• Competitor rules

    Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.

AMD VEGA confirmed for 2017 H1

Status
Not open for further replies.
This is what I have been saying for a long time. The top end Vega needs to be at least matching the Titan PX when it comes out or it will be a failure IMO.

Can anyone explain how new architecture + HBM2 + 14nm + extra time to develop = 1070 performance? Just does not compute.

It was not that long ago that top end AMD (Fury X) and NVIDIA (Titan X) were like not hugely apart in performance. How did AMD suddenly go from that to it will be surprising if they are able to beat a 1070/80?

AMD have had an extra year to develop Vega so it needs to be beating Titan XP and bringing that tier of performance to Fury X price range at the least.

It is because they do not necessarily have to develop the largest possible die for any architecture generation. That Vega is coming a year later doesn't mean much if they simply decided not to aim for a performance target that necessitates a very big die (relatively speaking).

Not that I think they will aim as mediocrely as they did with Polaris. They clearly abandoned the development of a larger Polaris die, perhaps quite early before skipping 20nm (power?) or after (economic viability bumping up against Vega roadmap?). But it's worth noting that just because they have had an extra year (in terms of Vega vs what will be older/outgoing Pascal), it does not mean it must therefore beat Titan XP.
 
It is because they do not necessarily have to develop the largest possible die for any architecture generation. That Vega is coming a year later doesn't mean much if they simply decided not to aim for a performance target that necessitates a very big die (relatively speaking).

Not that I think they will aim as mediocrely as they did with Polaris. They clearly abandoned the development of a larger Polaris die, perhaps quite early before skipping 20nm (power?) or after (economic viability bumping up against Vega roadmap?). But it's worth noting that just because they have had an extra year (in terms of Vega vs what will be older/outgoing Pascal), it does not mean it must therefore beat Titan XP.

I think the smaller Vega die is probably going to be similar in shader count to the the GPU in the XBox Scorpio GPU,just like the shader count of Polaris 10 and the PS4 PRO GPU are the same,and the noise is it has 4096 shaders like Fiji does. Now add two generations of uarch improvements,two generations of memory compression tech,more VRAM and hopefully more ROPs and a higher clockspeed,you should be looking at a sub 400MM2 die,with probably great than GTX1080 performance. I kind of expect GTX1080 beating performance,but not as fast as GTX1080TI for the smaller Vega chip.
 
Last edited:
HBM2 in its current state on the Tesla cards will throttle cards capable of very high fps like the Pascal Titan.

Latency is more important than huge bandwidth when dealing with very high fps.

Bandwidth is more important than latency when dealing with extremely high fps.
 
Last edited:
It is not lower than GDDR5X

Comparing HBM2 to GDDR5X is like a lorry to a sports car.

You will get a lot more work done with the lorry but the sports car will get you from A to B quicker.

With respect do you even know what you are referring to when you talk about Latency?

Yep,the physical traces are more closer to the GPU unlike GDDR5X. Fiji was bottlenecked by not having enough ROPs and probably CPU software bottlenecks under DX11. Under DX12 and Vulkan it does far better.

In fact if you look at Polaris 10,the improved command processor and primitive discard accelerator has improved performance under DX11 a decent amount. This is why in newer games a RX480 with significantly lower bandwidth is pushing ahead of even a R9 390 or R9 390X.

He does, he also understands the issues with Fiji.
 
Not worth upgrading from a 290X to a Fury (air cooled model with disabled shaders). It would be 10-15% faster at best, hardly worth the expense IMO.

I opted for a 1070 to tide me over until Vega launches.

I went from a 290x to a Fury pro but that was back when the Fury was relatively new, I wouldn't bother doing it now considering what else is around and due.
You all have low expectations.
If big Vega does exceed the 1080 by a huge amount it'll be a failure too. By the time Vega comes out the 1080 would have been on the market for close to a year. R&D gets cheaper to hit a performance level with time so has to exceed the 1080. BY then Nvidia might have released more cards or be very close to their next gen.

I'm hoping AMD are using this time to do a lot of R&D and their Vega ends up being close to Nvidia's next gen, not current gen, so that means blowing away the current TX too. If AMD can't do this then it'll be another rather poor showing IMO , even taking their lower R&D budget into consideration.
I'm just being a realist.
 
Bandwidth is more important than latency when dealing with extremely high fps.

With respect do you even know what you are referring to when you talk about Latency?



He does, he also understands the issues with Fiji.

You do realise that both the Tesla GP100 (with HBM2) and the Pascal Titan (with GDDR5X) have almost identical FP32 performance.

The Pascal Titan has 11 TFLOPS compared to the 10.6 TFLOPS of the Tesla GP100. The slight difference is due to the Titan boost clockspeed being slightly higher 1480mhz v 1531mhz.

Or putting it another way HBM2 has done absolutely nothing for FP32 performance which is something that can be measured on both cards.

The best part about the Pascal Titan is if you water cool it you can then run it at close to 2100mhz for a big performance increase.
 
You do realise that both the Tesla GP100 (with HBM2) and the Pascal Titan (with GDDR5X) have almost identical FP32 performance.

The Pascal Titan has 11 TFLOPS compared to the 10.6 TFLOPS of the Tesla GP100. The slight difference is due to the Titan boost clockspeed being slightly higher 1480mhz v 1531mhz.

Or putting it another way HBM2 has done absolutely nothing for FP32 performance which is something that can be measured on both cards.

The best part about the Pascal Titan is if you water cool it you can then run it at close to 2100mhz for a big performance increase.

Because HBM has the benefit of smaller footprint on the PCB, Lower power usage and increased bandwidth. If you increase the bandwidth over the PCI_E lanes does that magically increase performance? Probably not for one card but
start using two card then you will saturate the bandwidth. Ideally its the same for higher resolutions. High Bandwith memory.

Thing is with GDDR5x is that it also has increased bandwidth over GDDR5 however it has better clockspeeds too vs GDDR5.
Things work slightly different with HBM however.
 
Because HBM has the benefit of smaller footprint on the PCB, Lower power usage and increased bandwidth. If you increase the bandwidth over the PCI_E lanes does that magically increase performance? Probably not for one card but
start using two card then you will saturate the bandwidth. Ideally its the same for higher resolutions. High Bandwith memory.

Thing is with GDDR5x is that it also has increased bandwidth over GDDR5 however it has better clockspeeds too vs GDDR5.
Things work slightly different with HBM however.

So you agree HBM2 is doing nothing for performance on these single cards I mentioned about.
 
It is because they do not necessarily have to develop the largest possible die for any architecture generation. That Vega is coming a year later doesn't mean much if they simply decided not to aim for a performance target that necessitates a very big die (relatively speaking).

Not that I think they will aim as mediocrely as they did with Polaris. They clearly abandoned the development of a larger Polaris die, perhaps quite early before skipping 20nm (power?) or after (economic viability bumping up against Vega roadmap?). But it's worth noting that just because they have had an extra year (in terms of Vega vs what will be older/outgoing Pascal), it does not mean it must therefore beat Titan XP.

Sure they may not have necessarily gone for the largest die, but it is very likely big Vega's die will be at least as big as Titan PX. It may not mean having an extra year they will definitely beat Titan PX but my guess is they can and will. Others are being cautious purely in my opinion because they do not want to get their hopes up again and be burned, also you have others who cannot see past their green tinted shades :p:D
 
This is what I have been saying for a long time. The top end Vega needs to be at least matching the Titan PX when it comes out or it will be a failure IMO.

Can anyone explain how new architecture + HBM2 + 14nm + extra time to develop = 1070 performance? Just does not compute.

It was not that long ago that top end AMD (Fury X) and NVIDIA (Titan X) were like not hugely apart in performance. How did AMD suddenly go from that to it will be surprising if they are able to beat a 1070/80?

AMD have had an extra year to develop Vega so it needs to be beating Titan XP and bringing that tier of performance to Fury X price range at the least.

There won't be surprise from me if it just about beats a custom 1080, I think if the card can match a custom 1080 card for £500 or less it will please many people and many cards will end up in new Zen rigs.

I think AMD's target is to get Broadwell-E 8 Core/1080 performance to enthusiasts at significantly lower cost than what Intel/Nvidia are asking for right now.

I think this is realistic, I think the above posts about failure are way wide of the mark and are missing the pent up demand to refresh Sandy Bridge rigs at a decent price.

Product does not need HBM for the wow factor, if I don't need to fork out £500 for an Intel motherboard and £1000 for a CPU in my Q1 refresh, that will be wow enough for me thank you.

4k is not the requirement for most users in 2017, its fast 1440p/widescreen performance and VR and I am confident that AMD will more than meet this, to the extent that I would trade out of Nvidia stock and in AMD stock.
 
Last edited:
So you agree HBM2 is doing nothing for performance on these single cards I mentioned about.

Kaap, from the point of view of an unbiased observer, you are *obsessed* with slating HBM. I have no idea why you've decided to go on a one-man crusade against this technology, but anybody who follows these threads will basically be ignoring anything you have to say about it by this point.

Let it go.
 
There won't be surprise from me if it just about beats a custom 1080, I think if the card can match a custom 1080 card for £500 or less it will please many people

Won't impress me one bit, quite the opposite actually. If it only manages to just about beat out a 1080 by summer, it needs to be like £300-£400 not £500.
 
Won't impress me one bit, quite the opposite actually. If it only manages to just about beat out a 1080 by summer, it needs to be like £300-£400 not £500.

Agree with that too. People seem to be forgetting the time frames involved here and the development lifecylces. 1080 performance at £500 by summer 2017 from AMD will be utter tripe to be honest, a whole year after we've had the same performance for £619. I expect much more better than that else Nvidia will be wholly justified in releasing an 1180 for +30% more £ if the best the competition can do is just match a 1080 a whole year later than its release for £500.
 
Last edited:
So you agree HBM2 is doing nothing for performance on these single cards I mentioned about.

Well not looked into it in great detail as in Pascal Titanx and the Quadro with HBM. But its certainly not worse in performance using HBM. If its the same FP32 then its the same performance.
 
Well not looked into it in great detail as in Pascal Titanx and the Quadro with HBM. But its certainly not worse in performance using HBM. If its the same FP32 then its the same performance.

There is a lot about what makes up performance in a video game that on paper FP32 numbers can't tell you mind.

Generally though we seem to be a bit of a way off yet from GPUs where HBM type memory really brings significant advantages.
 
I don't think that AMD is going to have nothing between a £250 RX480 and a £500-600 (or greater) RX490. That leaves a massive gap in the performance market for Nvidia to cut a few prices and drop into. AMD has to have a product in the £300-400 market.
 
so the expectation now is that Vega will beat a 1080 and cost up to £400?

if that happens, performance wise, I don't see this card being cheaper than 500, probably more. Fury x was above 500 on release wasn't it? With the pound taking a beating, chances are prices will be quite a bit higher possibly edging towards £600.
 
to be honest I am actually tempted to give Vega a go, provided it is as good as some people think it will be. Now if I only I could get access to MM to have a loot around and sell my 1080 ... :)
 
Status
Not open for further replies.
Back
Top Bottom