Vega Rumors

[Simplyfun]

Completely agree. Great Post! I would like to add in there the only other benefit of the 1070 in mining though is that is can do other coins faster (cuda based mining programs), like Zcash than radeons for now, I don't know if the radeons can catch up to this, cause I haven't tried to mine zcash yet, it might be something to do with the cuda algorithm. So once the Eth crazy goes which should be early next year when they shift over to POS, the 1070's have options to go to more coins.

Speaking of the best case for Nvidia which is Zcash
I've already experimented with this stuff, a lot. The 1070 needs basically to be unrestricted and wide open to get a hash rate of around 455. I mean you have to push it over 2100 core and over 4500 RAM and then it's sucking down 185 watts plus. If you dial it back reasonably to get into that magic 150 watt mark, after the tweaking and underclocking, you get a hash rate of around 390 which is actually quite good at 150W ... zero complaints. .385 watts per hash/s This is the best case for a 1070 imo. (per watt the 1080 is a bit better than this)

With the 480's I cut the vcore to .91 or .905 (less if stable ) drop the frequency to 1100-1150 on GPU and 2200-2250 on RAM and I get about 290-295 ish with an unmodded BIOS. modded BIOS and hacked drivers I get more. All this in about a 105-115watt envelope. .396watts per hash/s . This is the worst case for a 480 imo.

The above figures do not represent days of intensive scrutiny, that would be stupid. It was the best I could do on the bench over a couple days.

So it's a coin flip for Zcash and for everything else Nvida gets eaten alive in H/s per watt. There are no coins the AMD's don't mine well (in comparison to Nvidias) so right now they remain the better more flexible choice for long term usability in all kinds of alt coins. What I'm hoping we'll see is more and better custom built cuda miners that take advantage of the extremely high GPU clocks of Nvidias in some way I can't really think of.

Anyhow, the point of this mental exercise is simply as SighTurtle stated;

Well that scenario sounds horrifying for anyone wanting a Vega card.

edit: For gaming.

Yeah, absolutely, I think the midrange Vega is going to be a mining beast with extreme efficiency once those clocks are dialed back to match the sweet spot on the Glo Flo fabrication process. Because we all know it's going to hit the market already wrung out with no head room to go up, that means going down and it'll still probably slay all the other GPU's out there in consumer land as far as mining goes. So launch prices are going to suck, they're not going to drop and availability is going to be painful - unless the mining market suffers a horrible doom between now and launch.

 

[N4CR]

The real AMD issue with power use is their voltage binning is too wide. Once you put both on equal footing, Polaris is more than competitive.. it's actually more efficient, which is a hard pill to swallow after all the shitting on Polaris.

Is there any way AMD can realistically tighten up binning for better efficiency in production? I wonder if people would pay a slight premium for a more power efficiently binned card, out of the box. Or you just DIY.

 

[Simplyfun]

The real AMD issue with power use is their voltage binning is too wide. Once you put both on equal footing, Polaris is more than competitive.. it's actually more efficient, which is a hard pill to swallow after all the shitting on Polaris.

Is there any way AMD can realistically tighten up binning for better efficiency in production? I wonder if people would pay a slight premium for a more power efficiently binned card, out of the box. Or you just DIY.

I can't speak to gaming, because I don't game much but I can't really see an advantage in gaming scenarios where bringing a Polaris into it's efficient sweet spot is a good thing for gamers (it would mean lower framerates). Who would want that? Only miners really. And AMD doesn't make cards for mining, just for gaming. So far.

 

[razor1]

The real AMD issue with power use is their voltage binning is too wide. Once you put both on equal footing, Polaris is more than competitive.. it's actually more efficient, which is a hard pill to swallow after all the shitting on Polaris.

Is there any way AMD can realistically tighten up binning for better efficiency in production? I wonder if people would pay a slight premium for a more power efficiently binned card, out of the box. Or you just DIY.

not really, binning is luck of the draw, the main problem is the architecture, they need to make the architecture require less voltage. Once that is done, the binning becomes less of an issue.

 

[N4CR]

I can't speak to gaming, because I don't game much but I can't really see an advantage in gaming scenarios where bringing a Polaris into it's efficient sweet spot is a good thing for gamers (it would mean lower framerates). Who would want that? Only miners really. And AMD doesn't make cards for mining, just for gaming. So far.

It's not a huge difference but Polaris was being crucified for a while (even in passing) regarding efficiency. So such a change would nullify that difference.

not really, binning is luck of the draw, the main problem is the architecture, they need to make the architecture require less voltage. Once that is done, the binning becomes less of an issue.

So this is where all the Navi eggs in one basket come to roost I guess...

 

[razor1]

probably, but Navi is going up against Volta, which if nV reaches anything close to what they are claiming, man that is going to be hard to beat. They are proclaiming twice the efficiency on essentially the same 16nm node (tweaks).

 

[CSI_PC]

Speaking of the best case for Nvidia which is Zcash
I've already experimented with this stuff, a lot. The 1070 needs basically to be unrestricted and wide open to get a hash rate of around 455. I mean you have to push it over 2100 core and over 4500 RAM and then it's sucking down 185 watts plus. If you dial it back reasonably to get into that magic 150 watt mark, after the tweaking and underclocking, you get a hash rate of around 390 which is actually quite good at 150W ... zero complaints. .385 watts per hash/s This is the best case for a 1070 imo. (per watt the 1080 is a bit better than this)

With the 480's I cut the vcore to .91 or .905 (less if stable ) drop the frequency to 1100-1150 on GPU and 2200-2250 on RAM and I get about 290-295 ish with an unmodded BIOS. modded BIOS and hacked drivers I get more. All this in about a 105-115watt envelope. .396watts per hash/s . This is the worst case for a 480 imo.

The above figures do not represent days of intensive scrutiny, that would be stupid. It was the best I could do on the bench over a couple days.

So it's a coin flip for Zcash and for everything else Nvida gets eaten alive in H/s per watt. There are no coins the AMD's don't mine well (in comparison to Nvidias) so right now they remain the better more flexible choice for long term usability in all kinds of alt coins. What I'm hoping we'll see is more and better custom built cuda miners that take advantage of the extremely high GPU clocks of Nvidias in some way I can't really think of.

Anyhow, the point of this mental exercise is simply as SighTurtle stated;



Yeah, absolutely, I think the midrange Vega is going to be a mining beast with extreme efficiency once those clocks are dialed back to match the sweet spot on the Glo Flo fabrication process. Because we all know it's going to hit the market already wrung out with no head room to go up, that means going down and it'll still probably slay all the other GPU's out there in consumer land as far as mining goes. So launch prices are going to suck, they're not going to drop and availability is going to be painful - unless the mining market suffers a horrible doom between now and launch.

I appreciate that is your experience but some considerations.
For ZCash and Nvidia you really need one of the solutions that uses CUDA and ideally CUDA8 - mentioning generally and I appreciate maybe this is what you were doing.
Importantly though and this is missed by many the only way to see the actual power demand on recent AMD GPU is to isolate it and measure usually with an oscilloscope or something similar (next closest to this is hardware.fr).
What you see with programs/utilities is not the actual power demand for the GPU, this was made pretty clear by PCPer/Tom's Hardware/Hardware.fr who are the only ones to measure cards correctly and accurately.
Some 'professional' miners who really know how to setup your mining algo/ensure supports CUDA 8 ideally (avoiding OpenCL) and Nvidia card you can hit even 36 MH/s with Claymore and that is stable with multiple 1070s,generally you still can hit 32 MH/s without same level of knowledge.
I notice one article that went on to test this and used CUDA 7.5 instead of CUDA 8 (introduced ideally for Pascal) with the Pascal 1070 even though it was possible.

With the right implemention you should not see Nvidia hitting 180+ watts and a 580 at 150W matching it (although like I mention you only get accurate figures as done by 3 sites and these are higher than what is shown by utilities/etc for AMD GPUs); tbh the ideal can be to overclock memory and undervolt the 1070 (that keeps it below 160W) - this is from some of the professional crypto mining blogs.
The difference is that you really do not have much choice but to undervolt the 580 relative to the 1070.

One reason probably not best for most to use the 1070 is it seems to require a fair bit of fiddling to get the optimum performance and importantly one has more options with AMD hardware, but the 1070 does have the performance/efficiency where those with high knowledge optimise the mining setup.
And worth noting the power demand has gone through the roof for the 580 relative to the 480.
This was the envelope for 480 and 1060 (unfortunately not 1070):

And the compute figures 480 with 5.1 TFLOPs at reference boost while 1070 is 5.7 TFLOPs at reference boost (but in reality the figure is quite a bit higher).
Reducing clock rates of any GPU lowers its compute capability and one cannot get around that, although fair to say some/maybe many of the algos are more memory sensitive in terms of tweaks.

I agree it is simpler,easier, and with more options to use the AMD GPUs but it does not necessarily equal Nvidia GPUs (more expensive to go 1070 but 580s are now getting expensive) in terms of performance/efficiency when both are setup ideally.
And I agree if one is already using 480/580s it makes sense to continue buying these to expand the operation even if the price is not as competitive as it used to be.
Cheers

 

[Bahanime]

probably, but Navi is going up against Volta, which if nV reaches anything close to what they are claiming, man that is going to be hard to beat. They are proclaiming twice the efficiency on essentially the same 16nm node (tweaks).

Twice? Where the heck did NV claim that?

Volta is a pretty big leap though, ~40% more TFlops at the same power 300W, but not 100% more.

 

[CSI_PC]

Twice? Where the heck did NV claim that?

Volta is a pretty big leap though, ~40% more TFlops at the same power 300W, but not 100% more.

Yeah not quite unless one is talking about DL related operations/instructions.

40% more cores but depending upon the FP32 operation can be up to 1.8x performance gains with cuBLAS, other applications shown so far can be 1.5x to 1.75x
Also it can do simultaneously certain Int/floating/etc operations within that TDP.

This is ignoring the Tensor cores.
Also need to consider there is a higher TDP also due to the improvements with Mezzanine NVLink2 over what is in the P100; 50% more connections and greater BW per connection.
That aside of more interest to most of us will be GV102 as it is without FP64 cores that take a fair bit of space and power/TDP relative to FP32 functioning cores.
Cheers

 

[Simplyfun]

I appreciate that is your experience but some considerations.

but the 1070 does have the performance/efficiency where those with high knowledge optimise the mining setup

SNIP
I agree it is simpler,easier, and with more options to use the AMD GPUs but it does not necessarily equal Nvidia GPUs (more expensive to go 1070 but 580s are now getting expensive) in terms of performance/efficiency when both are setup ideally.
And I agree if one is already using 480/580s it makes sense to continue buying these to expand the operation even if the price is not as competitive as it used to be.
Cheers

The wall of text of is huge but, I was using the cuda miner, I isolated the cards onto their own power supply measured the draw with a known load at 200w and then did the cards. my readings are accurate and account for PSU efficiency. I would have bought 100 if I could have convinced myself it was worth it so no point messing around.

In the ETH mining sample you gave... sure it will do that. Now drop the H/s per watt to what we get out of detuned 480's and it won't come within 15% of that efficiency.

As I said before, 1070's are a strong second choice, regardless of cost of acquisition. I had no complaints and they are a better choice than a Fury.

 

[razor1]

Twice? Where the heck did NV claim that?

Volta is a pretty big leap though, ~40% more TFlops at the same power 300W, but not 100% more.

Can't look at server parts, cause server parts have different requirements. But as a server part, they have much more than x2 the efficiency. As CSI_PC stated, 1.8x is what they are predicting.

The x2 is what was in previous time lines and nV's CEO saying. Probably a rounded off figure or target figure.

 

[CSI_PC]

The wall of text of is huge but, I was using the cuda miner, I isolated the cards onto their own power supply measured the draw with a known load at 200w and then did the cards. my readings are accurate and account for PSU efficiency. I would have bought 100 if I could have convinced myself it was worth it so no point messing around.

In the ETH mining sample you gave... sure it will do that. Now drop the H/s per watt to what we get out of detuned 480's and it won't come within 15% of that efficiency.

As I said before, 1070's are a strong second choice, regardless of cost of acquisition. I had no complaints and they are a better choice than a Fury.

Do you know if it was CUDA 7.5 or CUDA8?
Still do not see how you are working out greater efficiency for a 480 vs a 1070 both tuned one is around 28MH/s and the other 33MH/s (1070) without being fully overclocked if using CUDA8 and setup well; fully overclocked an etherminer who does a blog was hitting stable 36MH/s at a clock speed of 1900Mhz with multiple and single GPU 37MH/.
The chart above is using an oscilloscope measuring a 480, their values are pretty much in line with PCPer who also used a 480; at 1200Mhz it is around 145W and below reference boost so compute drops below 5 TFLOPs.
A 580 is roughly anywhere from 200W to 230W but with higher clocks and compute, again measured and isolated by a scope.

Decided to take some time to find a reasonable reference for my point.
This is from a blog that used CUDA 7.5 instead of CUDA 8 that improves this if using right algo and setup and they still managed this:

we came out with an optimized configuration that hashes at 32.4 MH/s, pulling 146 watts under load, and getting up to a maximum of 63 degrees in out test bed case.

http://bitcoinist.com/gtx-1070-optimized-ethereum/

And their own chart analysis on perf/efficiency (albeit using CUDA 7.5 like I mentioned) and is based upon their earlier review before they managed to fully optimise the 1070 so is 25.6 MH/s rather than the 32.4 MH/s at 140W.
The 32.4 MH/s required only 6 watts more at 146W in their optimised test later on, so the chart below improve much more for the 1070 when also taking the optimised final results into consideration.

But like I said, it is not easy to do mining with Nvidia products compared to AMD.
Cheers

 

[Drewis]

+1, did the math too. exactly on point Simplyfun.

The wall of text of is huge but, I was using the cuda miner, I isolated the cards onto their own power supply measured the draw with a known load at 200w and then did the cards. my readings are accurate and account for PSU efficiency. I would have bought 100 if I could have convinced myself it was worth it so no point messing around.

In the ETH mining sample you gave... sure it will do that. Now drop the H/s per watt to what we get out of detuned 480's and it won't come within 15% of that efficiency.

As I said before, 1070's are a strong second choice, regardless of cost of acquisition. I had no complaints and they are a better choice than a Fury.

 

[CSI_PC]

+1, did the math too. exactly on point Simplyfun.

I guess it comes down to who does the setup *shrug*.
I can provide another separate professoinal mining blog backing up my last post regarding power efficiency and 1070 actually being better; price and flexibility I agree less so.
Cheers

 

[Simplyfun]

I guess it comes down to who does the setup *shrug*.
I can provide another separate professoinal mining blog backing up my last post regarding power efficiency and 1070 actually being better; price and flexibility I agree less so.
Cheers

I don't blog, but I do mine, prior to the launch of LTC ASICS I was drawing around 255,000 watts (three phase, 800 amp service - I bought an empty grade school with bitcoin cash - I still have the school ).

Everyone you quote is an expert, far be it from me to get tangled up with experts. I encourage everyone to find a practical means of doing their own assessment.

If you're mining with 1-10 cards, who really cares though as these percentages have negligible impact. at that point it's more cost of acquisition and we see how that's going and how it's going to impact Vega.

 

[razor1]

I don' t know man with my test 1070 stock I'm getting 25 mhs, overclocked 31 mhs, remember this is a reference board dual mining. So if I only do eth, I'm getting 33 mhs, I haven't even tried pushing it to max, cause I don't really care to do that, only upped the power usage to max, and frequency of the core to +200mhz, I haven't pushed it to the max either.

PS all the while its only pulling 110 watts.

Now my 580 which is getting 29mhs/s maxed out, dual mining, is pulling 150 watts.

I'll have 5 1070 rigs up and going this coming week, so I'll keep ya posted on em, but I'm expected to get 35mhs per card dual mining, at 150 watts.

 

[TheLAWNoob]

I don't blog, but I do mine, prior to the launch of LTC ASICS I was drawing around 255,000 watts (three phase, 800 amp service - I bought an empty grade school with bitcoin cash - I still have the school ).

Everyone you quote is an expert, far be it from me to get tangled up with experts. I encourage everyone to find a practical means of doing their own assessment.

If you're mining with 1-10 cards, who really cares though as these percentages have negligible impact. at that point it's more cost of acquisition and we see how that's going and how it's going to impact Vega.

Damn, an empty grade school.

My plan is to work for AMD or nVidia for a few years after I graduate, then buy a factory and fill it up with GPUs.

I want to make my own drivers and miners, along with custom designed motherboards.

 

[Simplyfun]

Damn, an empty grade school.

My plan is to work for AMD or nVidia for a few years after I graduate, then buy a factory and fill it up with GPUs.

I want to make my own drivers and miners, along with custom designed motherboards.

I used to make motherboards, we made the first serious OC'd 386 mobos, 16 Mhz to 22Mhz , which we guaranteed. We even binned the co processors for that speed, which was hard, not many would run at that. Then IBM snagged me and it got a little more boring.

 

[razor1]

Ya ever come to the famous IBM tent conferences in Endicott?

 

[Simplyfun]

Ya ever come to the famous IBM tent conferences in Endicott?

No, I mostly stayed in Canada during that period, my career took off after that and I left IBM and traveled a lot... too much. So I retired.

 

[razor1]

ah cool!

 

[noko]

Don't we have a mining thread for this? Vega Rumors should be about Vega Rumors.

 

[KazeoHin]

Don't we have a mining thread for this? Vega Rumors should be about Vega Rumors.

Well, then the thread would be dead. We all know what Vega is going to be:








A disappointment

 

[Maddness]

Don't we have a mining thread for this? Vega Rumors should be about Vega Rumors.

True. But the problem there is, there are no rumors. AMD has been so tight lipped. Good or Bad.

 

[N4CR]

probably, but Navi is going up against Volta, which if nV reaches anything close to what they are claiming, man that is going to be hard to beat. They are proclaiming twice the efficiency on essentially the same 16nm node (tweaks).

Don't see how that's possible but with that R&D budget you'd sort of hope so. 1.4-1.5x performance perhaps, with expected uarch + bigger die up from 480mm² to 550-600mm² 250W+ I can see that happening.

Yeah not quite unless one is talking about DL related operations/instructions.

40% more cores but depending upon the FP32 operation can be up to 1.8x performance gains with cuBLAS, other applications shown so far can be 1.5x to 1.75x
Also it can do simultaneously certain Int/floating/etc operations within that TDP.

This is ignoring the Tensor cores.
Also need to consider there is a higher TDP also due to the improvements with Mezzanine NVLink2 over what is in the P100; 50% more connections and greater BW per connection.
That aside of more interest to most of us will be GV102 as it is without FP64 cores that take a fair bit of space and power/TDP relative to FP32 functioning cores.
Cheers

They're going up in die size from 480mm² to ~600mm² which will account for much of the performance gain too.

 

[noko]

Welp, I've pulled down all my images of benchmarks with overclocked CPU :\

Apparently when speeds are adjusted from outside of the BIOS (like I was doing, using K17TK program), the Real Time Clock issue with Windows 8/10 causes benchmarks to incorrectly report their performance. So my changing of speeds to 3.8GHz inside Windows would result in performance numbers that are incorrect.

Here are my real numbers
View attachment 27418

I am no longer a special snowflake. *sniff*

Well, then the thread would be dead. We all know what Vega is going to be:








A disappointment

True. But the problem there is, there are no rumors. AMD has been so tight lipped. Good or Bad.

With no new rumors it just remains inactive until further ado.

https://www.fool.com/investing/2017/06/11/in-another-blow-to-nvidia-amd-sets-its-sights-on-t.aspx

Increase market shares and plan to take away from Nvidia the high end.

 

[razor1]

Don't see how that's possible but with that R&D budget you'd sort of hope so. 1.4-1.5x performance perhaps, with expected uarch + bigger die up from 480mm² to 550-600mm² 250W+ I can see that happening.

the uarch of pascal is very much unknown at this point. And so far nV has hit every single project in the past 8 generations of cards. -1 the 2xx series.

They're going up in die size from 480mm² to ~600mm² which will account for much of the performance gain too.

That won't account for efficiency though.

Just the increase in cuda cores should give nV 40% increase in performance, So add on top of that efficiency and throughput increases. 80% seems likely with the increase in cache and register amounts along with the increased frequency nV has already stated. I think the node will accommodate part of the drop in power draw as well there was mention of this from TSMC and nV but can't remember the exact figure.

 

[Maddness]

With no new rumors it just remains inactive until further ado.

https://www.fool.com/investing/2017/06/11/in-another-blow-to-nvidia-amd-sets-its-sights-on-t.aspx

Increase market shares and plan to take away from Nvidia the high end.

Haha, I hope that's true. But the name of the websited doesn't give me much belief in there credability. Time will tell I guess.

 

[noko]

Haha, I hope that's true. But the name of the websited doesn't give me much belief in there credability. Time will tell I guess.

Well it is established the workstation or pro version runs at 1600mhz - think of FuryX 1.6x+. Gaming cards, at least the top end should be faster then that plus we still don't know how far above that speed it can really go, how temperature sensitive it is (Ryzen on this same process is very temperature sensitive) meaning some headroom can be earned with cooler configurations etc. Looks like the odds of it triumphing or beating the Pascal lineup is there.

Plus Nvidia frantic pace at trying to push up poor Volta is not for ought. Maybe some fear has finally crept up into their ranks.

 

[Maddness]

I would truly love to see it happen. Even if Vega can compete against the 1070, 1080 and the 1080Ti. In all the price brackets, would be great. Beat them and it will be some kind a statement to the direction AMD is heading in. Especially considering how well RyZen is competing against Intels CPU's. It would be amazing to see them so competitive in both markets. Especially with the limited funds they have compared to there rivals.

 

[noko]

I would truly love to see it happen. Even if Vega can compete against the 1070, 1080 and the 1080Ti. In all the price brackets, would be great. Beat them and it will be some kind a statement to the direction AMD is heading in. Especially considering how well RyZen is competing against Intels CPU's. It would be amazing to see them so competitive in both markets. Especially with the limited funds they have compared to there rivals.

I am pretty sure AMD will beat Nvidia Perf/$ other wise who will care? Getting the performance crown decisively would be the big win if that happens. I see that as possible but if it is not by much then it probably will make a small splash.

 

[CSI_PC]

Plus Nvidia frantic pace at trying to push up poor Volta is not for ought. Maybe some fear has finally crept up into their ranks.

Nvidia timeline has always been around what we are seeing; quite a few of us have been arguing for awhile Volta would launch and be in production by summer 2017 for Tesla and quite plausible for a GV104 to be sometime Q4, this comes back to product cycle and project obligations.
Nvidia also accelerated Pascal launch and changed the strategy by launching on the largest possible die even back then, followed in August 2016 the Titan Pascal and there was no challenge to GP104 back then in May let along GP102.

Nvidia has a lot of synergy between the various segments, but what is primarily driving this accelerated cycle is HPC/DL/Tegra automobile also with other solutions.
This has a knock on effect on consumer GPUs as well by bringing them forward due to the synergy in the product cycle process across segments (including R&D).

Pascal was a technical risk milestone towards Volta as it introduces some of the technical aspects in stages instead of all at once with Volta; NVLink/HBM2/Unified Memory/etc.
Quite a few of those have now evolved more to what was expected for Volta (such as NVLink 2 not only expanded capability but also actual cache level coherency between all accelerators and CPU), however it would had been an incredible technical risk to introduce all of these technologies and solutions in one go, especially with other core technologies in Volta such as Tensor/simultaneous Int32+FP32/re-write of the processor architecture/dispatch/compiler/etc.

Nvidia's challenge is consolidating HPC/scientific/cloud momentum and especially the new market of DL in its various forms against various competitors such as Intel or those more bespoke HW solutions.
AMD has a chance to get a piece of the pie in this space but for now they are not the main concern I would say for Nvidia.
Cheers

 

[CSI_PC]

I don't blog, but I do mine, prior to the launch of LTC ASICS I was drawing around 255,000 watts (three phase, 800 amp service - I bought an empty grade school with bitcoin cash - I still have the school ).

Everyone you quote is an expert, far be it from me to get tangled up with experts. I encourage everyone to find a practical means of doing their own assessment.

If you're mining with 1-10 cards, who really cares though as these percentages have negligible impact. at that point it's more cost of acquisition and we see how that's going and how it's going to impact Vega.

Well I can only mention the performance efficiency figures they are getting with the Nvidia GPUs and they were better than what you suggested *shrug*; specifically 1070 and depending upon algo 1080ti.

These guys understand the Linux (when needed)/CUDA version importance/settings and what needs to be disabled to get the most out of Nvidia, so yeah they are experts in that way; the ones I reference have managed similar performance efficiency beyond that of the 480 and 580 and are reaching similar figures on those to what most report at the upper end.
And tbh if scientific lab oscilloscopes point out how inefficient 480/580 is (I included actual chart showing this in earlier post) relative to 1060 and indirectly 1070 when it comes to clocks/watts/performance then it is pretty hard to argue against such data.
Like I said you cannot break the clocks/compute relationship, but that was in the "wall of text".
Problem for many doing their own assessment is it requires knowing how to set up the various HW to its optimum, and Nvidia HW is not that user friendly in the context of CUDA mining and disabling GUI interfaces, and that is compounded with accurately measuring performance/efficiency of GPU and indirectly CPU if really wanting to debate product X vs product Y.
Cheers

 

[Simplyfun]

Well I can only mention the performance efficiency figures they are getting with the Nvidia GPUs and they were better than what you suggested *shrug*; specifically 1070 and depending upon algo 1080ti.

These guys understand the Linux (when needed)/CUDA version importance/settings and what needs to be disabled to get the most out of Nvidia, so yeah they are experts in that way; the ones I reference have managed similar performance efficiency beyond that of the 480 and 580 and are reaching similar figures on those to what most report at the upper end.
And tbh if scientific lab oscilloscopes point out how inefficient 480/580 is (I included actual chart showing this in earlier post) relative to 1060 and indirectly 1070 when it comes to clocks/watts/performance then it is pretty hard to argue against such data.
Like I said you cannot break the clocks/compute relationship, but that was in the "wall of text".
Problem for many doing their own assessment is it requires knowing how to set up the various HW to its optimum, and Nvidia HW is not that user friendly in the context of CUDA mining and disabling GUI interfaces.
Cheers

They do indeed point out how inefficient Polaris is - the way they have them set up. Perhaps they need to get a little better at it, maybe if it really mattered and they were building income based on their ability to do so, they might work a little harder at it. Keep in mind, I was within 4 watts at the same performance level as your expert with a 1070. 4 watts - That's merely luck of the card lottery. Where you keep coming back to me is that the expert couldn't hit my numbers on Polaris . After all AMD HW is not that user friendly in the context of OpenCL mining and disabling GUI interfaces.

I know every watt that comes off my sub panels and where it's going.

Anyhow, I think we're not doing this thread any favors. Let's move on.

 

[noko]

Nvidia timeline has always been around what we are seeing; quite a few of us have been arguing for awhile Volta would launch and be in production by summer 2017 for Tesla and quite plausible for a GV104 to be sometime Q4, this comes back to product cycle and project obligations.
Nvidia also accelerated Pascal launch and changed the strategy by launching on the largest possible die even back then, followed in August 2016 the Titan Pascal and there was no challenge to GP104 back then in May let along GP102.

Nvidia has a lot of synergy between the various segments, but what is primarily driving this accelerated cycle is HPC/DL/Tegra automobile also with other solutions.
This has a knock on effect on consumer GPUs as well by bringing them forward due to the synergy in the product cycle process across segments (including R&D).

Pascal was a technical risk milestone towards Volta as it introduces some of the technical aspects in stages instead of all at once with Volta; NVLink/HBM2/Unified Memory/etc.
Quite a few of those have now evolved more to what was expected for Volta (such as NVLink 2 not only expanded capability but also actual cache level coherency between all accelerators and CPU), however it would had been an incredible technical risk to introduce all of these technologies and solutions in one go, especially with other core technologies in Volta such as Tensor/simultaneous Int32+FP32/re-write of the processor architecture/dispatch/compiler/etc.

Nvidia's challenge is consolidating HPC/scientific/cloud momentum and especially the new market of DL in its various forms against various competitors such as Intel or those more bespoke HW solutions.
AMD has a chance to get a piece of the pie in this space but for now they are not the main concern I would say for Nvidia.
Cheers

Nvidia has been reacting to Vega, cost cuts, 1080Ti and new Titan Xp upgrade . AMD has them running it seems. Plus the fancy Borg Cube Raja was holding in his hands (maybe a bluff) that put a highly dense Vega arsenal aimed at Nvidia deep learning progression. It is a matter of AMD delivering the goods and now it really looks like it will be a very strong delivery when ready. The pro card is at 1600mhz - what are the gaming cards going to run at? 1700? 1800? 1900? As those numbers get higher the further Nvidia will be behind. Is Vega really aiming at Volta? To much is not known but what is is actually impressive.

 

[Ieldra]

Nvidia has been reacting to Vega, cost cuts, 1080Ti and new Titan Xp upgrade . AMD has them running it seems. Plus the fancy Borg Cube Raja was holding in his hands (maybe a bluff) that put a highly dense Vega arsenal aimed at Nvidia deep learning progression. It is a matter of AMD delivering the goods and now it really looks like it will be a very strong delivery when ready. The pro card is at 1600mhz - what are the gaming cards going to run at? 1700? 1800? 1900? As those numbers get higher the further Nvidia will be behind. Is Vega really aiming at Volta? To much is not known but what is is actually impressive.

View attachment 27504

Never ceases to amaze me how an absence of information (and physical market presence) on behalf of amd with regards to Vega can be twisted into an indication that they have some kind of Ace up their sleeve.

Koduri stated that the consumer cards would be gaming optimized (drivers) and cheaper.

Aassuming stock clocks of 1525( 12.5 tflops) then a 15% OC would just bring it in line with 1080Ti throughpu - and that's already 1.77GHz

How you think this is even relevant to volta, which is virtually guaranteed to replicate the usual generational 50/60% leap, is beyond me

We are looking at the 20tflop mark for GV102.

Edit:

I'm not sure we agree on the definition of the word reaction, a reaction is a response to an external stimulus. How did nvidia react to Vega 6 months before we are even likely to see the first Vega cards in the wild?

Edit:

Last generation a Fury X was close to 40% ahead of a 980Ti on paper. You needed around a 30% OC to match the fury X and its 8.6tflops.

This generation Vega needs 15% OC to match 1080Ti on paper


Thats quite a major role reversal, AMD have a hill to climb

 

[noko]

Never ceases to amaze me how an absence of information (and physical market presence) on behalf of amd with regards to Vega can be twisted into an indication that they have some kind of Ace up their sleeve.

Koduri stated that the consumer cards would be gaming optimized (drivers) and cheaper.

Aassuming stock clocks of 1525( 12.5 tflops) then a 15% OC would just bring it in line with 1080Ti throughpu - and that's already 1.77GHz

How you think this is even relevant to volta, which is virtually guaranteed to replicate the usual generational 50/60% leap, is beyond me

We are looking at the 20tflop mark for GV102.

Edit:

I'm not sure we agree on the definition of the word reaction, a reaction is a response to an external stimulus. How did nvidia react to Vega 6 months before we are even likely to see the first Vega cards in the wild?

Edit:

Last generation a Fury X was close to 40% ahead of a 980Ti on paper. You needed around a 30% OC to match the fury X and its 8.6tflops.

This generation Vega needs 15% OC to match 1080Ti on paper


Thats quite a major role reversal, AMD have a hill to climb

Why would stock clocks be at 1525mhz? Where did you pull that up? The Frontier Edition is at 1600mhz, gaming cards are normally way faster then the Pro cards. How much faster is the question? Of course we have no clue on OC ability or if cooling can make a big difference. Once Frontier Edition is available we maybe able to extract some good data to have a better prediction on the gaming card performance. I suspect it will surpass the 1080Ti, just not sure by how much.

Vega looks to be sufficiently different over Fiji that comparing floating point performance may not paint an accurate picture at all. Still if Vega Extreme is 1750mhz it will be 1.66 x the clock speed of Fiji plus design increases - this could be a very very fast card.

 

[razor1]

That not the way Raja put his statements, Raja specifically stated compute wise the frontier edition will be highest compute performance board.

 

[Ieldra]

Why would stock clocks be at 1525mhz?

Well it's a 12.5 tflop card with 4096 shaders, that's pretty self explanatory

Where did you pull that up?

The Frontier Edition is at 1600mhz, gaming cards are normally way faster then the Pro cards.

1600MHz is the maximum boost clock detected in the compubench( I think thats what its called) leaks. Where are you pulling this number from? Cause it's not on any official AMD slides.

Gaming cards are normally clocked way faster than the pro cards? Where did you pull that up? A quick glance at the specs for the FirePro cards shows that the Polaris based SKUs have the same boost as consumer, same with Hawaii , and same with Fiji.

How much faster is the question?

Based on the specs, none at all.

Of course we have no clue on OC ability or if cooling can make a big difference.

It's a 300W card at 12.5tflops, with a CLC. You tell me.

I suspect it will surpass the 1080Ti, just not sure by how much.

By what reasoning ?

Vega looks to be sufficiently different over Fiji that comparing floating point performance may not paint an accurate picture at all.

Yet here Vega 10 is shown to be 40% faster than Fiji while it is 45% faster in terms of theoretical tflop/s

Still if Vega Extreme is 1750mhz it will be 1.66 x the clock speed of Fiji plus design increases - this could be a very very fast card.

If it reaches those clockspeeds it will match a 1080Ti on paper, as I said earlier.

 

[Ieldra]

That not the way Raja put his statements, Raja specifically stated compute wise the frontier edition will be highest compute performance board.

I think between the watercooler and the 300W rating he didn't even need to specify this would be the case