AMD VEGA confirmed for 2017 H1

Dave2150 · 21 Nov 2016 at 19:01

Bonjour said:
I'm thinking £600+ for the top-end tbh.

If Vega does indeed come with 16GB HBM2, it will be more like £800.

I'd imagine it will also come in a GDDR5X version, otherwise AMD would have a massive gap in their product stack between the RX480 8GB (£250) and the top end Vega with 16GB HBM2 (£800+ IMO).

Doogles · 21 Nov 2016 at 19:29

I never look at AMD now as they have only ever really done failures in recent years, R9 290x was excellent but that had to be fixed as they ballsed up with the cooling. With AMD everyone seems to expect something good from them me included so from now on I'll just take anything that comes good from them when it actually arrives, buying freesync and waiting for Vega taught me that!

Griffildur · 21 Nov 2016 at 19:30

Dave2150 said:
If Vega does indeed come with 16GB HBM2, it will be more like £800.

I'd imagine it will also come in a GDDR5X version, otherwise AMD would have a massive gap in their product stack between the RX480 8GB (£250) and the top end Vega with 16GB HBM2 (£800+ IMO).

I would expect to see multiple cards covering a larger price range. A card at 800+ better be amazing to have any chance of stealing sales away from Nvidia at that level. What I want to see if what happens on the power consumption and temperature side.

Dave2150 · 21 Nov 2016 at 19:32

Griffildur said:
I would expect to see multiple cards covering a larger price range. A card at 800+ better be amazing to have any chance of stealing sales away from Nvidia at that level. What I want to see if what happens on the power consumption and temperature side.

Is electricity expensive where you live? I expect the top end Vega to have similar TDP to the Titan XP (250W stock, 300+ overclocked).

Griffildur · 21 Nov 2016 at 19:34

Dave2150 said:
Is electricity expensive where you live? I expect the top end Vega to have similar TDP to the Titan XP (250W stock, 300+ overclocked).

It's not about that. I hope to see some decent cards not requiring a small nuclear power plant to function, that's all

Dave2150 · 21 Nov 2016 at 19:43

Griffildur said:
It's not about that. I hope to see some decent cards not requiring a small nuclear power plant to function, that's all

What single GPU AMD card required a "small nuclear power plant" exactly?

Perhaps you're unaware, the Fury X (AMD's latest single GPU flagship) has a very similar peak gaming TDP to the Titan XP:

It was also quite close to the 980ti and Maxwell Titan X:

Griffildur, are you suggesting that a 20-30W TDP difference requires a "small nuclear power plant"? If so, I think you need to go back to school

Gerard · 21 Nov 2016 at 19:45

pete910 said:
Where is Lambchop? Not seen him about for a good while now

Don't jinx it :eek:

SKILL · 21 Nov 2016 at 20:42

Kaapstad said:
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.

That's probably one of the most bizarre arguments I've seen, you're using a purely theoretical number, FP32 TFLOPS, to determine that the entirely unrelated memory architecture is better/worse?

You even go on to mention that the difference between the two is down to the clockspeeds, which is right, because it's a theoretical number that's a combination of clock speed and core count.

You've claimed several times that HBM is inherently higher latency than GDDR5/X do you have any sources to back up that claim? As mentioned both are likely to be limited more by the underlying DRAM design/process than the bus type.

JediFragger · 21 Nov 2016 at 21:09

Dave2150 said:
What single GPU AMD card required a "small nuclear power plant" exactly?

Perhaps you're unaware, the Fury X (AMD's latest single GPU flagship) has a very similar peak gaming TDP to the Titan XP:

It was also quite close to the 980ti and Maxwell Titan X:

Griffildur, are you suggesting that a 20-30W TDP difference requires a "small nuclear power plant"? If so, I think you need to go back to school

FuryX is half the speed (or less!) of a TXP dude.... :confused:

Seanspeed · 21 Nov 2016 at 21:12

Dave2150 said:
What single GPU AMD card required a "small nuclear power plant" exactly?

Perhaps you're unaware, the Fury X (AMD's latest single GPU flagship) has a very similar peak gaming TDP to the Titan XP:

It was also quite close to the 980ti and Maxwell Titan X:

Griffildur, are you suggesting that a 20-30W TDP difference requires a "small nuclear power plant"? If so, I think you need to go back to school

I dont think people mind a nuclear reactor so long as it provides power like one.

Dave2150 · 21 Nov 2016 at 21:22

Seanspeed said:
I dont think people mind a nuclear reactor so long as it provides power like one.

I prefer cooler and quieter cards, hence why I'm using a 1070 in my main PC at the moment.

Diagro · 21 Nov 2016 at 21:29

JediFragger said:
FuryX is half the speed (or less!) of a TXP dude....

its also an older generation :rolleyes:

JediFragger · 21 Nov 2016 at 21:41

Diagro said:
its also an older generation

Exactly, unsure why he was comparing them.

Seanspeed · 21 Nov 2016 at 21:46

Dave2150 said:
I prefer cooler and quieter cards, hence why I'm using a 1070 in my main PC at the moment.

Sure. Efficiency has its benefits. I'm not interested at all in 230W+ cards at all.

I know AMD are a ways behind on this right now, but Vega *could* well be AMD's equalizer. I've felt for a while, and have said so in the past, that Polaris was only a half step architecture.

Kaapstad · 21 Nov 2016 at 21:46

SKILL said:
That's probably one of the most bizarre arguments I've seen, you're using a purely theoretical number, FP32 TFLOPS, to determine that the entirely unrelated memory architecture is better/worse?

You even go on to mention that the difference between the two is down to the clockspeeds, which is right, because it's a theoretical number that's a combination of clock speed and core count.

You've claimed several times that HBM is inherently higher latency than GDDR5/X do you have any sources to back up that claim? As mentioned both are likely to be limited more by the underlying DRAM design/process than the bus type.

Nothing theoretical about it, they are NVidia's own figures.

If anyone knows how to measure performance of their cards it will be NVidia.

flopper · 21 Nov 2016 at 21:54

Kaapstad said:
Nothing theoretical about it, they are NVidia's own figures.

If anyone knows how to measure performance of their cards it will be NVidia.

Sure, 4gb cards sold (no 3,5gb)
sold cards defective that they knew would burn for Apple.
Adding a extra 100 tax to cards sold without any reason to.
Didnt have a sli driver for the 580 for over a year really messed up those that bought two.

Cant trust Nvidia to measure anything and telling you about it to be accurate.

Seems easy enough.

Dave2150 · 21 Nov 2016 at 22:05

JediFragger said:
FuryX is half the speed (or less!) of a TXP dude....

JediFragger said:
Exactly, unsure why he was comparing them.

Best to read the thread before posting. Griffildur implied AMD cards required 'small nuclear generators' to function. I corrected him by demonstrating that AMD's previous flagship (FuryX) had vey similar power consumption to the 980ti/TitanX/TitanXP.

Literally no one mentioned performance, as even a small child has the intelligence to know that the 980ti, TitanX, 1070, 1080, TitanXP all obliterate the 4GB FuryX in 99% of all games and benchmarks.

Lokken86 · 21 Nov 2016 at 22:08

Pretty sure Vega will have higher performance/watt than Fiji so who cares. Titan X> FuryX in performance/watt but at a cost everyone knows this.

Kaapstad · 21 Nov 2016 at 22:21

flopper said:
Sure, 4gb cards sold (no 3,5gb)
sold cards defective that they knew would burn for Apple.
Adding a extra 100 tax to cards sold without any reason to.
Didnt have a sli driver for the 580 for over a year really messed up those that bought two.

Cant trust Nvidia to measure anything and telling you about it to be accurate.

Seems easy enough.

Flopper you are brilliant (at scoring own goals). :eek:

My argument about the FP32 performance of NVidia cards is their figures are correct.

You are agreeing with the people who think the Tesla HBM2 card should perform better than the NVidia quoted figure.

CAT-THE-FIFTH · 21 Nov 2016 at 23:21

JediFragger said:
FuryX is half the speed (or less!) of a TXP dude....

Its around 60% to 70% according to TPU. However,in that same review the GTX1070 is around 15% faster than a Fury X. However,in the latest review the GTX1070 is now around 4% faster according to them.