http://wccftech.com/nvidia-pascal-specs/



@+
*_*

---

Déjà vu on PrimeGrid, Zarck is everywhere ;) Thanks Zarck!

---

[CSF] Thomas H.V. Dupont
Founder of the team CRUNCHERS SANS FRONTIERES 2.0
www.crunchersansfrontieres

Will be interesting to see which versions of Generation Pascal (GP) ship first.
Last time round the GTX750 and GTX750Ti Generation Maxwell (GM) cards turned up early (Feb 2014) compared to the main high end GM cards GTX970 and GTX980 (Sept 2014) and a Titan Z (GK) turned up after the GM107 releases. In Mar 2014 mobile variants of GM also turned up GM108 and GM107 in the GF800M series. This all made good business sense so I'm expecting something similar this time round.

Also worth noting that the GeForce 700 line-up (including Titans) summed 19 standard + OEM variants but the GF900 series only has 6, so far...
Probably at least another 2 or 3 months before anything GM ships to POS though.
About a month ago I read speculation of a GTX950 SE or LE (Second Edition or Light Edition), but nothing since. Could have been an Engineering Sample (ES) based on a GTX950 GM card but intended for GP, suggesting something like a GTX950Ti might turn up based on GP, say GP107, just ahead of a couple of similar mobile GP models.

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

http://wccftech.com/nvidia-pascal-specs/



@+
*_*

It looks like on the surface these cards are going to be twice as fast as their predecessors, while using a little less power.

But here is the catch, how much bigger is the WDDM lag going to be?

Also, on high CPU dependent WUs like the GERARD A2ARs, how much bigger is that lag going to be?

I guess, you have to take the good with the bad.

It looks like on the surface these cards are going to be twice as fast as their predecessors, while using a little less power.

Based on those specs I would guess a performance boost of around 45% but in terms of performance/Watt it could be around 2.2times better.

But here is the catch, how much bigger is the WDDM lag going to be?

There might be a lot of catches and things we don't know about yet. For example the acemd apps performance might not scale well on these high density cards. That's already the case with the GTX Titan X and GTX980Ti (to a lesser extent). The 7010 to 8000MHz DDR would reduce the Memory Controller Unit burden somewhat but probably not sufficiently to prevent a bottleneck in itself. The architecture might sort that out, or not... The memory bandwidth probably wouldn't be an issue with HBM2, but that seems to be limited to the supposed X80Titan, which might be a bit different in other ways too (double precision), and >>pricier.

Also, on high CPU dependent WUs like the GERARD A2ARs, how much bigger is that lag going to be?

Who knows what architectural advancements and cuda magic might reduce such CPU dependency?

I guess, you have to take the good with the bad.

I'm interested in what the NV-Link will bring to real world crunchers. The possibility of adding up to 8 GPU's via an NV-Link sounds like smallish devices might be great connected up, assuming the reliance on the CPU isn't as big.

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

Déjà vu on PrimeGrid, Zarck is everywhere ;) Thanks Zarck!

Souvent présent ici ->

http://forum.boinc-af.org

@+
*_*

---

Sincères salutations de toute l'équipe CRUNCHERS SANS FRONTIERES à L'Alliance Francophone ! :)

---

[CSF] Thomas H.V. Dupont
Founder of the team CRUNCHERS SANS FRONTIERES 2.0
www.crunchersansfrontieres

Wonder what they will cost also.

http://www.gputechconf.com/

Keep an eye out during the week of April 3rd for real Pascal Information. At 2015 GTC (GM200) Titan X launched. Maybe the GPU-Z database reveals a real Pascal soon. Si Software benchmarking database a good place to find new GPU's.

skgiven wrote:

Who knows what architectural advancements and cuda magic might reduce such CPU dependency?

A year and half ago in the "Maxwell now" thread I asked:

Has dynamic parallelism (C.C 3.5/5.0/5.2) been introduced to ACEMD? Or Unified Memory from CUDA 6.0? Unified memory is a C.C 3.0+ feature.
Quoted from newest CUDA programming guide-- "new managed memory space in which all processors see a single coherent memory image with a common address space. A processor refers to any independent execution unit with a dedicated MMU. This includes both CPUs and GPUs of any type and architecture. "

In you're opinion: how can GPUGRID occupied SM/SMX/SMM be further enhanced, and refined for generational (CUDA C.C) differences? Compatibility is important, as is finding the most efficient code path from CUDA programming. How can we further advance ACEMD? CUDA 5.0/PTX3.1~~~>6.5/4.1 provides new commands/instructions.

GPUGRID developer Matt (MJH) replied:

We have cc-specific optimisations for each of the most performance sensitive kernels. Generally don't use any of the features introduced post CUdA 4.2 though, nothing there we particularly need.
I expect the GM204 performance will be marked improved once I have my hands on one.

One difference is Maxwell's L1 cache size 80kB per SMM (SM 5.0/5.2) while Kelper L1 16kB per SMX (SM 3.0/3.5) possibly configured up to 48kB per SMX.
Maxwell L2 cache also larger than Kelper. (See Maxwell now thread for more info and CUDA toolkit. )

Pascal likely has a larger L1/L2 cache and more efficient instruction throughput as well more registers per SMP and per thread.

If the DP ratio is 1/3 than upping an SMP might be back to Kelper's 192c SMX size. Pascal with a 1/3 DP to SP ratio a 6 [32c] blocks per (1) warp schedulers for each 32c block or maybe (3) 64c blocks per SMP with 2 warp schedulers per block could work.

Maxwell's current 4 [32] blocks per SMM with One warp scheduler for each 32c block would only work with 1/2 or 1/4 and 1/8 DP to SP ratio.

Some rumors are pointing to a 1/3 ---> 1/24 (consumer Geforce GPU) DP ratio Pascal design. Although it's also possible the Tesla or Titan is 1/2 or 1/4.

Pascal could be 8 [32c] block with 8 warp schedulers per 256 core SMP - either 64/128DP cores per SMP at 8 or 16DP per 32c block. It possible a block can grow to 64c with 16 or 32DP in each block - 256c SMP with (4) 64c blocks.

Pascal "SMP" possible configurations with 16 bit cores (without TMU/ROP counts):

--- 256c SMP (4) 64c block - 8 warp schedulers - 2 warps per 64c block - 8/16/32 DP per block. 1/8 or 1/4 or 1/2 DP ratio. 32 bits cores (2x16bit)

--- 256c SMP (8) 32c block - 8 warp schedulers - 1 warp per 32c block - 8 or 16DP cores per (32c) block at 1/4 or 1/2 SP to DP ratio

--- 192 core SMP (6) 32c block - 16 DP core per 32c block for Tesla/Titans (96DP in an 192c SMP at 1/2 ratio) 1 warp scheduler per 32c block.

--- 192c SMP (3) 64c block - 32DP per block - 96DP core per SMP - 1/3 DP to SP ratio. Consumer's GPU knocked down to 8DP core per 64c block. 2 warps per 64c block.

--- 192c SMP with similar design as Kelper: 64 DP core per SMP with 32c block per warp.

--- 192c SMP with (4) 48c blocks (superscalar) 16bit cores (128 or 256bit SIMD lane) and (32) 32bit cores per block - 12DP core per block - 1/3 DP ratio

--- 128c SMP (2) 64 blocks and 2 warp schedulers per block - 16 or 32 DP cores per block - 1/4 or 1/2 DP ratio. Consumer's GPU's 8 DP cores per block.

--- 128 SMP (4) 32 blocks and 1 warp scheduler per block - 8 or 16 DP cores per block - 1/4 or 1/2 P ratio. Consumers GPU's 4 DP per block.

Maxwell's (4) 32bit 32c block at 1 warp per block (4 warp schedulers per 128SMX) design much more efficient than Kelpers 192c 4 warp schedulers SMX.

Kelper is superscalar 4 (32c) vec32 sets and 2 (32c) vec16 design per SMX. Maxwell is 32vec only so scalar. The 8 64bit core per SMX was trimmed to 4 per SMM in Maxwell.

Kelper GK104/110 golden 24/7 OC mark is 1.250/1.3GHz - GM107 1.4GHz - GM204/GM206 1.5GHz. 16nm GP104 probably able to equal GM107 clocks. I'm going to enjoy learning how 16nm transistors react when overclocked with water cooling and air while comparing temperature/power/OC 16nm profiles to 28nm Maxwell.

??Pascal compute Capability or "SM" version ??:
CUDA_ARCH=6.0
CUDA_ARCH=6.1
CUDA_ARCH=6.2
CUDA_ARCH=6.3

http://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html
http://docs.nvidia.com/cuda/parallel-thread-execution/index.html
http://docs.nvidia.com/cuda/gpudirect-rdma/index.html

7.5 CUDA has numerous updates. CUDA 8.0 even more so.

I'm interested in what the NV-Link will bring to real world crunchers. The possibility of adding up to 8 GPU's via an NV-Link sounds like smallish devices might be great connected up, assuming the reliance on the CPU isn't as big.

https://devblogs.nvidia.com/parallelforall/how-nvlink-will-enable-faster-easier-multi-gpu-computing/

Nv-livk whitepaper available.

NV-link (GPU >><< GPU) and (GPU >><< host) 5 to 12x more bandwidth than PCIe3.0. For IBM POWER8 and possible ARM CPU's they'll be CPU <<>> GPU <<>> GPU nvlivk while maybe (PCIe4.0?) Skylake Xeons get GPU <<>> GPU nvlink only. And maybe Kabylake i7/i5 also receive GPU <<>> GPU.

Programs that require PCIe bandwidth can be issue with Maxwell. (Kelper not so much)
A tip: run CUDA-Z without and with ACEMD processes and it will show the types of memory performance being affected.

Until consumer nv-link is mainstream: Skylake z170 and upcoming Kabylake z200 series chipsets platform offer the highest consumer WDDM OS ACEMD performance. Linux and XP also see benefits with Skylake compared to Haswell or Ivy Bridge.

Skylake CPU upgraded to DMI3 while Haswell/Ivy a DMI2 link. (doubling the bandwidth rate) PCIe/DMI clocks are also decoupled in Skylake with new (fclk) running 1 GHz or above. GPU runtimes with skylake chipsets are mostly faster with lower CPU times than Haswell.

(Some risers have an area where a DIY PCIe 6pin can soldered) on a Powered 15pin with (3 12V pins) SATA to 4pin Molex riser Maxwell GTX750 GPU runtimes lose around 50% performance running on a x1 PCIe 2.0 compared to x4 PCIe 3.0.

My (800MHz) 1.6GHz DDR3 64bit memory GT630 (384cores) secondary crunching GPU on the PCIe2.0 x1 slot > USB3.0 riser is around 30% slower than my 384 core GT650m GDDR5 128bit (1GHz/2GHz/4GHz) memory interface that's connected to PCIe2.0 x8. (My 970's show a 10% runtime difference with PCI3.0 x4 - vs. - x8 for certain WU.)

A note about power consumption:

Primegrid's Genefer opencl (4.3 PTX model) OCL3 and about to be released OCL4 puts an OCed 970 at 230W running n=21/22 WU. A n=20 WU will push be 210W with clocks at 1.4GHz or slightly above. Genefer uses the funnel shift instruction which is a C.C 3.5+ feature.

Having more power phases on a GPU and motherboard help's keep electrical power and heat lower. I find this important if one's computer 24/7 for years at a time crunching.

Both my GTX970's ran ACEMD at (5teraFLOPS) 1.5GHz core clock (145W to up 165W) - the Zotac 13 phase 970 consumes 8-15W less power on all ACEMD WU than an EVGA AVX2.0+ 3973 model with 8 phases when both GPU are on PCIe3.0 x8 or x4) WDDM isn't fully utilizing the ACEMD OC scaling.

My current MB is an all digital 20 phase (a refub z87 MSI Mpower Max I picked up for 79$ last July) that use's less power on CPU than my former 16 phase z97 Mpower (also bought as a refub for 80$) by about 3-7W and around 10-14W than a 12 all digital phase ASUS board I tried out. If you're building a Multi GPU board - a top notch phase count will help with long-term power consumption as does a custom PCB GPU compared to reference design. A 1200W (platinum) PSU with 102 AMP single 12V rail also helps.

@NV-Link: it's meant for servers, so don't expect any benefit for crunchers anytime soon.

@Price: the 1st card from the 28 nm generation provided lower power consumption and modern features, but were priced about as high as the similarly performing cards of the prior generation. I.e. there was no improvement of the price/performance ratio until about 1/2 to 3/4 years availability. The reason was limited fab capacity for the new process and initially higher costs (due to inferior yields), as well as refunding some R&D. Since 14/16 nm is going to be more expensive per area even at comparable yield I expect this situation to repeat, maybe even a bit worse (for us).

@WCCFTech: they're wrong so often that I call them "WTF Tech" :p

Sometimes they're right, but those specs don't make sense in too many points:

- just 1000 MHz for the X80? That's significantly below the GTX980, despite a major fabrication technology improvement. Unless the 16 nm process is extremly clock-unfriendly (and the mobile chips sugget otherwise) this looks wrong.

- the bigger chips clock higher? It's always been the other way around for technical reasons.

- X80Ti and X80 Titan based on the same flag-chip GP100, yet use different memory interfaces? And the smaller one comes with 1/6 of all units deactivated? If you're building a massive GPU (and 6000 shaders is massive!) every additional mm² hurts. The last thing you want on such a chip is a 2nd memory interface going unused. A 512 bit GDDR5 interface at extremly high clock speeds (8 GHz) eats a significant amount of die space. I suspect they'd rather build 2 different GPUs for that and equip one with strong FP64 horsepower for additional differentiation.

MrS

---

Scanning for our furry friends since Jan 2002

Based on shipping manifests there will be a Big card and a range of smaller cards. Logs show a range of GPU's with 'Insurance' values of between 30,000INR and 56,000INR and another GPU costing around 200,000INR (£2K). Assumes 699 represents the Pascal architecture, which might be wrong as it's also listed in mobile quadro's.

Date 	HS Code 	Description 	Origin Country 	Port of Discharge 	Unit 	Quantity 	Value (INR) 	Per Unit (INR)
26-Mar-2016	84733092	COMPUTER GRAPHICS CARDS, 699-1G413-0000-000 / NOT FOR SALE	United States	Banglore Air Cargo	NOS	2	61,380	30,690
26-Mar-2016	84733092	COMPUTER GRAPHICS CARDS, 699-1G610-0000-000	China	Banglore Air Cargo	NOS	5	291,803	58,361
26-Mar-2016	84733092	COMPUTER GRAPHICS CARDS, 699-1G411-0000-100	United States	Banglore Air Cargo	NOS	5	238,085	47,617
22-Mar-2016	84733092	COMPUTER GRAPHICS CARDS, 699-5G418-0503-000	United States	Banglore Air Cargo	NOS	2	483,089	241,544
22-Mar-2016	84733092	COMPUTER GRAPHICS CARDS, 699-5G418-0503-000	United States	Banglore Air Cargo	NOS	6	1,221,994	203,666
22-Mar-2016	84733092	COMPUTER GRAPHICS CARDS, 699-1G411-0000-100	United States	Banglore Air Cargo	NOS	15	657,083	43,806
19-Mar-2016	84733092	COMPUTER GRAPHICS CARDS, 699-1G411-0000-100 / NOT FOR SALE	United States	Banglore Air Cargo	NOS	2	112,473	56,237
18-Mar-2016	84733092	COMPUTER GRAPHICS CARDS, 699-1G411-0000-100	China	Banglore Air Cargo	NOS	10	381,192	38,119
16-Mar-2016	84733092	COMPUTER GRAPHICS CARDS, 699-12914-0000-100	China	Banglore Air Cargo	NOS	1	90,020	90,020

GM first appeared in the form of a GTX750Ti and GTX750. That was way back in Feb 2014. It’s now over 2years since the last Generation of NV GPU hit the streets. That was a successful launch within the entry level gaming GPU market and was followed up with similar mobile variants. 7 months later the high end cards (980 and 970) arrived - too long IMO.

This time I'm expecting a similar but not identical approach; entry level to mid range gaming cards to appearing first, as that's the biggest market area and NV will want to flood it before AMD release their 14nm cards. Similar mobile cards to follow and some time later the bigger cards. I'm not really expecting a replacement for the GTX980 or GTX980Ti yet as the big card could be a GP Quadro, GTX1080Ti, GTX 10-Titan, or even a Tesla. That said, and going by lots of GPU pairs shipping, a dual mid-range card based on Pascal could well outperform high end GM GPU's such as the GTX980 and offer many an early upgrade route. I'm expecting enough enticement for GTX700 card holders to buy sooner, rather than later. Just releasing entry level cards wouldn't do that, but mid range cards should.

Good chance Jen-Hsun Huang, NV’s CEO, will announce something GP when he speaks at the GPU Technology Conference on 5th April.

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

BTW was there any dual GTX750Ti variant ever released?

Not that I'm aware of. The GTX750Ti isn't SLI capable, so I doubt that a card would be built on GM107. Not being Sli capable probably improved the card's performance/Watt. Perhaps a dual GTX960 would be viable, but with GP so close I doubt anyone would want to build it.

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

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

Pascal Tesla P100 revealed at GTC (available 2017Q1 for OEM):

21.2 teraFLOPS FP16
10.6 () FP32
5.3 () FP64
3840 CUDA cores
GPU 15.3 billion transistors / 610mm²
4MB L2 cache / 14MB shared register file (6MB Maxwell GM200)
NVlink / HBM2
Unified Memory
Compute Preemption
Compute Capability 6.0
core clock 1328MHz / boost 1480MHz
64 cores per SM / (2) 32core blocks (warp for each block - dispatching two warp instructions per clock)
32 (Double precision) cores / 16 load/store / 16 SFU per SM
eight 512-bit memory controllers (4096 bits total)
16nm FinFET
6 Graphics Processing Clusters (10 SM's per GPC)
300W TDP

https://devblogs.nvidia.com/parallelforall/wp-content/uploads/2016/04/gp100_SM_diagram-624x452.png

https://devblogs.nvidia.com/parallelforall/wp-content/uploads/2016/04/gp100_block_diagram-1-624x368.png

https://devblogs.nvidia.com/parallelforall/cuda-8-features-revealed/

CUDA 8 will be available in August 2016 and there will be a release candidate available around June.

Well, this is not the product we are waiting for.

That GP100 is the Tesla (or similar manifestation) which doesn't concern most crunchers. They will probably be subject to pre-contracts and ALL be going to data centres first (AKA Titan 3 years ago), then to OEM's (maybe in the form of Quadro's) and then possibly/eventually in the form of a GTX1080i to everyone else, if it makes sense in 2017 to do that.

Would have been nice to see a midrange card announced 1050Ti or 1060 but that didn't happen.

For now, the GPU's we are interested in will likely be GP104 and down; from the GTX1080 (or whatever name it will be) down to perhaps a GTX1040.
So while there probably won't be a direct replacement for the GTX980Ti any time soon, there will likely be smaller models (GTX1080) that might offer up similar or competitive performances to the GTX980Ti, and if not it would be highly likely that two lesser GP cards would outperform a GTX980Ti. Even today, two GTX970's (£360) match a GTX980Ti (£520). The 980Ti didn't even ship until after the Titan X. While the GTX980Ti is a flagship gaming card and is only 10months old it was almost an afterthought in the GM line-up and never really made financial sense.

The latest rumour suggests a GTX1080 and 1070 will turn up in June (with an announcement between May 31st and June 4th at Computex, or not...), but mass shipment probably wouldn't happen until July 2016. The same rumours also suggest 7400M to 7900M transistors (not far off a GTX980Ti's 8000M and well over a GTX980's 5000M). Even at 74000M transistors (92.5% of a GTX980Ti's) it would only take a slight architectural improvement to match a GTX980Ti, or an 8% shader clock boost. That said, there could also be bottlenecks and architectural constraints that would inhibit performance and it probably won't work here straight out of the box.

I guess if you really want another card now, your electric is cheap, and you are not bothered about it losing value or being one of the fastest cards around in 6months then a GTX980Ti might still be an option but it's not something I would go for.

The best time to buy a GTX10xx to crunch here would be when we know they work, then when AMD release their 14nm line range, which will probably be some time after NV's launch (price drop).

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

I picked up a new Asus GTX 980 TI on Ebay for £450 free P&P. I figured that it would be a while before a competitor to come out and then for GPUGrid to make it work.

At £450 I should be able to resell at not too much of a loss.

As a by the way, graphics cards go for crazy prices on Ebay had to wait ages to get a price like £450

Completely of track,

"The Druids Nephew" EW in The Grand National

If your money's backing GP to be a late runner with poor odds of an early app appearance, £450 was a decent punt for an each way bet and a heads length better than sticking one on the nose at £520 so to speak.

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

HaHa, well done SK :-)

This GP100 is intended for professional use. I don't think it will be ever released in a form of a GeForce card. It's a "replacement" for the GK110 (GTX Titan, Titan Black, Titan Z), as there was no such chip in the Maxwell product line. The lesser Pascal chips we are waiting for probably won't have that many DP units, NVlink and HBM2. The leaked specifications are surely inaccurate regarding the clocks, as it could be as high as of the GP100. By using 16nm technology theoretically it is possible to have 3 times as much components over the same area as using 28nm (28/16=1.75; 1.75^2=3.0625), but I don't think NVidia wants to produce that large chips for the gaming market (to achieve higher yields), so I expect physically smaller chips than the GP100 in the high-end segment of the gaming cards. Still they could be twice as fast as the GTX980Ti, which is pretty enough (depending on what AMD will have). The effect of the WDDM overhead could be even more deteriorating than on the present high-end cards, also maybe there won't be Windows XP drivers for the Pascal series at all, and in this case I will have to switch to Linux on some of my hosts.