Personal computing discussed

Moderators: renee, morphine, SecretSquirrel

 
Kougar
Minister of Gerbil Affairs
Topic Author
Posts: 2306
Joined: Tue Dec 02, 2008 2:12 am
Location: Texas

Volta TI Edition Musings

Sat Jun 03, 2017 1:22 pm

With Volta likely to launch in Q4 and after that purported prototype Volta Titan leak, I've been wondering. GV100 has a not-insignificant portion of the die taken up by tensor cores. But those I would think are useless to gamers that gobble up the Ti editions.

So does this mean the Titan and Ti editions of a future Volta part are going to be respun to launch sans Tensor cores, or still have some left in? I'm rather curious as from a pure gaming standpoint Tensor cores would make the card less appealing to gamers. On the other side of the coin I can kind of see NVIDIA wanting to add them to a Titan model card, and the Ti's have always shared the same core. What do you Gerbils think?
 
USAFTW
Gerbil
Posts: 32
Joined: Wed Feb 27, 2013 2:14 am

Re: Volta TI Edition Musings

Sat Jun 03, 2017 1:43 pm

The recent picture that an intern leaked out, to me, suggests that the card is ready to go depending on the stock build-up.
Judging by how a full GP102 is basically a GP100 with the FP16 and FP64 functionality physically removed, I'd expect the same to be the case for GV102, meaning a full 5376 ALUs in the fully enabled die. Although at this point we don't know much of anything about the architecture other than the fact that it's supposed to be 50% more power efficient than Pascal, the specs may turn out differently.
One thing I could point out is how consistent Nvidia has been with generational improvement since Fermi.
GTX 680/GTX 770 > GTX 580 (GK104 vs. GF110)
GTX 980 > GTX 780 Ti
GTX 1080 > GTX 980 Ti/Titan X
If this cadence is maintained, GV104 could be noticeably faster than the fully enabled GP102 aka Titan Xp.
 
Airmantharp
Emperor Gerbilius I
Posts: 6192
Joined: Fri Oct 15, 2004 10:41 pm

Re: Volta TI Edition Musings

Sat Jun 03, 2017 11:32 pm

What USAFTW said, mostly, but I'll add:

Pay attention to potential GV100 versus GV102. GV100 has your compute resources and 'Tensor cores', but GV102 should be a similarly-sized part that's as lean as the mainstream GV104 parts are expected to be. It'll basically be a double-wide GV104, just like GP102 is to GP104 [Titan Pascal/1080Ti/Titan XP to the 1080].

So, possibly like this:
GV100- commercial products, Quadros, and possibly a Titan, but this is not very likely as the consumer argument for this part is very thin.
GV102- Following the GP102 pattern, likely a Titan first, then a top-end Ti part.
GV104- 1070/1080 replacement.

We can't really know what parts beyond GV100 will come out first, but I'd expect a GV104-based 1080 successor to be the opening salvo. GV102 releases may follow the Pascal pattern that I applied above or may be put to purpose sooner if Vega is competitive.
 
the
Gerbil Elite
Posts: 941
Joined: Tue Jun 29, 2010 2:26 am

Re: Volta TI Edition Musings

Mon Jun 05, 2017 1:38 pm

There were quiet a few minor differences between the GP100 and GP102. First off was the memory interface where GP100 uses HBM with an undisclosed number of ROPs. Between the higher bandwidth and purportedly more ROPs, this should give the GP100 better performance at higher resolutions. The other difference is how 16 bit floats are handled on die. GP100 splits a 32 bit ALU to do two where as GP102 has specific hardware units that handle it (and at reduced throughput). GP100 has nvLink where as GP102 uses standard SLI. These add up to radically different die sizes as the GP102 is 471 mm^2 vs. 610 mm^2 for the GP100. Of course, these are both dwarfed by GV100 at a behemoth 815 mm^2 and ranks as the largest mass produced die (and with interposers and EMIB coming into fashion, it has a good chance of keeping that title for years into the future).

I strongly suspect that the picture of the Titian VX (as I'll call it for now) is GP102 due to the insane size of GV100. So what would I expect out of a GV102? Probably a die size in the 600 mm^2 range with a nice increase in ALU count vs. the GP102. Memory bus width would be 384 bit wide bus native but I'll go out on a limb and say that nVidia is only going to ship the Titan with a 320 bit wide bus and 20 GB of GDDR6 memory. Using 16 Ghz GDDR6 memory will still offer a 20% bandwidth increase over 11 Ghz GDDR5X on the Titan XP. Yields are always an issue so I'd expect at least two clusters of ALUs to be disabled. nvLink will be included and this time around nVidia may again permit 3 way and 4 way setups to be fully supported. It wouldn't surprise me if they did so only with Titan VX though while keeping normal Geforce cards limited to a pair.

For the GTX 1180 TI part, for lack of a better name, I would expect 10 GB of GDDR6 running at 15 Ghz on a 320 bit bus. Or if nVidia follows Pascal, the Ti version will be similar to the original Titan with the Titan VX2 arriving a month or two later.
Dual Opteron 6376, 96 GB DDR3, Asus KGPE-D16, GTX 970
Mac Pro Dual Xeon E5645, 48 GB DDR3, GTX 770
Core i7 [email protected] Ghz, 32 GB DDR3, GA-X79-UP5-Wifi
Core i7 [email protected] Ghz, 16 GB DDR3, GTX 970, GA-X68XP-UD4
 
Kougar
Minister of Gerbil Affairs
Topic Author
Posts: 2306
Joined: Tue Dec 02, 2008 2:12 am
Location: Texas

Re: Volta TI Edition Musings

Wed Jun 07, 2017 4:10 pm

The funny thing is last NVIDIA said anything about consumer Volta it was still HBM2. Of course they first announced Volta way back in 2013 for that matter. Did they officially state GDDR6 at Computex or are people going off the SK Hynix information? I'd think the "1180" will undoubtedly be GDDR6, but what about a Titan model with half again the shader resources? Either would be great as it means more of the power budget goes back to the GPU.

It appears consumer Volta will make use of NVLink 2.0 which would be a boon for workstation use. But I really don't see NVIDIA utilizing it to make 3 or 4-way SLI a thing again. NVIDIA nixed that in the first place due to the cost & headaches of coding drivers for it when such a small percentage of users actually had 3-4 GPUs in SLI, and changing the interface from PCIe to NVLink isn't going to change that. Linking cards at a low level for a compute cluster is practically no work at all, it's up to the developers or the program running on the cluster to support the extra bandwidth/links.
 
the
Gerbil Elite
Posts: 941
Joined: Tue Jun 29, 2010 2:26 am

Re: Volta TI Edition Musings

Fri Jun 09, 2017 4:26 pm

Hynix did state that there was a 2018 product using a 384 bit wide GDDR6 bus. After that, it was presumed that it was the GV102 following traditional iterations.

Any AMD part needing bandwidth is simply using HBM2.
Dual Opteron 6376, 96 GB DDR3, Asus KGPE-D16, GTX 970
Mac Pro Dual Xeon E5645, 48 GB DDR3, GTX 770
Core i7 [email protected] Ghz, 32 GB DDR3, GA-X79-UP5-Wifi
Core i7 [email protected] Ghz, 16 GB DDR3, GTX 970, GA-X68XP-UD4

Who is online

Users browsing this forum: No registered users and 1 guest
GZIP: On