Nvidia Pascal Announcement

GP100 is so different from the other Pascals (this is obviously assuming GP102 shares GP104 DNA, but why would they build more or less two similar sized chips [GP100, GP102], if it didn't?), that at least I don't see even prosumer-versions of GP100 happening.
Because they can build it as an in-between size of GP104 and GP100 and use it in a similar way to the GK110 that spanned Tesla/Quadro/Titan.
It would have less DP than the GP100.
The challenge I see for Nvidia is at some point they will need to create a Pascal with greater FP32 and FP16 than the P100, this should be feasible but they are pushing P100 into the DL crowd currently as well and so might cause some backlash, unless they have been made aware of potential products - I doubt it would have much of a criticism in the business world but putting it out there, especially if the 5kish price is correct for the lowest PCIe P100 model.
And they are probably aware, obviously the Swiss National Supercomputing Center knew about the plan/timeline of producing the PCIe model when the NVLink P100 was announced.

Cheers
 
Older article with some interesting info ...
Nvidia’s architects, on the other hand, chose to increase focus on increased performance in the Pascal GPU architecture, which lead to an increased design complexity, transistor count, and die size of the “data center” version of Pascal. With over 15 billion transistors and a die size of 610 mm^2, this first Pascal GPU is one of the largest processors ever manufactured, and the largest so far in a 16 nanometer FinFET process.

The Pascal GPU implements a scalable compute pipeline. Nvidia organizes their CUDA cores into Streaming Multiprocessors, or SMs, which makes it easy for software developers to code and debug parallel processing resources at manageable scale. Pascal’s 32-bit single precision floating point (FP32) core design supports performing two simultaneous half precision floating point operations (FP16) at the same clock speed as one FP32 operation, using the same compute path. Pascal implements FP32 and double precision floating point (FP64) units in a 2:1 ratio. So, in the time one FP64 operation can be executed, the Pascal architecture can execute four simultaneous FP16 operations. However, only one type of FP instruction – FP16, FP32, or FP64 – may be executed simultaneously in within a single Pascal SM.

Note that Nvidia’s FP16 compute format is fully compliant with IEEE 754-2008, using round-to-nearest-even for all arithmetic. FP16 fully supports subnormal values and they run at the same speed as normalized values.

Also new to the Pascal architecture is unified memory across CPU and GPU physical and virtual memory. A 49-bit virtual address space enables GPU and CPU memory to exist in the same address space while a hardware “page migration engine” globally manages page faults across the unified memory space.

Nvidia also added “compute preemption” to the Pascal architecture – which is a fundamental enabling feature for an operating system or virtual machine to control task execution on Pascal GPUs. Compute preemption coupled with unified memory across CPUs and GPUs will enable Pascal generation GPUs to look and act like a virtualized, composable pool of physical compute resources. The result is that HPC customers can start talking about GPUs in the context of sharable cloud resources, just like CPUs. This will have a huge impact from a software development point of view.
...
The P100 module is as much of an innovation as HBM memory. Instead of mounting the CoWoS package to an add-in board with a standard PCIe Gen3 x16 interconnect, Nvidia surface mounted the CoWoS package to a dedicated module and connectors. This enabled Nvidia to:



  • Vastly increase the number of signal pins to support four NVLink connections instead of only 16 PCIe Gen3 lanes
  • Lower the profile of the host chassis
  • Supply 300W of power through the motherboard instead of via separate “over the top” wires for power delivery
  • The P100 module assumes the cost of the 300W power transfer to the CoWoS package via the large, dedicated power management chips (PMICs, shown in the photo)
  • No over the top power supply enables use of mechanically simple large passive heat sinks for both the CoWoS package and the PMICs
Air cooling these 300W modules at 35°C ambient temperature is a significant feat of design.

http://www.nextplatform.com/2016/04/21/nvidias-tesla-p100-steals-machine-learning-cpu/
 
Last edited by a moderator:
Do we know that the GP100 and GP102 have similar size? I was expecting the GP102 to be about 1.5x the GP104, which would make it around 450 mm^2.
If you double GP104, you still have fewer transistors than P100 so it's quite fab-able. But a 1.5 scaleup is almost certain due to wattage limits. 180 watts for GTX 1080, scaled to a 2x GP102, is too high at 360 watts. 250 watts is a pretty hard limit before you have to really add extra PCIE power and extra cooling, so NVidia pretty much only does that for dual-GPU boards. The Titan Z dual board used 375 watts and was a triple-slot card for its cooler.
 
Especially if you're about to run into a power wall, it might make sense to consider going broader, i.e. fully double GP104, because then you can keep frequencies down which in turn lowers the voltage requirements which in turn lowers power drastically. IIRC, our GTX 1080 sample had around 1.07 volts at 1.800 MHz, but ran at a mere 0.8 volts for 1400 MHz. Might be worth running your simulator over with. Of course, it depends whether or not you really need to distinguish yourself from your own mid-range part or if 20% extra perf would be enough.

edit:
re-checked with Excel:
0.800 Volt -> 1455 MHz
1.062 Volt -> 1886 MHz, but still

FWIW, the least Millivolts required per GHz with our sample was at 1607 MHz with 0.862 Volt.
 
Last edited:
So about that availability...

https://twitter.com/killyourfm/status/744945604195979264?ref_src=twsrc^tfw

ClDf8hBUgAMhRpy.jpg
 
Especially if you're about to run into a power wall, it might make sense to consider going broader, i.e. fully double GP104, because then you can keep frequencies down which in turn lowers the voltage requirements which in turn lowers power drastically. IIRC, our GTX 1080 sample had around 1.07 volts at 1.800 MHz, but ran at a mere 0.8 volts for 1400 MHz. Might be worth running your simulator over with. Of course, it depends whether or not you really need to distinguish yourself from your own mid-range part or if 20% extra perf would be enough.

edit:
re-checked with Excel:
0.800 Volt -> 1455 MHz
1.062 Volt -> 1886 MHz, but still

FWIW, the least Millivolts required per GHz with our sample was at 1607 MHz with 0.862 Volt.
Worth noting those hitting close to 2.1GHz have a voltage around 1.08.

Still it raises the question of where everyone thought the cap was 1.25V for Pascal, unless they are assuming it is the same as Maxwell, or that it still exists but to go above 1.1V and hit 1.25V ceiling you need a very cold solution such as LN2 because the Boost3 is being too conservative with its algorithm and temperature variable.
I appreciate the 1.25V has also been broken with LN2, but that required a fair bit of work, point is maybe the temperature variable is this time influencing the voltage limiter more than needed in Boost3, especially when seeing some custom AIB having temps around 70c.

At least this is something that can be resolved in software if it is deemed to be too conservative (seems to be), just ironic everyone is generally talking (not directed at you) about the 1.25V limit when OC is still limited to 1.1V with Boost3 - or so it seems.
Would be great if Nvidia could be put under pressure by publications to explain this.
Cheers
 
Last edited:
Yes but that is a scalper and not an official retailer.
Although custom AIB are going out as soon as they come in it seems.
Cheers
You didn't check the tweet? According TweakTowns editor the cheapest GTX 1080 he can find atm is $859, that's not from scalpers.
 
You didn't check the tweet? According TweakTowns editor the cheapest GTX 1080 he can find atm is $859, that's not from scalpers.
You didn't look at the Amazon pageToTTenTranz linked?
That is a scalper trying to sell at stupid prices, we had other stupid prices at time of launch and why EVGA had to keep responding when their product was getting 1* because people complaining about price or trying to stop the scalper; back then they were trying to charge $1,099.
Demand must be pretty excessive in North America, I remember reading a post by an EVGA member (not here) suggesting 30% of the cards they get are reserved for North America.
And while demand has been crazy here in the UK, it is still possible to buy them; custom AIB though needs to be pre-order and on list as they sell instantly here.
Cheers
 
Last edited:
You didn't look at the Amazon pageToTTenTranz linked?
That is a scalper trying to sell at stupid prices, we had other stupid prices at time of launch and why EVGA had to keep responding when their product was getting 1* because people complaining about price or trying to stop the scalper; back then they were trying to charge $1,099.

Cheers
Of course that was a scalper, but point was that just like the tweet says, availability is so poor even the non-scalpers are asking $859 minimum now
 
Of course that was a scalper, but point was that just like the tweet says, availability is so poor even the non-scalpers are asking $859 minimum now

I think it was pretty clear if one looks at both Amazon he linked and the twitter, I was talking scalping in terms of Amazon :)
Anyway I added further info about availability in previous post compared to say UK, and what EVGA member mentioned on another forum.
If critical of my response maybe be critical of the OP for putting both together, that post started to skew context by showing a ridiculous selling price that only scalpers (buying stock from retailers and then selling even higher than them) are asking.
What is the current lead time if buying a 1080FE from Nvidia directly?
Cheers
 
Last edited:
In my country the cheapest 1080 can be found for 775€, which is around $875.

Has anyone ever been able to purchase a GTX 1080 for less than $700, let alone its MSRP of $600?
 
699 € (non-FE) atm for Germany. With 19% VAT deducted, that's 587,39 €.

edit: for the pic-or-it-didn't-happen-faction:
BZQGdSr.png
 
Last edited:
You didn't check the tweet? According TweakTowns editor the cheapest GTX 1080 he can find atm is $859, that's not from scalpers.

So why doesn't TweakTown's editor check here first:

http://www.nowinstock.net/computers/videocards/nvidia/gtx1080

GTX 1080 Combo: EVGA GTX 1080 FE + RAZER Abyssus Gaming Mouse
for $734.98 at NewEgg
http://www.newegg.com/Product/ComboBundleDetails.aspx?ItemList=Combo.2981487

GTX 1080 Combo: Gigabyte GTX 1080 Founders Edition + Gigabyte Pro-Laser Gaming Mouse
for $729.98 at NewEgg

http://www.newegg.com/Product/ComboBundleDetails.aspx?ItemList=Combo.2980727

or get it directly from Nvidia for $699
http://www.geforce.com/hardware/10series/geforce-gtx-1080
 
Last edited:
Have not looked much but 1080FE here in uk available now for £619.
https://www.overclockers.co.uk/msi-...dr5x-pci-express-graphics-card-gx-320-ms.html
Next price up where there is much more in stock is £630-£634 for 1080FE.

Gibbo has been pretty open about pricing comparison between here and North America, it is pretty naunced because it is not just currency prices (this has some charges involved as well), but international shipping,distributors,regions tax,margins for all parties involved.
In an ideal world yes but sometimes the margin expectation of the retailer from the manufacturer is too low, for example sub 5% at which point on new product we simply refuse to hit MSRP on majority of stock and will maybe just use a cheap brand or OcUK brand to hit MSRP and then other cards will be £10 or so higher to get us upto around 10% margin.

This happens only sometimes though, other times we get margins we are happy with and can hit MSRP no problem.

Typically the safe formula is USD cost, divided by dollar rate given by bank (not XE.com) and then add 30%-40% is a safe bet, as that covers taxes, shipping and margins.

That £629 is not too far off from the NewEgg prices albeit NewEgg seems to keep selling out and also creep possibly because they keep selling out.
The issue here though is trying to get cheaper non-FE cards, need to do pre-order and be on a list as that is what everyone it seems now wants.
Cheers
 
So to answer my question, supply is so short that no one is selling the card for less than $700, not even the non-founder editions.


That £629 is not too far off from the NewEgg prices
That's well over $900.
 
So to answer my question, supply is so short that no one is selling the card for less than $700, not even the non-founder editions.



That's well over $900.
How are you working that out using Gibbo's example?
Exchange rate of $1.45 to £1 say
your $900 = £620
Then apply the 30% for reasons Gibbo says that I quoted and that gives £806 (£170 over actual buying figure), way above the £630 figures I mentioned.
The up to 40% is used when MSRP margins are tight, which they are meant to be for 480.
Cheers
 
Okay then, what's the MSRP in the UK and/or rest of the EU?

We're certainly not getting anywhere near the MSRP in the states, so how is the EU doing?
 
Back
Top