Nvidia Pascal Announcement

[Razor1]

I don't think they will do that, they tend not to switch architectures over one line of products. Plus there might not be a need to because DP units are the same units doing SP now.....

 

[Berek]

Computex is end of May/beginning of June, so they may announce/release then:

http://www.eventsforgamers.com/event/computex-2016/

 

[OlegSH]

Going from 8 to 15 B transistors only 512 more cores ?

2.5x registers
1.7x shared memory
1920 additional FP64 cores
512 additional FP32 cores
+NVlink
+HBM2

 

[fellix]

I don't think they will do that, they tend not to switch architectures over one line of products. Plus there might not be a need to because DP units are the same units doing SP now.....

Nope, they are dedicated. And I guess Nvidia will iterate the consumer (desktop/mobile) SKUs into their own compute capability version (6.1 and 6.2) with some additional ISA changes.

 

[Voxilla]

That's pretty high Turbo clock for the big Pascal -- 1480MHz. I can only imagine how high the smaller consumer SKUs will reach.

Also pretty high 300 Watt.

 

[kukreknecmi]

Reborn of the Fermi, lol?

 

[pjbliverpool]

So it seems they are using the majority of the new nodes benefit to service the professional market rather than the gaming market. Dissapointing.

 

[Voxilla]

2.5x registers
1.7x shared memory
1920 additional FP64 cores
512 additional FP32 cores
+NVlink
+HBM2

Add to that 32 extra texture units, and no word on ROPs.
It's definitely made for compute first. Kind of first gen Maxwell.

 

[xpea]

Hopefully, they drop the DP stuff for the consumer models and add more shaders instead.

rumored GP102 makes sens. GP100 is too much HPC oriented to be viable as a consumer product. It also means that GP100 is the first GPU exclusively dedicated to HPC. At such performance leap with DGX1 on trendy deep learning market, they must have a long queue of customers wanting for this new toy.
hmmm maybe time to buy some NVDA shares $$$$$$$

 

[ninelven]

?

So it seems they are using the majority of the new nodes benefit to service the professional market rather than the gaming market. Dissapointing.

Seems rather rash to judge an entire generation's gaming performance based upon the specs of a single chip aimed at the professional market.... but that's just me.

 

[steveOrino]

So it seems they are using the majority of the new nodes benefit to service the professional market rather than the gaming market. Dissapointing.

Those high margin markets are very important to them right now so I think it makes sense.

 

[fellix]

Well, last time I heard Nvidia get most of their revenue from the embedded and HPC markets. Gaming can wait a bit longer for this generation release.

Add to that 32 extra texture units, and no word on ROPs.
It's definitely made for compute first. Kind of first gen Maxwell.

The render backend configuration for P100 is probably 128/1024 color/depth samplers, judging from the MC and L2 design.

The primitive setup rate seems unchanged (6 GPC clusters, just as GM200). Tessellation could be improved a bit, because of the finer distribution of the compute resources.

 

[Razor1]

what is interesting is with a 25% increase in core counts they are getting a 74% increase in SP performance.....

 

[iMacmatician]

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

From the dev blog:

Tesla P100 accelerators have four 4-die HBM2 stacks, for a total of 16 GB of memory, and 720 GB/s peak bandwidth

So the memory speed would be 1.4 Gbps.

 

[sebbbi]

The amount of SMs has been doubled in GP100, it has 2x of registers and 1.5x of shared memory per lane

Great news. ALUs are great for marketing, but big+fast register files and LDS (including fast LDS atomics since Maxwell) are more important for actual compute performance.

As game workloads are shifting more to compute shaders, it is good that NVIDIA's focus has also shifted towards compute once again. NVIDIAs graphics frontend has been way ahead AMDs for long time (and still improving), but they haven't managed to beat GCN in compute. Maxwell and Kepler are both great for compute (huge improvements over Kepler).

 

[ninelven]

what is interesting is with a 25% increase in core counts they are getting a 74% increase in SP performance.....

Clock speed.

 

[fellix]

Shared memory no longer just being a slice of L1? About time for that...
Means Pascal might actually allow mixed graphic/compute loads now.

Huh? This has been the case since Maxwell v1.

 

[fellix]

Great news. ALUs are great for marketing, but big+fast register files and LDS (including fast LDS atomics since Maxwell) are more important for actual compute performance.

As game workloads are shifting more to compute shaders, it is good that NVIDIA's focus has also shifted towards compute once again. NVIDIAs graphics frontend has been way ahead AMDs for long time (and still improving), but they haven't managed to beat GCN in compute. Maxwell and Kepler are both great for compute (huge improvements over Kepler).

Yep, Kepler design was very inefficient, particularly the shared memory bank organisation, resulting in a record low 32% efficiency. Maxwell improved both the throughput and latency of the shared memory by leaps and bounds on top of the overall SMM re-design.

 

[silent_guy]

Well, last time I heard Nvidia get most of their revenue from the embedded and HPC markets.

The numbers are all available in the CFO notes. Tesla sales accounted for ~$100M of revenue last quarter, out of a total of $1.2B. Embedded is something similar. Doesn't come close to GeForce which is at $800M or $900M or so.
Margins are a different story, of course. They must be very nice for that $129K DXI appliance.

 

[Ailuros]

There's a "told you so" someone owes to the above gentleman