be using Ashes of Singularity for testing this myself, it is an interesting read. Anandtech posted some benchmark testing a feature in DirectX 12 c...
Denial
Senior Member
Posts: 12340
Joined: 2004-05-16
Senior Member
Posts: 12340
Joined: 2004-05-16
#5292647 Posted on: 06/20/2016 02:23 PM
I wish they had a vendor list of companies that are buying this, for my own curiosity. I know Google has their own Tensor Processing Unit which they claim is the best in the industry for deep learning. I wonder how these stack up.
I wish they had a vendor list of companies that are buying this, for my own curiosity. I know Google has their own Tensor Processing Unit which they claim is the best in the industry for deep learning. I wonder how these stack up.
Cave Waverider
Senior Member
Posts: 880
Joined: 2010-07-25
Senior Member
Posts: 880
Joined: 2010-07-25
#5292661 Posted on: 06/20/2016 03:14 PM
I assume we could see something similar as the PCIe versions for the Geforce Ti and Titan Pascal variants, respectively.
I assume we could see something similar as the PCIe versions for the Geforce Ti and Titan Pascal variants, respectively.
BD2015
Member
Posts: 46
Joined: 2015-08-08
Member
Posts: 46
Joined: 2015-08-08
#5292662 Posted on: 06/20/2016 03:21 PM
Exactly what I hope for and can't wait to.
I assume we could see something similar for the Geforce Ti and Titan Pascal variants, respectively.
Exactly what I hope for and can't wait to.
schmidtbag
Senior Member
Posts: 4411
Joined: 2012-11-10
Senior Member
Posts: 4411
Joined: 2012-11-10
#5292671 Posted on: 06/20/2016 03:45 PM
I'm not sure I understand. Teslas are as similar to TPUs as they are to CPUs. Teslas are meant for high-precision highly-parallel number crunching. The TPUs, to my knowledge, are meant for rapid approximations.
In the server world, you buy what does your workload fastest, and that's why architectures like PPC, SPARC, and AMD's Bulldozer are still relevant. Intel is really the only company that has any interest in general-purpose servers.
I wish they had a vendor list of companies that are buying this, for my own curiosity. I know Google has their own Tensor Processing Unit which they claim is the best in the industry for deep learning. I wonder how these stack up.
I'm not sure I understand. Teslas are as similar to TPUs as they are to CPUs. Teslas are meant for high-precision highly-parallel number crunching. The TPUs, to my knowledge, are meant for rapid approximations.
In the server world, you buy what does your workload fastest, and that's why architectures like PPC, SPARC, and AMD's Bulldozer are still relevant. Intel is really the only company that has any interest in general-purpose servers.
Denial
Senior Member
Posts: 12340
Joined: 2004-05-16
Senior Member
Posts: 12340
Joined: 2004-05-16
#5292674 Posted on: 06/20/2016 03:52 PM
I'm not sure I understand. Teslas are as similar to TPUs as they are to CPUs. Teslas are meant for high-precision highly-parallel number crunching. The TPUs, to my knowledge, are meant for rapid approximations.
In the server world, you buy what does your workload fastest, and that's why architectures like PPC, SPARC, and AMD's Bulldozer are still relevant. Intel is really the only company that has any interest in general-purpose servers.
Tesla's were meant for high precision workload crunching, but with Pascal they added mixed precision (FP16) high performance crunching too. That's why they give an FP16 (Half Precision) number of 21Tflops.
Nvidia specifically added that functionality for deep learning. Now obviously, I'd imagine that an entire chip dedicated to deep learning would perform better than one that is good across a wide variety of stuff. But regardless, I'd like to see how the Nvidia chips would stack up against Google's custom built hardware.
I'm not sure I understand. Teslas are as similar to TPUs as they are to CPUs. Teslas are meant for high-precision highly-parallel number crunching. The TPUs, to my knowledge, are meant for rapid approximations.
In the server world, you buy what does your workload fastest, and that's why architectures like PPC, SPARC, and AMD's Bulldozer are still relevant. Intel is really the only company that has any interest in general-purpose servers.
Tesla's were meant for high precision workload crunching, but with Pascal they added mixed precision (FP16) high performance crunching too. That's why they give an FP16 (Half Precision) number of 21Tflops.
Nvidia specifically added that functionality for deep learning. Now obviously, I'd imagine that an entire chip dedicated to deep learning would perform better than one that is good across a wide variety of stuff. But regardless, I'd like to see how the Nvidia chips would stack up against Google's custom built hardware.
XenthorX
Senior Member
Posts: 2697
Joined: 2014-06-17
Senior Member
Posts: 2697
Joined: 2014-06-17
#5292676 Posted on: 06/20/2016 04:07 PM
I might be flamed but... Can you play games with this card?
I know about all the deep-learning focus, but "what if" ?
I might be flamed but... Can you play games with this card?
I know about all the deep-learning focus, but "what if" ?
schmidtbag
Senior Member
Posts: 4411
Joined: 2012-11-10
Senior Member
Posts: 4411
Joined: 2012-11-10
#5292677 Posted on: 06/20/2016 04:08 PM
Tesla's were meant for high precision workload crunching, but with Pascal they added mixed precision (FP16) high performance crunching too. That's why they give an FP16 (Half Precision) number of 21Tflops.
Ah I wasn't aware of FP16 or Nvidia targeting deep learning. But yeah, I can't imagine that performing as well as a chip dedicated toward it. On the other hand, Google's TPUs are ASICs. They were meant for Google's own specific purposes. What Nvidia is doing is allowing for a general-purpose deep-learning design. So, you sacrifice a little bit of performance but the hardware will still work for other tasks, while also probably being a little more lenient about development. I suppose it's a lot like comparing games developed for consoles vs PC. The console games ditch the graphics APIs, which give them better performance on weaker hardware. Meanwhile, PCs sacrifice performance for freedom of choice.
I might be flamed but... Can you play games with this card?
I know about all the deep-learning focus, but "what if" ?
Probably, but it doesn't have any display connectors.
Tesla's were meant for high precision workload crunching, but with Pascal they added mixed precision (FP16) high performance crunching too. That's why they give an FP16 (Half Precision) number of 21Tflops.
Ah I wasn't aware of FP16 or Nvidia targeting deep learning. But yeah, I can't imagine that performing as well as a chip dedicated toward it. On the other hand, Google's TPUs are ASICs. They were meant for Google's own specific purposes. What Nvidia is doing is allowing for a general-purpose deep-learning design. So, you sacrifice a little bit of performance but the hardware will still work for other tasks, while also probably being a little more lenient about development. I suppose it's a lot like comparing games developed for consoles vs PC. The console games ditch the graphics APIs, which give them better performance on weaker hardware. Meanwhile, PCs sacrifice performance for freedom of choice.
I might be flamed but... Can you play games with this card?
I know about all the deep-learning focus, but "what if" ?
Probably, but it doesn't have any display connectors.
XenthorX
Senior Member
Posts: 2697
Joined: 2014-06-17
Senior Member
Posts: 2697
Joined: 2014-06-17
#5292679 Posted on: 06/20/2016 04:14 PM
Probably, but it doesn't have any display connectors.
lol, ok, i guess not xD
Most expensive PhysX card on the market lol
Probably, but it doesn't have any display connectors.
lol, ok, i guess not xD
Most expensive PhysX card on the market lol
Denial
Senior Member
Posts: 12340
Joined: 2004-05-16
Senior Member
Posts: 12340
Joined: 2004-05-16
#5292681 Posted on: 06/20/2016 04:28 PM
It's also passively cooled by the server fans. So unless you have some wind tunnel case cooling going on, you're going to have a bad time. Then there is also driver support and stuff. Oh and it probably costs over $10,000+ per card.
You probably can but it would require some dedication.
It's also passively cooled by the server fans. So unless you have some wind tunnel case cooling going on, you're going to have a bad time. Then there is also driver support and stuff. Oh and it probably costs over $10,000+ per card.
You probably can but it would require some dedication.
Koniakki
Senior Member
Posts: 2840
Joined: 2009-09-15
Senior Member
Posts: 2840
Joined: 2009-09-15
#5292710 Posted on: 06/20/2016 05:34 PM
It's also passively cooled by the server fans. So unless you have some wind tunnel case cooling going on, you're going to have a bad time. Then there is also driver support and stuff. Oh and it probably costs over + per card.
You probably can but it would require some dedication.
Or simply wait for the gaming Ti variant of the card for 1/10 of the cost.
It's also passively cooled by the server fans. So unless you have some wind tunnel case cooling going on, you're going to have a bad time. Then there is also driver support and stuff. Oh and it probably costs over + per card.
You probably can but it would require some dedication.
Or simply wait for the gaming Ti variant of the card for 1/10 of the cost.
fantaskarsef
Senior Member
Posts: 10706
Joined: 2014-07-21
Senior Member
Posts: 10706
Joined: 2014-07-21
#5292712 Posted on: 06/20/2016 05:38 PM
So my question is, and I know I'm hitting a nerve here probably with asking, but are the Titans and 1080TIs as we expect them going to be based on that GP100 chip, or are we getting some kind of cut down GP102 chip, missing the double precision part and nothing else, as in full single precision and memory bandwidth and everything?
So my question is, and I know I'm hitting a nerve here probably with asking, but are the Titans and 1080TIs as we expect them going to be based on that GP100 chip, or are we getting some kind of cut down GP102 chip, missing the double precision part and nothing else, as in full single precision and memory bandwidth and everything?
Denial
Senior Member
Posts: 12340
Joined: 2004-05-16
Senior Member
Posts: 12340
Joined: 2004-05-16
#5292713 Posted on: 06/20/2016 05:51 PM
I expect it to be different.
There is very little point of doing HBM2 on a gaming card. It's added cost and it yields very little in return, the extra bandwidth is unnecessary, the power savings is minimal compared to GDDR5x on 16nmFF (10-15w at most). The biggest reason is the die size (memory controller on HBM2 is much smaller on die) but I'm not even sure that is worth the cost of manufacturing it over GDD5x.
Further, GP100 has some under the hood architectural changes that didn't get carried over to GP104. In GP100 the SM is split in half and nearly every part of the SM got doubled, registers and all. GP104 isn't split -- they did change the number of SM's per GPC but that's about it.
I expect cut down GP100's used in the Tesla P100 will end up in some other Tesla variant (they are already kind of doing it here with the HBM). I'm not really sure what form the Titan will take. I expect the Ti to have a 384 bit GDDR5x bus and 3840 cuda cores (or less if it's cut down) and probably also run at a similar clockspeed as the 1080 does.
So my question is, and I know I'm hitting a nerve here probably with asking, but are the Titans and 1080TIs as we expect them going to be based on that GP100 chip, or are we getting some kind of cut down GP102 chip, missing the double precision part and nothing else, as in full single precision and memory bandwidth and everything?
I expect it to be different.
There is very little point of doing HBM2 on a gaming card. It's added cost and it yields very little in return, the extra bandwidth is unnecessary, the power savings is minimal compared to GDDR5x on 16nmFF (10-15w at most). The biggest reason is the die size (memory controller on HBM2 is much smaller on die) but I'm not even sure that is worth the cost of manufacturing it over GDD5x.
Further, GP100 has some under the hood architectural changes that didn't get carried over to GP104. In GP100 the SM is split in half and nearly every part of the SM got doubled, registers and all. GP104 isn't split -- they did change the number of SM's per GPC but that's about it.
I expect cut down GP100's used in the Tesla P100 will end up in some other Tesla variant (they are already kind of doing it here with the HBM). I'm not really sure what form the Titan will take. I expect the Ti to have a 384 bit GDDR5x bus and 3840 cuda cores (or less if it's cut down) and probably also run at a similar clockspeed as the 1080 does.
Koniakki
Senior Member
Posts: 2840
Joined: 2009-09-15
Senior Member
Posts: 2840
Joined: 2009-09-15
#5292717 Posted on: 06/20/2016 06:14 PM
I expect it to be different.
There is very little point of doing HBM2 on a gaming card. It's added cost and it yields very little in return, the extra bandwidth is unnecessary, the power savings is minimal compared to GDDR5x on 16nmFF (10-15w at most). The biggest reason is the die size (memory controller on HBM2 is much smaller on die) but I'm not even sure that is worth the cost of manufacturing it over GDD5x.
Further, GP100 has some under the hood architectural changes that didn't get carried over to GP104. In GP100 the SM is split in half and nearly every part of the SM got doubled, registers and all. GP104 isn't split -- they did change the number of SM's per GPC but that's about it.
I expect cut down GP100's used in the Tesla P100 will end up in some other Tesla variant (they are already kind of doing it here with the HBM). I'm not really sure what form the Titan will take. I expect the Ti to have a 384 bit GDDR5x bus and 3840 cuda cores (or less if it's cut down) and probably also run at a similar clockspeed as the 1080 does.
If history is to be trusted(it almost always is) something tells me we are gonna see the same/similar scenario as 970SLI vs 980Ti performance wise.
Which means GTX 1070 will offer equal or a bit more performance over GTX 1080Ti.
Which is almost 980Ti SLI performance in a single card! That's completely bonkers!
I expect it to be different.
There is very little point of doing HBM2 on a gaming card. It's added cost and it yields very little in return, the extra bandwidth is unnecessary, the power savings is minimal compared to GDDR5x on 16nmFF (10-15w at most). The biggest reason is the die size (memory controller on HBM2 is much smaller on die) but I'm not even sure that is worth the cost of manufacturing it over GDD5x.
Further, GP100 has some under the hood architectural changes that didn't get carried over to GP104. In GP100 the SM is split in half and nearly every part of the SM got doubled, registers and all. GP104 isn't split -- they did change the number of SM's per GPC but that's about it.
I expect cut down GP100's used in the Tesla P100 will end up in some other Tesla variant (they are already kind of doing it here with the HBM). I'm not really sure what form the Titan will take. I expect the Ti to have a 384 bit GDDR5x bus and 3840 cuda cores (or less if it's cut down) and probably also run at a similar clockspeed as the 1080 does.
If history is to be trusted(it almost always is) something tells me we are gonna see the same/similar scenario as 970SLI vs 980Ti performance wise.
Which means GTX 1070 will offer equal or a bit more performance over GTX 1080Ti.
Which is almost 980Ti SLI performance in a single card! That's completely bonkers!
WareTernal
Senior Member
Posts: 245
Joined: 2013-09-27
Senior Member
Posts: 245
Joined: 2013-09-27
#5292718 Posted on: 06/20/2016 06:16 PM
I don't think they will give you a list, but they claim they are selling a bunch to the U.S. Department of Energy. In 6 months, you can probably look at the top500 and see if anyone is using them.
http://www.top500.org/
https://www.youtube.com/watch?v=J6oE5Knk8K8
I wish they had a vendor list of companies that are buying this, for my own curiosity.
I don't think they will give you a list, but they claim they are selling a bunch to the U.S. Department of Energy. In 6 months, you can probably look at the top500 and see if anyone is using them.
http://www.top500.org/
https://www.youtube.com/watch?v=J6oE5Knk8K8
Click here to post a comment for this news story on the message forum.
Senior Member
Posts: 2897
Joined: 2008-11-29
Check spelling^