Nvidia Pascal Announcement

$699 = £475.90
+ VAT@20% = £571.08

So £620 isn't a huge markup buy why should there be any? They both need shipping from the far east. Surely there should be ample profit in the $699/ £475.90 base price for both US and UK retailers?

Last time I checked other PC products were fairly priced between the US and UK retailers. Although Overclockers are usually a bit more expensive than Scan, eBuyer, Dabs (BT Store now I think), etc.
 
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?
According to the list, Nvidia distributed shortly before launch:


GTX 1080 Founders Edition: 657 Euro plus VAT, which of course varies from country to country. For germany, it was 789 EUR, for France 789 EUR, for GB 619 GBP.
GTX 1080 Partner Boards: 559 Euro plus VAT, which of course varies from country to country. For germany, it was 665 EUR, for France 669* EUR, for GB 529 GBP.

*don't know why the partner boards are 4 EUR higher in FRA than in GER, while the FE is not.
 
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?

What is strange is just how quickly stock is out in North America with main retailers, earlier I mentioned EVGA representative on another thread suggested it is around 30% (even if it is a bit lower say 20-25% that still must be a lot) reserved for that region of their 1080 cards.
Cheers
 
Last edited:
What is strange is just how quickly stock is out in North America with main retailers, earlier I mentioned EVGA representative on another thread suggested it is around 30% (even if it is a bit lower say 20-25% that still must be a lot) reserved for that region of their 1080 cards.


Because there are very few cards to begin with?
 
Because there are very few cards to begin with?
Even though they are getting close to 30% of the cards provided by an AIB?
Yeah that was an EVGA representative finger in the air estimate, but suggests there are plenty of cards if other partners follow a similar formula.

Following some of the other forums, I am hearing of people able to get their custom AIB in North America easier than we are here in the UK, we are having to wait for those after initial filling in North America it seems.
So something is up, and I do think these are selling much faster than 980 and 980ti combined at launch, which does not help.

OverclockersUK was one of the largest or actual largest 980ti launch retailers, and they are saying for them at launch for 1st 2-3 weeks the 1080 is actually selling better than previous Maxwell2 launches, not just 980ti and 980 but also 970.
Just emphasising the key point of several weeks as I (and also Gibbo in a post) doubt it will sustain sales comparable to more mainstream such as 970.
Gibbo is the lead at OverclockersUK for GPUs,monitors,etc.
And yeah I appreciate this is not necessarily replicated globally.

Cheers
 
Last edited:
Because there are very few cards to begin with?
Judging from what was written by many when the pricing of GTX 1080 became public, only a very tiniy minority of gamers could even afford such a card, while most of those would already have a 980 Ti or a Titan X, so not a really worthwhile upgrade-path with GTX 1080.

And of those who could afford it, only a fraction would consider buying it, because what those people really wanted was the "Big Pascal" or at least the "Big Gamer Pascal".

So the question is, how many cards would one consider "very few" with respect to the TAM? There is one shop here in germany (mindfactory dot de - not to make any advertisements) who publicly shows how many of each products were already sold. As for the correctness of those numbers, I obviously cannot vouch, but the 1080s for far sum up to aroun 2400. And to my knowledge, albeit one of the larger e-tailers in germany, said outlet is by far not the largest.

So, how many is very few?
 
Last edited:
What is strange is just how quickly stock is out in North America with main retailers, ...
With websites like nowinstock, that's not really surprising, is it? It removes the friction of people having to browse around.

No matter how big the supply, as long as demand is higher, these kind of websites pretty much guarantee that cards will sell out immediately.
 
Judging from what was written by many when the pricing of GTX 1080 became public, only a very tiniy minority of gamers could even afford such a card, while most of those would already have a 980 Ti or a Titan X, so not a really worthwhile upgrade-patch with GTX 1080.

And of those who could afford it, only a fraction would consider buying it, because what those people really wanted was the "Big Pascal" or at least the "Big Gamer Pascal".

So the question is, how many cards would one consider "very few" with respect to the TAM? There is one shop here in germany (mindfactory dot de - not to make any advertisements) who publicly shows how many of each products were already sold. As for the correctness of those numbers, I obviously cannot vouch, but the 1080s for far sum up to aroun 2400. And to my knowledge, albeit one of the larger e-tailers in germany, said outlet is by far not the largest.
Between the number of countries and the amount of stores that sell it, I think the volumes are much higher than people think. If a website like that can sell 2400 in a month, a worldwide volume of 100k is not unreasonable.
 
So I got an email from NVIDIA this morning. We can finally lay the question of FP16 execution to rest once and for all.

GP104 has a single, dedicated FP16x2 core per SM. The FP32 cores cannot execute FP16x2.

This is basically identical to how NVIDIA does FP64, except GP104 has more FP64 units (4 per SM). This is where the 1/128 instruction rate comes from, and since it's capable of executing 2 FP16 ops in a vec2, the resulting 1/64 FLOP rate. This also means that it takes 32 clocks to actually execute a single instruction of a single warp.
 
With websites like nowinstock, that's not really surprising, is it? It removes the friction of people having to browse around.

No matter how big the supply, as long as demand is higher, these kind of websites pretty much guarantee that cards will sell out immediately.
Yeah like I said here in UK one of the most popular retailers is doing better at launch than 970, which is quite incredible and shows Nvidia knows their market very well.
But even here we have sites that make it very easy to check the main retailers, even Nvidia makes it moderately easy for UK as it lists models and UK sites-retailers selling, albeit with no stock info but provides easy access to a diverse range of major stockists.
And still quite a few places still have various 1080FE in stock now, which for North America seems very different with stock out.

If I can find a few major retailers with 1080FE in stock (not all but not difficult to find) with sales for at least one better than 970 launch (I would assume the trend would be across all UK stockists), then one would expect a similar situation in North America but maybe sales there are better than even that.
As you say we need more info, but strange can be that selling patterns and warehouse stocking-logistics are unusual, and exacerbated by the huge buying interest.

Just to add as Carsten mentioned there is a fine balance for manufacturers between internal manufacturing line-stock levels and storage, it has been ages since I was told about it but there is a formula many of the large international manufacturers use to calculate how efficient their line (manufacturing and logistics) is, and deviations from the figure means business-manufacturing efficiency is down for that product.
For a new product they would have some kind of financial-sales forecasting and probably based upon 980/980ti with a buffer for additional interest, but I doubt even Nvidia expected the 1080 to do better than 970 in some regions at launch (key point 1st several weeks).
The 1080 would be deemed a small footprint enthusiast product, so they would not go nuts producing much beyond their forecast, otherwise if sales were not strong the manufacturing line would need to be slowed down including purchasing of components required from 3rd parties to build the product, impact this has on the logistics side-shipping batchment process,etc.

AMD will have a different forecast challenge with the 480, but its sales will still be more sustained after initial launch due to larger demographics of potential customer due to targetting mainstream product.
Cheers
 
Last edited:
So I got an email from NVIDIA this morning. We can finally lay the question of FP16 execution to rest once and for all.

GP104 has a single, dedicated FP16x2 core per SM. The FP32 cores cannot execute FP16x2.

This is basically identical to how NVIDIA does FP64, except GP104 has more FP64 units (4 per SM). This is where the 1/128 instruction rate comes from, and since it's capable of executing 2 FP16 ops in a vec2, the resulting 1/64 FLOP rate. This also means that it takes 32 clocks to actually execute a single instruction of a single warp.
Any chance you can nudge them for an answer regarding GP102 on same issue :)
I guess this will need to be asked again when that launches.
Is there even any point to putting in a single dedicated FP16x2 core (is this even a mixed-precision FP32/16 core or a unique FP16x2) per SM?
Also raises questions again just like 970 why this information was not shown with the Cuda cores/SM diagrams and architecture given to press for the reviews of the 1080/1070.
Cheers.
 
Last edited:
Is there even any point to putting in a single dedicated FP16x2 core (is this even a mixed-precision FP32/16 core or a unique FP16x2) per SM?
Compatibility with programs written for GP100, for developers who are writing code to deploy on DGX-1 and such. The same reason NVIDIA bothers to include FP64 CUDA cores. NV puts just enough hardware to allow execution, but not enough that it's even remotely fast at it, that way these cards don't undermine Tesla sales. Never mind the fact that the vast majority of GP104 cards will never touch FP16x2 code, so spending too much die space would be wasteful for most cards.
 
Compatibility with programs written for GP100, for developers who are writing code to deploy on DGX-1 and such. The same reason NVIDIA bothers to include FP64 CUDA cores. NV puts just enough hardware to allow execution, but not enough that it's even remotely fast at it, that way these cards don't undermine Tesla sales. Never mind the fact that the vast majority of GP104 cards will never touch FP16x2 code, so spending too much die space would be wasteful for most cards.
Although from a compatibility perspective wouldn't you also ideally want multiple cores to test efficiency from a Cuda Thread Block/Grid perspective?
You cannot do much parallelism, which is pushed with Cuda.
Maybe Spworley can share a little insight how useful it will be.

I think a few who purchased the 1080FE from a CUDA programming perspective may feel a bit hard done by, why I feel the capabilities or at least the Cuda Cores of the 1080 should had been communicated better.
But then Nvidia were crud communicating the capabality of the Kepler Titan/780ti/780 in relationship to Tesla.
CHeers
 
Last edited:
Compatibility with programs written for GP100, for developers who are writing code to deploy on DGX-1 and such. The same reason NVIDIA bothers to include FP64 CUDA cores. NV puts just enough hardware to allow execution, but not enough that it's even remotely fast at it, that way these cards don't undermine Tesla sales. Never mind the fact that the vast majority of GP104 cards will never touch FP16x2 code, so spending too much die space would be wasteful for most cards.

Also, for FP64 CUDA cores, FP64 and FP16 ( or 10) are needed for be certified DX12 ( optional on DX11_0, DX11_1, required in DX12 ).. Theres not minimum performance value, but should be at least in theory capable of it.

@CSI_PC : Its calling marketing...
 
Last edited:
So I got an email from NVIDIA this morning. We can finally lay the question of FP16 execution to rest once and for all.

GP104 has a single, dedicated FP16x2 core per SM. The FP32 cores cannot execute FP16x2.

This is basically identical to how NVIDIA does FP64, except GP104 has more FP64 units (4 per SM). This is where the 1/128 instruction rate comes from, and since it's capable of executing 2 FP16 ops in a vec2, the resulting 1/64 FLOP rate. This also means that it takes 32 clocks to actually execute a single instruction of a single warp.
Yes, please put those rumors to rest about FP16 throttling on GP10x.
 
Last edited:
Because there are very few cards to begin with


Originally Posted by Greebo
My guess is that so long as you have decent quantity that ignoring the pre-orders on 1070/80s you will ship more 480s in the first week than you have 70/80s in the first month
Originally Posted by Gibbo (June 20)
Doubtful, we have shipped well over 1000 on 1080 already, we have had a lot of 1080 stocks, but the sales are totally insane on that part, even beyond the sales what 970 was at launch. Kudos to NVIDIA for launching a product at twice the MSRP and still selling more.?
https://forums.overclockers.co.uk/showpost.php?p=29653512&postcount=4754

Originally Posted by Gibbo (June 22)
We had nearly 1000 (GTX 1070) land now and they go on special promo today (£419.99), they are flying out!
https://forums.overclockers.co.uk/showpost.php?p=29665183&postcount=6352
 
Last edited:
Back
Top