Nvidia announces Volta based Titan V

tServo · 2017-12-08, 15:08

Available today at Nvidia's online store for a mere 3000 US dollars. It has 12GB HBM2 memory, 5120 Cuda cores, blower style cooling shroud, and 1/2 DP for 6.9 teraflops of DP compute power !
It sucks a tiny 250W of power. Most impressive to me is that it's based on 12 nanometer fab process !
Sorry, guys, but there is a limit of 2 boards per order.

Anandtech has more details :
https://www.anandtech.com/show/12135...r-3000-dollars

firejuggler · 2017-12-08, 16:39

tomshardware link
http://www.tomshardware.com/news/nvi...ops,36085.html

Also is the memory is 3072-bit wide instead of 4096 for the tesla. how does that affect performance?
is it 33% *slower* or did i misunderstood something again?

axn · 2017-12-08, 16:41

Quote:

Originally Posted by tServo

1/2 DP for 6.9 teraflops of DP compute power !

Bandwidth of 652.8 GB/s compared to, say, 484 GB/s of 1080 Ti. It will probably the fastest LL tester by quite some margin, but not worth the money. Other, less bandwidth-dependent, workloads will benefit.

CRGreathouse · 2017-12-08, 17:52

Seems to be aimed at the low-end machine learning community with the Tensor cores. Not sure it's a good fit for our crunching needs.

tServo · 2017-12-08, 19:21

Quote:

Originally Posted by CRGreathouse

Seems to be aimed at the low-end machine learning community with the Tensor cores. Not sure it's a good fit for our crunching needs.

Nvidia has made it clear that this is the community this board is aimed at.
The tensor cores do FP16 multiplies into an FP32 accumulator.
Still, for other purposes, it DOES have lots of DP power.

preda · 2017-12-09, 04:41

FP64 looks impressive, but the gain there will be bottlenecked by memory for LL/PRP.

Nevertheless, I'm curious by the time-per-iteration one gets for $3K (at 4M FFT size, let's say).

TheJudger · 2017-12-09, 19:23

Quote:

Originally Posted by preda

Nevertheless, I'm curious by the time-per-iteration one gets for $3K (at 4M FFT size, let's say).

Just scale the available memory bandwidth: http://mersenneforum.org/showpost.ph...postcount=2652

Oliver

preda · 2017-12-10, 02:16

Quote:

Originally Posted by TheJudger

Just scale the available memory bandwidth: http://mersenneforum.org/showpost.ph...postcount=2652

Oliver

So I'd guess between 0.8ms/it - 1ms/it at 4M FFT.. ?

TheJudger · 2017-12-10, 16:04

Quote:

Originally Posted by preda

So I'd guess between 0.8ms/it - 1ms/it at 4M FFT.. ?

I would place my bet close to 1ms/it for 4M FFT.

2017-12-08, 15:08	#1
tServo "Marv" May 2009 near the Tannhäuser Gate 3·269 Posts	Nvidia announces Volta based Titan V Available today at Nvidia's online store for a mere 3000 US dollars. It has 12GB HBM2 memory, 5120 Cuda cores, blower style cooling shroud, and 1/2 DP for 6.9 teraflops of DP compute power ! It sucks a tiny 250W of power. Most impressive to me is that it's based on 12 nanometer fab process ! Sorry, guys, but there is a limit of 2 boards per order. Anandtech has more details : https://www.anandtech.com/show/12135...r-3000-dollars

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Nvidia announces Titan Xp card	ixfd64	GPU Computing	10	2017-05-17 15:19
Nvidia announces Volta GPU	tServo	GPU Computing	2	2017-05-13 07:58
Nvidia announces Titan X	ixfd64	GPU Computing	20	2015-04-28 00:27
NVIDIA TITAN 320.59 driver with cuda 5.5	Manpowre	GPU Computing	43	2013-08-22 12:28
2x AMD 7990 or 2x Nvidia Titan ??	Manpowre	GPU Computing	27	2013-05-12 10:00

2017-12-08, 16:39	#2
firejuggler "Vincent" Apr 2010 Over the rainbow 101101101000₂ Posts	tomshardware link http://www.tomshardware.com/news/nvi...ops,36085.html Also is the memory is 3072-bit wide instead of 4096 for the tesla. how does that affect performance? is it 33% slower or did i misunderstood something again?

2017-12-08, 17:52	#4
CRGreathouse Aug 2006 2²×3×499 Posts	Seems to be aimed at the low-end machine learning community with the Tensor cores. Not sure it's a good fit for our crunching needs.

2017-12-09, 04:41	#6
preda "Mihai Preda" Apr 2015 2²·3·11² Posts	FP64 looks impressive, but the gain there will be bottlenecked by memory for LL/PRP. Nevertheless, I'm curious by the time-per-iteration one gets for $3K (at 4M FFT size, let's say).