Nvidia Pascal Reviews [1080XP, 1080ti, 1080, 1070ti, 1070, 1060, 1050, and 1030]

So did nvidia just completely cannibalized all the Quadro P5000 and most of P6000 sales, or is there a catch here?
The catch is the same as with Vega FE: acceleration is enabled but the driver is not certified and has no technical support for these apps.
 
That makes sense, but then this post is wrong.
I don't think it's decided yet whether the drivers will be certified (most likely not), but could be a selling point differentiating the TitanXp from it's certified siblings. One way or the other it's a fantastic opportunity.

Edit: The fact that the drivers are WHQL even though they carry the Beta name tag may indicate certification may not be forthcoming.
 
Screen_Shot_2017-08-01_at_7.03.03_AM.png

Why is Maya decreasing in performance by ~10%? Isn't that the one they said would increase?

You're being quite funny today :) Good thing for end-users that AMD created this "non-crippled-but-not-certified" niche. :D

AMD Created prosumer niche? Hasn't that been what the Titan line has always been about? Nvidia markets it that way.
 
Not sure I'd call it legitimate, but a lot of people seemed to be under the mistaken impression that adding more than 10% performance with drivers was impossible.
Wasn't that particular issue focused on whether the performance demonstrated by Vega FE would only be improved by so much for Vega RX?

If the Vega RX driver can be credited with the Titan performance gains, AMD should talk to their driver group.
 
Wasn't that particular issue focused on whether the performance demonstrated by Vega FE would only be improved by so much for Vega RX?
That was one aspect of it, but the other was magical drivers built from the horns of unicorns to make Vega competitive against Nvidia with comparable hardware. That even with improved drivers Vega was unlikely to gain much more than 10% even after a year as that's never happened before. Simply because drivers can't have that much influence. Completely ignoring the part where it occurs somewhat regularly. I'm not expecting 654% across the board, but enough to get near Titan and possibly exceed with future features. More on par with the pro benchmarks and larger than usual gains over time.
 
That was one aspect of it, but the other was magical drivers built from the horns of unicorns to make Vega competitive against Nvidia with comparable hardware. That even with improved drivers Vega was unlikely to gain much more than 10% even after a year as that's never happened before. Simply because drivers can't have that much influence. Completely ignoring the part where it occurs somewhat regularly. I'm not expecting 654% across the board, but enough to get near Titan and possibly exceed with future features. More on par with the pro benchmarks and larger than usual gains over time.
While probable that Vega should improve over time, the Titan situation was the result of purposeful segmentation.
There was little speculative about the idea that one SKU using GP102 being able to perform similarly to another, if permitted.

It's almost a non-sequitor to use a decision to reduce segmentation as a way to bolster a far more speculative position. It's also not a strong refutation of the skeptics' position, or the rather stretched version provided.
 
It's almost a non-sequitor to use a decision to reduce segmentation as a way to bolster a far more speculative position. It's also not a strong refutation of the skeptics' position, or the rather stretched version provided.
Not if the argument is that drivers can't make a significant difference to performance. Obviously they did in this case, even though deliberately crippled. I'm not suggesting AMD will see these kind of gains, but that drivers can in fact have a huge influence on performance and efficiency. Big moves are seen with linux drivers all the time where driver development has historically been behind Windows. Benchmarks are showing that Vega FE hasn't been performing very efficiently. It's not 600% but a mere 40% that is at issue here. We already know packed math could do 25-30% of that. Add in a rather pessimistic 10% in game optimizations over the course of a year and they're roughly equal. It's not a stretch to say that currently known improvements will do just what I've suggested.
 
Not if the argument is that drivers can't make a significant difference to performance. Obviously they did in this case, even though deliberately crippled.
I consider the deliberate choice by the vendor to be more critical than the specific method for hobbling the same silicon. The driver was the most convenient method to do so, but there are other methods that could be used and have been used in the past for graphics hardware or professional products in general.

One claim is that best-effort attempts to improve a product given the same silicon is unlikely to produce major improvements in limited time given that the product performed consistently.
A vendor's very purposeful non-improvement or throttling of its product when it is plainly clear the hardware is fully performant when it is convenient highlights how important the vendor's choice is.

Benchmarks are showing that Vega FE hasn't been performing very efficiently. It's not 600% but a mere 40% that is at issue here. We already know packed math could do 25-30% of that. Add in a rather pessimistic 10% in game optimizations over the course of a year and they're roughly equal. It's not a stretch to say that currently known improvements will do just what I've suggested.
To note, the largest example given for the "drivers have a huge influence" position is something requiring developer intervention, and my understanding of the origins of the 10% claim is that it didn't include a time frame of one year.
 
Last edited:
We already know packed math could do 25-30% of that.
Adoption rate is important, and it doesn't come from paid partnerships, it comes from widespread hardware support. Who is going to push for RPM when it's available on a very limited slice of the market (Vega GPUs)? Vega will probably sell like FuryX, (ie, in very limited quantities. So no, don't expect widespread support. Just like DX12 and Async (which supported a wider range of hardware), developer adoption remains very limited 2 whole years after DX12's introduction.

More on par with the pro benchmarks and larger than usual gains over time
Pro benchmarks place Vega FE slightly behind pro GP104 in all but very CPU limited scenarios. So now you have both gaming and pro at the level of 1080.
8c4adf8d42d1fb43e74207ebb37441b1b3d22f15b80bbd999267a417f2718aff.png
 
One claim is that best-effort attempts to improve a product given the same silicon is unlikely to produce major improvements in limited time given that the product performed consistently.
A vendor's very purposeful non-improvement or throttling of its product when it is plainly clear the hardware is fully performant when it is convenient highlights how important the vendor's choice is.
In most cases there aren't significant features left non-functional. There are estimates with DSBR so we have some idea as to the benefit, but the feature is disabled on most titles currently. That's not far off of deliberately hobbled. Until RX officially launches there is no way to know for certain. I wouldn't call that far off from deliberately crippling a product, even if reasons exist.

To note, the largest example given for the "drivers have a huge influence" position is something requiring developer intervention, and my understanding of the origins of the 10% claim is that it didn't include a time frame of one year.
Few sides to this: 1) New architecture so likely more potential for improvement. 2) While more API than drivers, packed math and intriniscs aren't readily exposed. That is likely to change in the near future.

The one year remark is just a rough estimate from typical gains for reference. The goal was enough changes to get near 40% uplift. Situationally DSBR could do that given provided figures. Improvements to cache usage could provide significant gains. Compiler optimizations using new hardware or cache. Point being there would seem many avenues to get that 40% without necessarily using packed math. Packed could simply be a crutch in the short term as drivers tune performance to get equivalent performance.

Between DSBR and packed I think AMD will manage to get near Titan, then surpass it over the course of a year.
 
Back
Top