Nvidia Pascal Announcement

[CSI PC]

This is kinda strange... beginning to smell like a paper launch, but the day has a couple more hours.

Just to clarify that they have plenty of stock.

Gibbo overclockersuk's employee who sources products there said:

Nope EVGA is excluded as we have over 100 of those too, we also have more of the other brands, just not in the photo.

We have over 500 cards IN STOCK right now ready for 14:00.

That quote was in reponse to a photo they posted and a member saying "is that all there is".
He has said pre-orders will be accepted for custom AIB 1080 from 2pm, so I got a feeling they launch soon after Computex.
Cheers

 

[CarstenS]

Sorry for self-quote:

This is kinda strange... beginning to smell like a paper launch, but the day has a couple more hours.

Over last 1-1.5 hours or so, cards seem to have become availabe here too.

 

[CSI PC]

LOL overclockersuk sold out that 500 1080FE stock by 14:07.......
Going to be a nightmare for back-end stock system and online shop to be aligned short term I think.
Ah well.
Cheers

 

[CarstenS]

Hm, so much for overpriced, it seems. Apparently, there actually IS a market.

 

[silent_guy]

Hm, so much for overpriced, it seems. Apparently, there actually IS a market.

Was there really ever any doubt?

 

[trinibwoy]

Was there really ever any doubt?

Compared to other hobbies $1000 is really nothing. There are a lot of people out there willing and able to spend that and more.

Obviously not a huge market but definitely big enough to clean out day 1 inventory.

 

[DuckThor Evil]

LOL overclockersuk sold out that 500 1080FE stock by 14:07.......

They still seem to have more than 10 pieces of the Zotac model left and one Gigabyte.

 

[Clukos]

First shots from the MSI PCB, seems much better than the FE, although it still has passively cooled VRM (like other MSI GPUs in the past) which i'm still not a fan of.

From here: https://www.overclockers.co.uk/msi-...dr5x-pci-express-graphics-card-gx-323-ms.html

And probably the one i would go for if i went for a 1080 (actively cooled VRM on this one): https://www.overclockers.co.uk/giga...phics-card-gv-n1080g1-gaming-8-gx-184-gi.html

 

[Deleted member 2197]

EVGA website is down ....

 

[CSI PC]

They still seem to have more than 10 pieces of the Zotac model left and one Gigabyte.

Yes but as I said back-end services are not necessarily synch'd with online shop.
I knew someone would pick up on this so quoting their employee in charge of ordering these products .

All FE cards are sold out, more due next week.

https://forums.overclockers.co.uk/showpost.php?p=29549956&postcount=1761

Edit:
It could be made more complicated as well by those that cancel the order.
Which may also be another headache for up to date data between front-backend services.
Cheers

 

[Grall]

although it still has passively cooled VRM

Not sure how you deem it "passively cooled"; the VRMs have a heatsink which sit underneath one of the main heatsink fans. And it's a large heatsink. My R390X board's VRM heatsink is way smaller than that, for a much higher power draw (and might not even have as efficient power regulation hardware either.)

 

[Clukos]

Not sure how you deem it "passively cooled"; the VRMs have a heatsink which sit underneath one of the main heatsink fans. And it's a large heatsink. My R390X board's VRM heatsink is way smaller than that, for a much higher power draw (and might not even have as efficient power regulation hardware either.)

Passively cooled means that it doesn't have a dedicated fan blowing on that heat-sink. Even this, w/o any fans is considered passive cooling

By comparison you can see the heat-sink on the G1 is actively cooling vram/vrm/core (fan blowing air directly on the hs responsible for cooling)

 

[Clukos]

And you can see the results in the respective reviews from guru3d

MSI: http://www.guru3d.com/articles_pages/msi_geforce_gtx_980_ti_gaming_oc_review,10.html

GB G1: http://www.guru3d.com/articles_pages/gigabyte_geforce_gtx_980_ti_g1_gaming_soc_review,10.html

I'm sure it's not a big deal in general but if i'm spending almost $1k on a GPU i'd want the best thing around, i went with a G1 last time around just because of the cooling solution and the overall experience has been great. That's all irrelevant if you are going with water of course.

Edit:

Not sure how you deem it "passively cooled"; the VRMs have a heatsink which sit underneath one of the main heatsink fans. And it's a large heatsink. My R390X board's VRM heatsink is way smaller than that, for a much higher power draw (and might not even have as efficient power regulation hardware either.)

The 390x you mentioned has active cooling for the vrm and passive for the vram

https://www.techpowerup.com/reviews/MSI/R9_390X_Gaming/images/cooler2.jpg
https://www.techpowerup.com/reviews/MSI/R9_390X_Gaming/images/cooler3.jpg

So it might be better than the one used for the 1080.

 

[Grall]

The 390x you mentioned has active cooling for the vrm and passive for the vram

I didn't mention any model/brand, but FYI, ASUS R9 390X DC-II has a smaller discrete VRM heatsink located under the GPU heatsink just like the MSI board you posted a pic of previously, except it's a really dinky little heatsink compared to the MSI sink. Literally it's barely larger than the board area the VRM chips themselves occupy, and yet the card functions just fine.

Seriously, if the sink sits directly in the airflow of one of the cooling fans, it's not a passive cooling system. And VRMs typically handle a lot of heat, from what I've seen of spec sheets, they are typically rated up to 125C at full power draw. You're worrying over nothing methinks, and the heat camera images you posted seem to back that up. Temps are nowhere near redline.

 

[CSI PC]

waiting to see 1070 with custom PCB....

While not 1070, here in the UK looks like the first custom 1080s will be going on sale in next couple of weeks, although sounds like stock supply will be smaller than the FE models.

Hi there

I am proud to announce OcUK has the worlds largest range of custom GTX 1080 for pre-order, stocks shall be arriving over next few weeks throughout June and we expect quantities to be small. So pre-ordering early shall secure your place nice and early in the queue:

https://forums.overclockers.co.uk/showthread.php?t=18732799

Anyway nice range of prices as well (due to small numbers and popularity not sure if that is the rrp or with a bit of excess applied).
Cheers

 

[Clukos]

I didn't mention any model/brand, but FYI, ASUS R9 390X DC-II has a smaller discrete VRM heatsink located under the GPU heatsink just like the MSI board you posted a pic of previously, except it's a really dinky little heatsink compared to the MSI sink. Literally it's barely larger than the board area the VRM chips themselves occupy, and yet the card functions just fine.

Seriously, if the sink sits directly in the airflow of one of the cooling fans, it's not a passive cooling system. And VRMs typically handle a lot of heat, from what I've seen of spec sheets, they are typically rated up to 125C at full power draw. You're worrying over nothing methinks, and the heat camera images you posted seem to back that up. Temps are nowhere near redline.

Yeah i don't think it's a huge issue either but active cooling on the VRM is flat out better design, no matter if it makes a real world difference or not. It's just another thing i don't have to worry about when pushing the GPU to its limits

 

[pixelio]

So native fp16 is in GP104 (sm_61) but the throughput isn't known yet and is not 256 ops/clock/SMP:

https://devtalk.nvidia.com/default/...e-gtx-1080-amp-gtx-1070/post/4889379/#4889379

 

[CarstenS]

something is in there, yes. And it seems to be slower than emulation through FP32.

 

[pixelio]

something is in there, yes. And it seems to be slower than emulation through FP32.

The exact same SASS instructions are generated for sm_60 (GP100), sm_61 (GP104) and sm_62 (GP???):

sm_60: HFMA2.FTZ R10, R0, R5, R6;                     /* 0x5d0003200057000a */
sm_61: HFMA2.FTZ R10, R0, R5, R6;                     /* 0x5d0003200057000a */
sm_62: HFMA2.FTZ R10, R0, R5, R6;                     /* 0x5d0003200057000a */

The open question is throughput.

We now know it's not 256 ops/clock like in the Tegra X1 (sm_53) and assumed to be in the GP100 (sm_60).

128 fp16 FMAs per clock would be excellent but at this point I'm betting that it's closer to 16 since anything higher should've been trumpeted!

 

[Ryan Smith]

The exact same SASS instructions are generated for sm_60 (GP100), sm_61 (GP104) and sm_62 (GP???):

sm_60: HFMA2.FTZ R10, R0, R5, R6;                     /* 0x5d0003200057000a */
sm_61: HFMA2.FTZ R10, R0, R5, R6;                     /* 0x5d0003200057000a */
sm_62: HFMA2.FTZ R10, R0, R5, R6;                     /* 0x5d0003200057000a */

The open question is throughput.

We now know it's not 256 ops/clock like in the Tegra X1 (sm_53) and assumed to be in the GP100 (sm_60).

128 fp16 FMAs per clock would be excellent but at this point I'm betting that it's closer to 16 since anything higher should've been trumpeted!

I was saving this for my review, but since I'm a bit behind schedule...

Under CUDA, FP16 performance (in FLOPs) on GTX 1080 is artificially limited to 1/64th the FP32 rate. Technically this is actually 1/128th the instruction rate, but due to vec2 packing you get twice the FLOPs out of a single instruction.

FP16 support on GTX 1080 is solely for binary compatibility reasons, and suffice it to say, NVIDIA made sure that's all it is good for.