SemiAccurate is probably on AMD's whitelist ever since their exposure of the woodscrews Fermi saga all those years ago: http://semiaccurate.com/2009/10/01/nvidia-fakes-fermi-boards-gtc/
Follow along with the video below to see how to install our site as a web app on your home screen.
Note: This feature may not be available in some browsers.
Agree, Vega have a very short window to grab the huge customers, not saying there isn't a market for Vega but large data centers probably don't care about hardware cost so thew probably will opt for the what is best out there.
Eh, GV 100 is a massive chip, price wise there's probably plenty of room for Vega. If the performance difference is only 20-30 percent or so, there's plenty of room for them to compete on price. They're beating the P100 by a pretty huge margin in DeepBench.
This first version of the benchmark will focus on training performance in 32-bit floating-point arithmetic. Future versions may expand to focus on inference workloads as well as lower precision arithmetic.
Seriously, when Computex rolls around then you can make this argument, but this was an investors meeting so NO you will not see the comparisons you keep pointing out. They showed their new products in a positive manner without overstepping. We all know to wait for reviews as most every company never tells you everything in every manner that can be told.Because DeepBench is only FP32 and P100 is designed really for FP64 and FP16 around good FP32 performance in general (albeit surpassed by GV102 with regards to FP32).
If AMD wants to do FP32 benchmark, they should use the P40 that has just over 12TFLOPs rather than the P100 with 9.3TFLOPs as PCIe model.
Here is the info relating to DeepBench:
I am wondering why AMD never showed FP16 actual results with DL from AlexNet or ResNet.
The V100; well that has raw performance of 42% increase over P100 in the traditional sense of general cores increased, but then it has additional HW that goes beyond this and one aspect is still to be seen how it pans out in general (parallel Int32 and FP32 computation) along with 8 Tensor cores per SM.
And those 8 Tensor cores per SM is 4x greater peak throughput than P100 with FP16 instruction per SM doing DL, importantly that is per SM and remember the V100 has 80SM against 60SM with the P100, so an additional 33% increase on top of that when using all Tensor cores vs P100 cores.
In reality that is theoretical peak throughput which both AMD and Nvidia base their figures with FMA.
But even now before truly optimised real world benchmarks show 2x to 3.5x improvements over comparable operation from P100 for DL or matrix related math, and this is still not using all the available optimisation changes brought with Volta over Pascal.
If AMD could had launched Vega 6-months ago it would had possibly been truly disruptive, it is still impressive but the launch cycle now just does not help IMO for some of the segments.
Cheers
So why did AMD even bother showing the benchmark then using a comparison with the wrong and weaker Nvidia card in the Tesla lineup when it comes to FP32?Seriously, when Computex rolls around then you can make this argument, but this was an investors meeting so NO you will not see the comparisons you keep pointing out. They showed their new products in a positive manner without overstepping. We all know to wait for reviews as most every company never tells you everything in every manner that can be told.
SemiAccurate is probably on AMD's whitelist ever since their exposure of the woodscrews Fermi saga all those years ago: http://semiaccurate.com/2009/10/01/nvidia-fakes-fermi-boards-gtc/
That is true, the product shown is for the Pro market though I still believe the window for gaining traction in HPC market is small since the impending release of Volta. I am not sure how effective it will compete in the Pro scene due to existing Pascal Quadro line.The headache for AMD though is that Frontier is not large data centers or even actual HPC nodes/clusters.
Because those would use the passive Mi25 in chassis like Nvidia does with the DGX-1.
I think for that market this card is to get buy-in and acceptance towards convincing clients of Mi25 and scale-up actual HPC/scientific nodes.
But how far away is the Mi25 relative to the V100/GV102.
Cheers
So the one Raja was holding was the one with the 1 6in pin and 1 8 pin, anywhere from 300 watts to 375 watts? Still up in the air. But its matching up with the instinct cards.
My 1080 ti has 6 and 8 pin. Does it use 300 watt or 375? Hell no. To say how much power a card uses looking at power plug is not really fair. 8 pin will only give you 225w so not really any other option to stick another 6 pin or 8 pin to go beyond that.
However, it won't be the first time AMD marketing goof up. Seeing 2 8 pin does make me wonder how much juice it is really using?Highly doubt they can match Pascal's perf/watt, so I'm thinking its closer to the 300 watts TDP limit, that rendition with the 2 8 pin, don't think its by accident. Art departments get their "ideas" from what the engineer specs sheets have .
However, it won't be the first time AMD marketing goof up. Seeing 2 8 pin does make me wonder how much juice it is really using?
Lol you are assuming if AMD Marketing Department were run by competent folks, we know that is not true.Goof ups like that they would tell them to re render with a 6 pin lol, That is a easy fix. with Ryzen it should be done in 30 secs right lol.
Good stuff, AMD will cube gpus for density or more gpu/volume ratios for more compute per volume. As for single card looks like vega x2 will compete with V100. Still an uphill battle for AMD.Nice, that's pretty insane, we'll have to see how Vega lines up to it when it releases at the end of next month. I'm curious how much this damn thing is going to cost.
I want three 8 pin sockets if it OC wellMy 1080 ti has 6 and 8 pin. Does it use 300 watt or 375? Hell no. To say how much power a card uses looking at power plug is not really fair. 8 pin will only give you 225w so not really any other option to stick another 6 pin or 8 pin to go beyond that.
I love the very simple flat design of the card as in cut to the chase and let's get to business. I hope the gaming cards are similar in design.
I like more the performance that goshes out of a card. Card styling for me is ridiculous especially when it adds an extra $100 for the designers choice that clocks slower than the lower cost ones.I don't. It is a small but pet hate if mine when it comes to AMD cards. I think most of them look like an ugly brick. Each to his own i suppose.
I don't. It is a small but pet hate if mine when it comes to AMD cards. I think most of them look like an ugly brick. Each to his own i suppose.
Maybe the difference between the Rx versions and the Instinct ones will be Red vs. blue .personally I like the look, just eye of the beholder I guess. Love to the lighted cube too, goes back to the Vega cube they showed last year, prefect marketing.
Raja said the game demos were running on the frontier edition GPU. He also said binning rasterizer or culling I think requires no input from developer side and new geometry pipeline.
Also they are optimizing for gaming. I think it's just drivers. Frontier edition pro drivers currently may not have the binning rasterizer feature enabled. Also mentioned RX vega flavors will be faster than Frontier edition.
Is the geometry performance overkill with Nvidia? More than needed? How much is enough for current and future games? Just having X times more does not always = better overall or more importantly makes a difference for what it will be used for.He never stated anything about the binning rastorizer,
This is the full q and a.
Culling is better it seems to be akin to Polaris's for the most part, but they aren't going to get the advantages to really remedy the problems Polaris has when compared to Pascal or Maxwell or any nV hardware since Fermi, unless they use RPM for geometry. Polaris's geometry through put is better than previous GCN's, about twice the performance, but that only matches Maxwell (gtx 960). The reason why nV has this advantage is because their geometry units scale with their architecture, GCN doesn't do that, Polaris has 4, and Vega has 4. Now with GCN they have major front end issues that really hurt geometry through put even more, Polaris remedied that, I don't think Vega is going to be that much of step away from Polaris in this regard.
He never stated anything about the binning rastorizer,
This is the full q and a.
Culling is better it seems to be akin to Polaris's for the most part, but they aren't going to get the advantages to really remedy the problems Polaris has when compared to Pascal or Maxwell or any nV hardware since Fermi, unless they use RPM for geometry. Polaris's geometry through put is better than previous GCN's, about twice the performance, but that only matches Maxwell (gtx 960). The reason why nV has this advantage is because their geometry units scale with their architecture, GCN doesn't do that, Polaris has 4, and Vega has 4. Now with GCN they have major front end issues that really hurt geometry through put even more, Polaris remedied that, I don't think Vega is going to be that much of step away from Polaris in this regard.
Is the geometry performance overkill with Nvidia? More than needed? How much is enough for current and future games? Just having X times more does not always = better overall or more importantly makes a difference for what it will be used for.
For example Fuji and 980Ti at 4K. People I've seen has a tendency to equate more with better which can be true or not. Like having 600fps, 400fps or 200fps at 1080p has no bearing on game experience even though you can see a huge difference in frame rates. So far the geometry superiority of my 1070's made zero difference in my game experience, the overall performance otherwise has.
That was just a guess about frontier may not be fully optmized and raja quoting flavors of vega will be faster then frontier edition. That is why I was saying may be that is not being used in frontier edition.
This was my reasoning behind the other statement i made about primitive shader culling.
ooy787: I’m currently studying game programming at college (going to be a senior next year wooo) and I was wondering if you could discuss in what ways the new geometry pipeline will affect us programmers? What makes it more efficient than before?
RK: The new geometry pipeline in Vega was designed for higher throughput per clock cycle, through a combination of better load balancing between the engines and new primitive shaders for faster culling. As a programmer you shouldn’t need to do anything special to take advantage of these improvements, but you’re most likely to see the effects when rendering geometrically complex scenes that can really push the capabilities of the hardware.
There is no way of not using it, its part of the asic and part of the graphics pipeline, there really is no way to circumvent that part of the pipeline because of the way the API is set up. So if you are using DX or OGl, its using the primitive culling.
dude I am saying pro drivers not optimized for it. Common now! I never said like they cut it out or the card doesn't have it. Also everyone here kept saying that develop input will be required for primitive shaders but that doesn't seem to be the case.
its weird Kaduri flat out said there will be flavors of RX Vega faster then Frontier Edition. Also confirmed 2 stacks of HBM2 plus mentioned samsung and hynix as providers ramping up production to meet demand. I would not be surprised at all if AMD decided to get some memory from Samsung as well.
They are not that far apart in reality.The Quadro will not have a chance.