As expected, NVidia has released an additional GeForce card in the form of the GTX650Ti, filling out the GeForce 600 range.

The GK106 GTX650Ti has 768 cuda cores (shaders) and reference models are clocked at 925MHz, sport 1GB GDDR5 and have a TDP of 110W (so one 6-pin connector is sufficient).
Manufacturer variants will include 2GB versions and a range of clocks from 925MHz to 1071MHz (~16% range so far)…

As with other GK cards, these are PCIE3 compliant, but like the GTX650 they do not feature boost (get a fast one)!

Cost:
In the UK these begin at £115 and rise to ~£135 (at good online retailers).
In the USA the starting price is $145, and in the Euro zone they also start around 145Euro.

Performance:
With 768cuda cores these cards fall between the GTX650 and GTX660Ti (GTX660 is OEM), and should give roughly twice the performance of a GTX650.

In terms of performance per Watt, the GTX650Ti should slightly outperform the GTX650 (in-itself), but in terms of performance per outlay (purchase cost) it’s certainly worth the extra $40 for twice the crunching performance!

While it’s generally preferable for larger cards to have 2 or 3 fans, I think one fan should be sufficient for these cards. I would however be concerned about the size of the fan; in my experience small fans fail very frequently, while larger fans generally keep running for years.

Overclocking aside (if even possible), it’s generally the case regarding improved/performance cards (cards with higher frequencies) that they give as good a performance per Watt (at least) as reference models. So when you add the system Wattage overhead (motherboard, CPU, RAM, HDD...), they are better value for money. That said, for the same reason it’s usually more efficient overall to get a ‘bigger’ card.

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

Update: GTX660 with GK106 and 960 shaders at 980 MHz went retail a few weeks ago.

MrS

---

Scanning for our furry friends since Jan 2002

Several etailers are offering a free code for download of Assassins Creed III. This could likely be sold (eg on ebay) for at least $30 making this a VERY high value card IMO!

edit: just purchased evga GTX 650 Ti SSC at 1071 MHz, for $159.99 - $5.00 mail in rebate, plan to sell the game code for likely $30-$45 = appr. $114.99 to $129.99 total price :D

---

XtremeSystems.org - #1 Team in GPUGrid

yes i read about the release from this card to and was impressed by the number of cuda cores.

---

DSKAG Austria Research Team: http://www.research.dskag.at

Several etailers are offering a free code for download of Assassins Creed III. This could likely be sold (eg on ebay) for at least $30 making this a VERY high value card IMO!

edit: just purchased evga GTX 650 Ti SSC at 1071 MHz, for $159.99 - $5.00 mail in rebate, plan to sell the game code for likely $30-$45 = appr. $114.99 to $129.99 total price :D

well I was able to sell the Assassin's Creed 3 code for $43. But the buyer is claiming the code is "already used". So I may end up being out the $43 with nothing to show for it if the buyer requests a refund through PayPal. :(

---

XtremeSystems.org - #1 Team in GPUGrid

Ouch, that's pretty bad! By now you won't be able to use the code either, and you'll never know if that guy is happily running the game now.. or is there anything the game distributor could help you with here? Like invalidate all current installations using this code?

MrS

---

Scanning for our furry friends since Jan 2002

the buyer tried it again the next day and it worked!

---

XtremeSystems.org - #1 Team in GPUGrid

For the price its fantastic number cruncher. Its doing a long work unit in 8 hours!

My 650ti is currently posting slightly better averages than my 560 , and it has not peaked yet. So im not exactly sure what its peak performance is.

About 190k RAC using only long-run tasks.

MrS

---

Scanning for our furry friends since Jan 2002

How does it do for desktop lag? I usually use a dedicated card for DC projects, but this time can only free up the display card slot. I do some video editing, but otherwise just do web browsing with Firefox.

The GT 440 I tried (96 cores) on GPUGrid was not so great on either desktop lag, or speed. But it seems to be a good match for the Keplers, which is not true for all projects as you know.

worst case scenario you can just disable "Use GPU while computer is in use" in BOINC.

---

XtremeSystems.org - #1 Team in GPUGrid

With my GTX660Ti I didn't notice disturbing desktop lag, although HD video playback may sometimes have dropped below 25 fps for short periods. But it's got a lot more raw horse power than a GTX650Ti.

MrS

---

Scanning for our furry friends since Jan 2002

Short test with GTX650TI (MSI Power Edition/OC) & ACEMD2: GPU molecular dynamics v6.16 (cuda42), i7-3770K @3,8 Ghz, 9 threads in BOINC (8xSIMAP, 1xGPUgrid), no problem if this slows down the performance. The progress counting moves constantly over all nine threads.

3x @Stock (993/1350)
Runtimes in s
- 10,260
- 10,294
- 10,207 : mean 10254

1x @ Core +211 = 1202Mhz (Memory @stock)
Runtime in s
- 9,136
- 9,406
- (third in queue) : mean 9271 (-983)

OC +22% (temperatures + 5 degrees Celsius, VCore 1.087 V > 1.150 V)
Result +9% (983/10254)

Question
With previous GPUs it was "safe" to concentrate on shader-OC, overclocking the memory (I found) was not recommended to keep reliability.

What about the new Kepler GTX 650 TI (GK106-220), how safe is it to tune up the memory clock ? Experiences available or give it a try ?

I would prefer valid results, wouldn´t like to test a small increase but "respectable". In a related OC review they occed the memory from 1350 to 1550 (+200 Mhz), do you think it´s just extreme, lucky or necessary to push or support this according to the Core-OC (+211) ?

Can such an opinion be argued for regarding the special kind of calculation resp. requirements of GPUgrid workunits ? Does GPUgrid need memory performance ? More than Core or a balanced mix ? (I can guess your answers, balanced mix.)

In other words: how to get the missing 13 percent ? (22-9) ;-)

I have no experiences with Kepler GPUs, shader tuning is out-dated probably.

P.S. Yes of course, no testing with scientific workunits as a benchmark. I think of OpenCL benchmarks to verify the stability of ocing both core + memory. Temperatures are never a problem, that shows my short testing, the card runs @GPUgrid no more quietly such as @WCG/HCC1 but with a distinct noise (fan @47%=1600rpm). I bought the GTX 650 TI as a small & quiet cruncher with low idle consumption if paused, mission accomplished, now fine tuning shall make it perfect.

(edit timeout ?)

1x workunit (ACEMD2) @ Core +211 / Memory +198 = 1202/2898 Mhz (1.175V)
- 9,057 (-1197)

OC Core +22%
OC Memory +14%
Result +11,6% (1197/10254) - more done by core or memory ?

That should be the limit. Next days I look at undervolting and better balancing.

Your core-OC still yielded more, percentage wise, than your mem-OC. However, GT640, GTX650Ti and GTX660Ti are less balanced than usual cards in the way that they pack a surprising amount of shader power into a package with "usually just enough" memory bandwidth. In games these cards are love memory OC's, much more so than usual cards. Seems like on these cards GPU-Grid is finally also affected by the mem clock by a measureable amount.

You can take a look at the memory controller load via e.g. GPU-Z. Your numbers should be interesting, with and without OC. I've got a slightly core-OC'ed GT640 at ~66% mem controller load and a slightly core-OC'ed GTX660Ti at 44%.

MrS

---

Scanning for our furry friends since Jan 2002

Thank you, MC ~48% at the OC mode, switched to normal mode (=MSI Core OC, Memory stock) MC ~46%, underclocking to Core @928MHz not possible, the offset -65 was not accepted.


Latest runtimes in s with this oc setting (left screen):
- 8,791
- 8,822
- 8,765 : mean 8793

Result +14% (1461/10254). That´s o.k., as long as this setting produces valid results I can live with it.

Should I try a long run workunit ? The hostid 135701 i.e. never crunched one of them, maybe they are reserved field-tested to the really fast GPUs. I´m quite happy to join with this starter peace of Kepler.

Should I try a long run workunit ?

Yes...

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

Two long run workunits finished (GTX 650 TI; GK106, 1GB), looks like these are different, sorry but not enough samples to show effect of OCing:

- 29,753 s: 60,900 credits (run @stock frequencies) /VRAM used: 965 MB - influence of VRAM 2GB ?
- 41,032 s: 103,500 credits (run @Core +207 & Memory +400 Mhz) /VRAM used: 920 MB
(+14% faster as 48,030 s, not knowing his frequencies)

The memory usage is higher than with short run workunits, nearby the built-in limits of 1 GB.



For comparing I searched for related performance levels:

first example of ...GTX 650 TI (GK106; 1GB) ... Linux !
- 28,771 s: 60,900 credits
- 48,030 s: 101,400 credits (base for OC effect, see above, if valid)
- 60,746 s: 113,400 credits - one more step longer workunits are distributed ?

second example of ... GTX 650 TI (GK106; 2GB) ...
- 28,436 s: 60,900 credits
- 46,178 s: 103,500 credits
- 54,151 s: 101,850 credits - bonus missed ?

example of ... GTX 660M (GK107) ...
- 41,402 s: 60,900 credits
- 69,650 s: 101,400 credits

example of ... GT 640 (GK107) ... Linux !
- 42,481 s: 60,900 credits
- 70,912 s: 101,400 credits
- 68,785 s: 103,500 credits

example of ... GTX 560 (GF110 or 114; 336 or 384 shaders) ...
- 29,081 s: 60,900 credits
- 50,328 s: 101,400 credits
- 63,407 s: 113,400 credits

exmple of ... GTX 560 TI (GF 110 or 114; 352 or 384 or 448 shaders) ... Linux !
- 20,379 s: 60,900 credits
- 34,320 s: 101,400 credits

example of ... GTX 660 (GK106) ...
- 20,516 s: 60,900 credits
- 34,253 s: 101,400 credits

My conclusion
No errors so far, in spite of extensive OCing. Overvolting at highest level, probably that helped.
The GTX 650 TI runs respectable with stock or OCed. Between 1GB and 2 GB no apparent difference. I cannot understand the kind of workunits/credit association, but that doesn´t matter, more interesting I think is the time period to finish a certain credit level. The CPU sharing in the progress appears high (CPU time) but indifferent if Windows or Linux.

I am not able to analyze all factors, some linked examples may be OCed in one or all linked results, so this synopsis may contain a lot of mistakes. In the end this investigation isn´t significant at all. Hope you enjoyed the show .. ehm ... reading at least. ;-)

Thanks for collecting and sharing that data. This should be enough for anyone who wants to know how well this card (and some comparably fast ones) performs :)

MrS

---

Scanning for our furry friends since Jan 2002