Nvidia Pascal Speculation Thread

Status
Not open for further replies.
Were those chips process pathfinding chips? They (GPU vendors) won't be pioneers for the new node this time, right? Apple and others are handling the newborn this round, I think. Still, a 600mm2 chip seems risky.
Guess everyone learned their lesson: Concentrate on fewer process variations, adhere more strictly to design rules and off you go.
 
I'm seeing the MSI GT 710 listed right now where I usually buy hardware.
Very confusing : it appears identical to the GT 720.

In fact what I'm seeing here :
MSI GT 710 1GD3H LP : 954 MHz GPU, 800MHz RAM
Gainward GeForce GT 720 1 GB : 797 MHz GPU, 800MHz RAM
MSI GeForce GT 720 N720-2GD3HLP : 797 MHz GPU, 800MHz RAM

GT 710 faster than GT 720? What's that!
The better card still is that one I guess (overpriced, but only slightly more than the preceding MSI GT 720)
MSI GeForce GT 720 N720-2GD5HLP : 797 MHz GPU, 1250 MHz RAM
 
That sounds like a great card to use as dedicated PhysX processor.

If PhysX was still relevant, that is.
 
Really confusing, true. Same # of ALUs, faster wrt to clocks, but only available in DDR3 flavor. Main difference from what I can see: No GT710 with Display Port, thus only for lower res displays up to 2560 x 1600. Maybe there's a niche for just that but I'm having a hard time seeing it.
 
PhysX is cross-platform and a useful thing to run on the CPU these days - on CPU, it has multi-threaded SIMD code rather the dubious single-threaded, x87 code it used to have.
Nvidia-only, GPU-accelerated effects are rare and if you find yourself running them, a big Kepler or Maxwell GPU is powerful enough and has good enough multi-tasking to run the stuff by itself.
You might even imagine a dedicated "coprocessor" card could reduce performance by adding overhead and latency, though that would have to be tested.
 
Really confusing, true. Same # of ALUs, faster wrt to clocks, but only available in DDR3 flavor. Main difference from what I can see: No GT710 with Display Port, thus only for lower res displays up to 2560 x 1600. Maybe there's a niche for just that but I'm having a hard time seeing it.

I haven't seen any board with GK208 or GK107 that has Displayport, at least retail ones.
As far I know the lowliest nvidia board with Displayport is GTX 750, and before that 650 Ti Boost (based on GTX 660). AMD isn't much better (minimum is 250X, a.k.a. a fast Radeon 7770, or 260/360)

That's bad but technically on low-end boards you can do dual-head 2560x1440 (one on HDMI 1.4 and one on dual-link DVI). For some people that's even an upgrade : it's pretty commmon that they have a couple outputs on the motherboard (Intel graphics) that go up to 1920x1080/1920x1200 only.
Many millions people run on a ~$100 1080p monitor (I'll refrain from sharing judgement on the typical quality)
 
Yeah, that's strange too. Funny side-note: Cheapest Nvidia card with DP is a Quadro (GK107-based K600, 99 EUR)... from AMD you can get R7 250E here in Germany for around 80 EUR, or if you're into VLIW, a 6670 for 75. I thought though that HDMI carries a royalty cost where DP does not.
 
Yeah, that's strange too. Funny side-note: Cheapest Nvidia card with DP is a Quadro (GK107-based K600, 99 EUR)... from AMD you can get R7 250E here in Germany for around 80 EUR, or if you're into VLIW, a 6670 for 75. I thought though that HDMI carries a royalty cost where DP does not.
Correct. DP does require that you're a member of the VESA to get the specification, which is $5K/$10K depending on the corp size, so it's not absolutely free. But there's no per-device royalties.

However even with the cost of HDMI, it's seen as more desirable at the low-end. Low-end devices have HDMI, so therefore any new low-end device should have HDMI...
 
VC has a entry of GP104 Geforce 1800 series, which should be Pascal series card, according to VC, Pascal GP104 is clocked at 1200MHz, should have 5120 CUDA cores, 8GB HBM2 with 1TB/sec memory bandwidth.

Another VC news says Pascal is possibly water-cooled (not impossible considering the fact HBM should has more serious heat issue comparing to traditional layout), so maybe next generation work-station case should be redesigned for water-cooled professional cards.
 
VC has a entry of GP104 Geforce 1800 series, which should be Pascal series card, according to VC, Pascal GP104 is clocked at 1200MHz, should have 5120 CUDA cores, 8GB HBM2 with 1TB/sec memory bandwidth.

Another VC news says Pascal is possibly water-cooled (not impossible considering the fact HBM should has more serious heat issue comparing to traditional layout), so maybe next generation work-station case should be redesigned for water-cooled professional cards.

Well, outside heat issue .. the thing is "HBM gpu design" is just a perfect match for H2o..

As you can seen on this waterblock, the design of the water circuit is rather simple, close to a simple CPU waterblock, you have a central inlet and then PWM along the outlett.. compared to standard gpu waterblock, it is really simple, no need thave an external circuit around the central one who "run" over the ram position.

At contrario, if you use an air coooling ( like on the Fury "non X" ), you need too have large surfaces of fins for get efficient exchange between air and metals part.. ( so you will end with a pretty long gpu's ). Its so logic to use H20 with HBM design..

ekfc-r9-fury-nano_np_fill1_800.jpg
 
VC has a entry of GP104 Geforce 1800 series, which should be Pascal series card, according to VC, Pascal GP104 is clocked at 1200MHz, should have 5120 CUDA cores, 8GB HBM2 with 1TB/sec memory bandwidth.

GM204 -> GP104 = 2.5x the shaders? Sounds a little far fetched unless there have been big changes (Fermi to Kepler for example).
 
GM204 -> GP104 = 2.5x the shaders? Sounds a little far fetched unless there have been big changes (Fermi to Kepler for example).
NVIDIA already described Pascal as "Maxwell + Mixed Precision + 3D Memory + NVLINK" so unlikely.
Anyway, is VC VideoCardz? It has quite sketchy reputation unless they have actual slides from a manufacturer.
 
Shipping data on Zauba suggests, that the next performance chip (GP104?) has a bit smaller package than GK104/GM204, but much higher pin count.
Mixed-Precision will probably make the SMM bigger. So if were are looking on a 250-300mm² GP104 a better guess is 2048-2536CC @ 256-Bit GDDR5X.
We have to consider that GP102 is also planned, this could be the bigger ~400mm² HBM2 performance part for 2017.

I wonder if they will push Mixed-Precision ~ FP16 through GameWorks in to the games. If you can switch ~25% of the code to FP16 the GP104 I suggested will offer >8 "effective" TFLOPs .
 
NVIDIA already described Pascal as "Maxwell + Mixed Precision + 3D Memory + NVLINK" so unlikely.
Anyway, is VC VideoCardz? It has quite sketchy reputation unless they have actual slides from a manufacturer.

But VC also has a tradition to leak NDA info this way (a recent example is VC entry for AMD Fury series months before AMD release that card officially).

It is not unlikely that Pascal GP104 has 5120 cuda cores, since GP104 is most likely a pure-gaming card, with rather limited FP64 or even FP16 functionalities, and there are two-full node or at least 1.5 node maunfacturing gap between Pascal and the previous generation GPUs.

However the real problem is, how many more cuda cores Nvidia can manage to put into the big Pascal GP100 then.
 
Well, outside heat issue .. the thing is "HBM gpu design" is just a perfect match for H2o..

As you can seen on this waterblock, the design of the water circuit is rather simple, close to a simple CPU waterblock, you have a central inlet and then PWM along the outlett.. compared to standard gpu waterblock, it is really simple, no need thave an external circuit around the central one who "run" over the ram position.

At contrario, if you use an air coooling ( like on the Fury "non X" ), you need too have large surfaces of fins for get efficient exchange between air and metals part.. ( so you will end with a pretty long gpu's ). Its so logic to use H20 with HBM design..

Sure, water-cooled is a good match for HBM.

However, I was talking about the impact of water-cooling to current GPU workstation market, most of the high-density GPU computing platform at the moment is optimized for air-cooling (using high power airflow to cool the entire system, not just the GPUs), a Fury-X like water-cooling will get problem with such designs.

I just hope they can design new solutions quick enough since I am eager to try Pascal to handle my computing tasks.
 
I wonder if they will push Mixed-Precision ~ FP16 through GameWorks in to the games. If you can switch ~25% of the code to FP16 the GP104 I suggested will offer >8 "effective" TFLOPs .
Supposedly, they'll be called G-TOPS. ;)
 
New Really confusing, true. Same # of ALUs, faster wrt to clocks, but only available in DDR3 flavor. Main difference from what I can see: No GT710 with Display Port, thus only for lower res displays up to 2560 x 1600. Maybe there's a niche for just that but I'm having a hard time seeing it.
Maybe destined to be the replacement for GT210? It has been the cheapest card on the market for ages, but from April 2016 will have no driver support.
Some people still need those, for example, I need a display output for my P67 motherboard, and the current HD4670 is dying.
 
Status
Not open for further replies.
Back
Top