GTX750Ti... wait a sec...

SunnyD

SunnyD

2[H]4U
2FA
Joined
Jul 6, 2004
Messages
3,441
No PCIe power connector, 60w TDP... 250KH/s without optimizations, potentially 280-300KH/s with optimization+overclock.

Not sure how Nvidia's drivers play with many multiple cards, but I'm sure if they did their homework they want to kick AMD's ass. Sooooo....

Are we talking a motherboard, low end Ivy/Haswell CPU and say 6 of these GTX750Ti's being powered by a SINGLE decent 800 watt PSU pushing 1.5MH/s?

If only the price/performance on these cards was better. Expected market segment is going to be $149. Make it $100 and you got a whole new king for the mining wars. Heck, you're looking at whole 1MH+ mining rigs being built for peanuts at that level.
 
250KH/s from a low end Nvidia card with no power connector?

I'd have to see it to believe it.
 
It's interesting...

60w nets you 265 kh/s even without optimizations on a stock reference card which is what 1k boost or thereabouts? How much will bandwidth effect performance, it has little to no effect on my 780 and it's pretty clear from the 1400MHz Palit OC from Guru the card is severely bandwidth starved on the core at those clocks.

Then there is the 6 pin cards coming out, as well as the possibility of skynet bios unlocking the power.. How high are these cores going to go with more voltage and more power allowed? And how will that translate into kh/s and kh/w?

I think we'll need to see how that plays out, but I have an itch to grab a Maxwell from PhysX and when not using it for that making some of the cost back mining it (in secret since it will be silent the women folk never have to know).

$150 for the KH/s is steep, but whats build cost difference when you factor in several cards and the PSU needed to run them? Say you go 7 cards, right now that is 1855 kh/s @ 420w.

260x would be 1500 kh/s @ stock I believe, and 805w consumption.

That doesn't seem right though someone correct me!

My only real hesitation at getting a 750 Ti for PhysX and light mining is the idea that these cards are so efficient Nvidia could come out with something faster like 760 performance (760 Ti perhaps, OEM not withstanding) and then I'd feel stupid.
 
I agree, is it actually verified or speculation?

seeing that it has a 128bit mem bus, and scrypt is heavily memory dependant, unless you're talking about bitcoin(SHA256) mining :p
 
It is for sure with AMD, with my 7950s I overclocked the ram to 1700 because it made a fairly big difference.

With my 780 I set the memory to -500 and it has no impact on my kh/s.

I find it interesting, gonna have to keep up on it to see how it plays out.
 
BallaTheFeared said:
It is for sure with AMD, with my 7950s I overclocked the ram to 1700 because it made a fairly big difference.

With my 780 I set the memory to -500 and it has no impact on my kh/s.

I find it interesting, gonna have to keep up on it to see how it plays out.
Click to expand...

If you read the explination by thestilt on litecoin mining forum, the way he explains it is this(simplified)

the amount of work your gpu is capable of has to saturate your memory bandwidth completely before increasing memory speed can improve your hash rate.

EG: your 780 can't saturate your memory @ -500mhz from stock, where as your 7950 can saturate it all the way up to 1700.

Also you have to consider timing, the modified 290 bios actually allows LOWER memory speed now(Vs stock bios, without losing perf) since the timings have been adjusted, I'm currently running 1385mhz on memory vs 1500mhz before.

So the combination of timing + how much your gpu can push have a lot to do with it.
 
Digital Viper-X- said:
I agree, is it actually verified or speculation?

seeing that it has a 128bit mem bus, and scrypt is heavily memory dependant, unless you're talking about bitcoin(SHA256) mining :p
Click to expand...

cris buchner (he's the guy that updates cudaminer) has said that the 750Ti is getting 250KH/s with the current version of cudaminer... and he doesn't have access CUDA 6.0 yet. He's expecting close to 280KH/s once he gets access to it and updates cudaminer.

Tom's Hardware benched a 750Ti... on LiteCoin - obviously with the current version of cudaminer:

Litecoin.png


They obviously don't know how to optimize AMD cards, but anyway... cudaminer does a pretty good job optimizing Nvidia cards right out of the box. There ya go.
 
Digital Viper-X- said:
If you read the explination by thestilt on litecoin mining forum, the way he explains it is this(simplified)

the amount of work your gpu is capable of has to saturate your memory bandwidth completely before increasing memory speed can improve your hash rate.

EG: your 780 can't saturate your memory @ -500mhz from stock, where as your 7950 can saturate it all the way up to 1700.

Also you have to consider timing, the modified 290 bios actually allows LOWER memory speed now(Vs stock bios, without losing perf) since the timings have been adjusted, I'm currently running 1385mhz on memory vs 1500mhz before.

So the combination of timing + how much your gpu can push have a lot to do with it.
Click to expand...


Explain why my 780 gets 670 kh/s with 1250 mem and my 7950 got 700 kh/s with 1750 memory, 650 with 1500 mem?

670 kh/s not enough to saturate Kepler at 1250 but easily smothers Thaiti at similar kh/s?

The explanation was for AMD cards, I don't think you can apply it to cudaminer and Nvidia cards.
 
SunnyD said:
cris buchner (he's the guy that updates cudaminer) has said that the 750Ti is getting 250KH/s with the current version of cudaminer... and he doesn't have access CUDA 6.0 yet. He's expecting close to 280KH/s once he gets access to it and updates cudaminer.

Tom's Hardware benched a 750Ti... on LiteCoin - obviously with the current version of cudaminer:

Litecoin.png


They obviously don't know how to optimize AMD cards, but anyway... cudaminer does a pretty good job optimizing Nvidia cards right out of the box. There ya go.
Click to expand...

Well colour me shocked :)
 
SunnyD said:
cris buchner (he's the guy that updates cudaminer) has said that the 750Ti is getting 250KH/s with the current version of cudaminer... and he doesn't have access CUDA 6.0 yet. He's expecting close to 280KH/s once he gets access to it and updates cudaminer.

Tom's Hardware benched a 750Ti... on LiteCoin - obviously with the current version of cudaminer:

Litecoin.png


They obviously don't know how to optimize AMD cards, but anyway... cudaminer does a pretty good job optimizing Nvidia cards right out of the box. There ya go.
Click to expand...

When you say they don't know how to optimized the AMD cards, what are they getting?

That looks pretty good for hash/watt on the 750 ti
 
BallaTheFeared said:
Explain why my 780 gets 670 kh/s with 1250 mem and my 7950 got 700 kh/s with 1750 memory, 650 with 1500 mem?

670 kh/s not enough to saturate Kepler at 1250 but easily smothers Thaiti at similar kh/s?

The explanation was for AMD cards, I don't think you can apply it to cudaminer and Nvidia cards.
Click to expand...

I believe Nvidia GPU's have a lot more cache onboard, which masks a lot of any latency deficiencies that accessing external memory might have. That's part of the reason AMD GPU's have historically been smaller and why AMD went with GDDR5 quicker than Nvidia did - because they shifted how they dealt with latency issues to a more direct methodology. With Nvidia having more cache onboard per cluster, they can do more work without hitting the external memory, thereby not having nearly as much impact by memory speed/latency issues because there's less need to access the memory as often.

I'm mostly speculating here though - but I'm drawing from what I know of historical differences between the two competing architectures.

wand3r3r said:
When you say they don't know how to optimized the AMD cards, what are they getting?

That looks pretty good for hash/watt on the 750 ti
Click to expand...

Look at the chart. 292KH/s for a 270 and 302KH/s for a 270x kinda sucks.
 
BallaTheFeared said:
Explain why my 780 gets 670 kh/s with 1250 mem and my 7950 got 700 kh/s with 1750 memory, 650 with 1500 mem?

670 kh/s not enough to saturate Kepler at 1250 but easily smothers Thaiti at similar kh/s?

The explanation was for AMD cards, I don't think you can apply it to cudaminer and Nvidia cards.
Click to expand...

I thought the explination was for scrypt mining in general.
but I don't know enough to say otherwise.

http://products.ncix.com/detail/msi-geforce-gtx-750-ti-b5-94307.htm

So tempted :D
the problem is ofcourse, while these have a GREAT KH/W , they have a reallly crappy KH/$ and KH/SLOT, I assume people will buy these up and price will get gouged making it even worse.

But it's not a bad card to toss into your htpc, and mine it's value back.
 
Last edited:
Radeons and Nvidia cards do their calculations differently cause their architectures are different simple as that. Nvidia historically wanted to crunch simple, fast, and throw numbers results as fast as possible, Radeon wanted a steady input medium speed more complex end results. bitcoin or hashcracking in general shows the differences quite well going back many generations. I know Radeon is not VLIW based anymore and GTX are different then a couple years back, however, just running vectors, workunit total and such you can still tell there is a pretty distinct difference between both camps and gaming wise also proves this.

Yeh toms doesn't do very good testing with anything, its just a within the ballpark range and certainly nowhere close when it comes to Radeons mining performance with SHA or Scrypt.
 
Another concern is with lack of VRM cooling, this card will be running at 100% load 24/7, will it be able to last? even though it's only 68W being pulled.
 
Digital Viper-X- said:
So tempted :D
the problem is ofcourse, while these have a GREAT KH/W , they have a reallly crappy KH/$ and KH/SLOT, I assume people will buy these up and price will get gouged making it even worse.

But it's not a bad card to toss into your htpc, and mine it's value back.
Click to expand...

Not for $199 :eek:

Yes it's interesting from the standpoint that it uses so little power, produces decent enough kh/s it could pay for itself in a few months, with almost no noise pollution... HTPC comes to mind, dedicated PhysX card you can mine when not using it for that purpose while still gaming on your main card as well.
 
BallaTheFeared said:
Not for $199 :eek:

Yes it's interesting from the standpoint that it uses so little power, produces decent enough kh/s it could pay for itself in a few months, with almost no noise pollution... HTPC comes to mind, dedicated PhysX card you can mine when not using it for that purpose while still gaming on your main card as well.
Click to expand...

well, that's the MSI TwinFrozr version, the EVGA one is sitting @ $169 (also this is Canadian, current exchange rate difference is about 10%)

what is realllly interesting is what a big maxwell chip will be able to produce!
 
SunnyD said:
I believe Nvidia GPU's have a lot more cache onboard, which masks a lot of any latency deficiencies that accessing external memory might have. That's part of the reason AMD GPU's have historically been smaller and why AMD went with GDDR5 quicker than Nvidia did - because they shifted how they dealt with latency issues to a more direct methodology. With Nvidia having more cache onboard per cluster, they can do more work without hitting the external memory, thereby not having nearly as much impact by memory speed/latency issues because there's less need to access the memory as often.

I'm mostly speculating here though - but I'm drawing from what I know of historical differences between the two competing architectures.



Look at the chart. 292KH/s for a 270 and 302KH/s for a 270x kinda sucks.
Click to expand...
What's normal for a 270/x since you are saying those rates in the graphs are poor. I saw the charts. ;)
 
wand3r3r said:
What's normal for a 270/x since you are saying those rates in the graphs are poor. I saw the charts. ;)
Click to expand...

An optimized 270X should be pushing 400KH/s without an overclock, 450+KH/s with one. A 270 should be close to 400KH/s.
 
Well with Nvidia anything is possible, it might be running quite cool, but is probably not at all meant ot have a 24/7 load on it, I know past few gens of cards were quite energy efficient but also had high amps given to the core which did tend to make the Vreg run very warm. Wouldn`t put it past them to have skimped on cooling or a proper spec Vreg for the card.

Many buddies of mine from the 200 series all the way to the 600 series had cards just up and stop working for no apparent reason, could be a variety of things really, in this case, low energy card with a price premium because of the name attached to it and cause reality is, it is a very low energy card and seemingly excellent performance/power, anyways only time will tell I suppose.

Hope they made a great card in specs and actual pcb instead of skimping which even from a reference standpoint they like to do many times, low end cards does not matter quite as much but seems a lot to ask for in the context they charge a premium due to the Geforce name.
 
This falls between a Radeon 290 and a scrypt ASIC in power efficiency and kH/W efficiency. The hash per dollar is much better than the first gen scrypt ASICs, though.
 
IEC said:
This falls between a Radeon 290 and a scrypt ASIC in power efficiency and kH/W efficiency. The hash per dollar is much better than the first gen scrypt ASICs, though.
Click to expand...

In theory perhaps, not so much in practice. You need too many mobo slots to equal the hashrate which means much more overhead. It's also not really viable at 150$+ per card.

It does however indicate that perhaps when the high end cards get this architecture, nVidia could be the ones with supply/price issues for a change. I don't expect the high end to be quite this efficient, but it can afford to lose quite a bit as they scale it up and still be impressive.
 
SunnyD said:
I knew this from 15 years ago. Tell me something I didn't know.
Click to expand...

lol, its probably had three different owners over those 15 years.

Should be getting some user results in a few days, should shed some more light on it.

My 7950s did 1000 kh/s @ 330w or 1300 @ 550w, my 780 has pretty awful kh/w by comparison.

Probably wouldn't consider farms with the 750 Ti, but if I still had AMD for my main card I'd consider one for hybrid physx and mining.
 
A 7950 that does 1000KH/s?

Hrm, no. I'm calling BS.

The 270 cards will all do over 450 with almost no tweaking. My 270X cards will do 480-500KH/s pretty easily. Between the 270-(X) cards I have over a dozen of them.

Will be interesting to see what people get from these 750ti cards. I find it hard to believe you can get 260KH/s from them with only 70w. Even if this is true, we'll have the same problem we have now, which is running out of PCI-X slots on the motherboard. Even the Asrock BTC board only has 6.
 
If a 270 is 400-500 kHash this thing doesn't even look that special. At first glance with the skewed numbers (toms low number on the AMD cards) it looked like it was almost a decent hashrate with low power consumption. In the end you can only fit so many in a MB plus the hashrate isn't even close to the 4-500 hash.

Basically what it has going for it is low power consumption and that's about it. It can't touch 2 high end cards even if you fill the slots on your board, and with high end cards you can put 6 on a board. Once you factor in CPU/MB/RAM etc. it's nothing like the first impression I was getting about this card. Either way it's good to see some improvement.
 
Pitbully said:
A 7950 that does 1000KH/s?

Hrm, no. I'm calling BS.

The 270 cards will all do over 450 with almost no tweaking.
Click to expand...

Sorry, I had two and that was combined in total hashing power.

Didn't know the 270 hashed so well until Google.

With an OC 450 kh/s at probably ~180w not bad at all.
 
wand3r3r said:
If a 270 is 400-500 kHash this thing doesn't even look that special. At first glance with the skewed numbers (toms low number on the AMD cards) it looked like it was almost a decent hashrate with low power consumption. In the end you can only fit so many in a MB plus the hashrate isn't even close to the 4-500 hash.

Basically what it has going for it is low power consumption and that's about it. It can't touch 2 high end cards even if you fill the slots on your board, and with high end cards you can put 6 on a board. Once you factor in CPU/MB/RAM etc. it's nothing like the first impression I was getting about this card. Either way it's good to see some improvement.
Click to expand...

To be fair - I did point all of these things out going right in at the top.

#1 - the price is a tad high. I really needs to come closer to $100 to be competitive on a hash per dollar level.

#2 - the density issues - but again, that really factors directly into price per dollar.

But where it DOES shine is the potential to build out an entire relatively low TDP mining rig without the need to spend $200 on a PSU, possibly smaller enclosure area, etc. The idea of having less of a need for specialty parts is appealing.

Imagine if some enterprising manufacturer releases a single-slot PCIe 1x version of this card... you've just shaved off another $100 in riser costs and potentially the need for a custom chassis. It's the notion of what a low TDP/low draw card with a decent hash rate brings to the table that excites me. All it needs now is a more reasonable price expectation.

Hell, at that point you just put ASIC's on the back burner for a while.
 
wand3r3r said:
If a 270 is 400-500 kHash this thing doesn't even look that special. At first glance with the skewed numbers (toms low number on the AMD cards) it looked like it was almost a decent hashrate with low power consumption. In the end you can only fit so many in a MB plus the hashrate isn't even close to the 4-500 hash.

Basically what it has going for it is low power consumption and that's about it. It can't touch 2 high end cards even if you fill the slots on your board, and with high end cards you can put 6 on a board. Once you factor in CPU/MB/RAM etc. it's nothing like the first impression I was getting about this card. Either way it's good to see some improvement.
Click to expand...

290 wasn't looking good at release either, maybe give the program and users time to mess with it a bit before writing it off. 450 for the 270 is with an OC and considerably more power.

$260 for 300 kh/s before OC @ 125w vs $150 for 265 kh/s @ 60w with an unoptimized miner. Neither option is going to compete with the raw power of several high end solutions.
 
BallaTheFeared said:
290 wasn't looking good at release either, maybe give the program and users time to mess with it a bit before writing it off. 450 for the 270 is with an OC and considerably more power.

$260 for 300 kh/s before OC @ 125w vs $150 for 265 kh/s @ 60w with an unoptimized miner. Neither option is going to compete with the raw power of several high end solutions.
Click to expand...

Yep, the dev already ordered some 750 Tis and is planning on optimizing the miner for Maxwell. From Bitcointalk:

the 250 kHash/s is without specific optimizations for this platform and with factory clocks. The numbers I've seen @ nVidia were in the 270-280 kHash/s range. I expect to reach these numbers once I get access to the hardware and use the CUDA 6.0 toolkit.

Density will be achieved when OEMs create 2 or 4 chip versions of Maxwell GPUs on single PCBs. Four should be within reasonable power limits (240 Watts? can do!). Maybe ASUS will create a new iteration of their MARS series.

Christian
Click to expand...
 
You must log in or register to reply here.
Back
Top