[News] AMD to announce Vega tonight [Update: Vega Computex Announcement]

[vario]

Stick around with LordEC911.
I'll definitely call you up upon this.

By the way, Q1 2018 is far from very long time

6-8months ? Eh. I consider it a long time.
Its still to be seen how much better volta is going to be(than pascal). Things arent really what they seem.
I just want to know, what went wrong with vega, maybe its still fixable.484mm2 chip, on 14nm. Thats a big ass chip with absurd power consumption, and performance nowhere to be found.
I mean if they would shrink hawaii and make it bigger a bit faster and give it the same 300W tdp it would be faster.
Polaris is more than half the size ! And has slow bandwith.

[EternityZX9]

I just want to know, what went wrong with vega, maybe its still fixable.484mm2 chip, on 14nm. Thats a big ass chip with absurd power consumption, and performance nowhere to be found. I mean if they would shrink hawaii and make it bigger a bit faster and give it the same 300W tdp it would be faster. Polaris is more than half the size ! And has slow bandwith.

From my understanding, in the short version, the main problem is that AMD bet big on HBM2 and it back fired.

See below, for more details on the long version: Some info taken from Nvidia's SC15 GPU Tech Theater

On further explaining the next generation GPU architectures and efficiency, Stephen pointed out that HBM is a great memory architecture which will be implemented across Pascal and Volta chips but those chips have max bandwidth of 1.2 TB/s (Volta GPU). Moving forward, there exists a looming memory power crisis. HBM2 at 1.2 TB/s sure is great but it adds 60W to the power envelope on a standard GPU. The current implementation of HBM1 on Fiji chips adds around 25W to the chip. Moving onwards, chips with access of 2 TB/s bandwidth will increase the overall power limit on chips which will go from worse to breaking point. A chip with 2.5 TB/s HBM (2nd generation) memory will reach a 120W TDP for the memory architecture alone, a 1.5 times efficient HBM 2 architecture that outputs over 3 TB/s bandwidth will need 160W to feed the memory alone.

Source: Link

[vario]

From my understanding, in the short version, the main problem is that AMD bet big on HBM2 and it back fired.

See below, for more details on the long version: Some info taken from Nvidia's SC15 GPU Tech Theater



Source: Link

Hold on. HBM was supposed to be the "power saver" for the GPU memory architecture.
Anyhow, we have 484GB/s on vega, so thats less than half of this 60W power figure, that would put HBM power consumption in vega at around 30W maybe less.Thas less than GDDR5/x/6.
And The cards are supposed to eat up 300-400W. HBM doesnt really explain Vega for me.
Fiji had 4 stacks of HBM1, Vega has 2 Stacks of HBM2 at LOWER bandwith and still eats anyhwere between 25 and 100W more than Fiji! Something more is wrong here.
I know that HBM2 availability was a thing that surely went wrong for AMD. But i dont know about the power consumption and performance.

[EternityZX9]

Hold on. HBM was supposed to be the "power saver" for the GPU memory architecture.

That's the key..."supposed to" is turning out to not. One of the two main reasons nvidia hasn't switched to HBM2 for consumer grade products availability and power consumption. It's shown right in the link directly from nvidia this was a major concern for them.

[vario]

That's the key..."supposed to" is turning out to not. One of the two main reasons nvidia hasn't switched to HBM2 for consumer grade products availability and power consumption. It's shown right in the link directly from nvidia this was a major concern for them.

I get that HBM2 in the near future is going to be a power problem. But according to this slide, Vega is at most at the "unhappy" emoji phase, which means still better (lower) power consumption for the memory subsystem than the gddr5. Isnt it ?
I mean, remember, it has TWO SLOW HBM2 stacks, thats in contrast with FOUR FASTER HBM1 stacks of Fury X. its not even at 0.5TB/s yet, and you are talking about a problematic 1,2TB/s+ speeds.That cant explain a 100W+ power envelope over FURY.
I only see it as a major AMD fckup, or GLOFO`s fault.

@SinOfLiberty
Is vega 11 the same chip with disabled CU`s ,or is it a separate chip ?

[BeepBeep2]

I get that HBM2 in the near future is going to be a power problem. But according to this slide, Vega is at most at the "unhappy" emoji phase, which means still better (lower) power consumption for the memory subsystem than the gddr5. Isnt it ?
I mean, remember, it has TWO SLOW HBM2 stacks, thats in contrast with FOUR FASTER HBM1 stacks of Fury X. its not even at 0.5TB/s yet, and you are talking about a problematic 1,2TB/s+ speeds.That cant explain a 100W+ power envelope over FURY.
I only see it as a major AMD fckup, or GLOFO`s fault.

@SinOfLiberty
Is vega 11 the same chip with disabled CU`s ,or is it a separate chip ?

Unhappy emoji phase is actually 4 stacks, look at the projected bandwidth (1000 GB/s), 2 stacks HBM2 should be AT HBM 4 stack power consumption or less. Anyone thinking the HBM2 is using lots and lots of power is probably wrong, the 2 stacks should be at least 2x the perf / watt of GDDR5 @ <250 GB/s and 3x+ perf / watt of GDDR5 @ ~500GB/s.

Further backing this up is the mem VRM on the Vega FE card, OnSemi NTMFD4C86N which is only going to do about 25-30 amps @ 1.2-1.3v <40w

It is absolutely mind-blowing to think that Zen and Vega are both on 14nm LPP considering how great Zen is and how poorly Vega is performing.

It is way more cut down w/o HBM2

It is all a load of bull.. As in practice, Power Consumption goes out of the window once the card is clocked beyond certain frequencies. In case of AMD, dat huge аss number of W`s comes from exactly that. The chip is running at its highest freq, that the hardware is capable of.

But again, this is not how it will be marketed as. Just an extra option for OC`ers and reviewers to demonstrate what the card is really capable of when pushed to max.

How do you know this?

Look - there aren't even ES samples of Vega 11 yet nor have there ever been and I don't think we'll see one.
Vega 10 should easily scale down to the 150w TDP range Polaris 10 occupies. My bet is that Vega 11, if a design actually existed, has been cancelled.

Leaked Q4'16 slides with Vega 11 show it slotting in for inference applications below Vega 10 but today we have the Radeon Instinct MI8 instead, a 175w R9 Nano in new clothing. The MI8 for FP32 most likely matches or beats the MI25 in performance / watt at stock performance levels.

Vega 11 wouldn't make sense unless it is equipped with the same FP16 capability as Vega 10 and extremely huge GDDR5X bus and 20GB+ GDDR5X...but then again, it would use way more power than Vega 10 due to the GDDR5X power consumption.

[vario]

...

Look - there aren't even ES samples of Vega 11 yet nor have there ever been and I don't think we'll see one.
Vega 10 should easily scale down to the 150w TDP range Polaris 10 occupies. My bet is that Vega 11, if a design actually existed, has been cancelled...

I dont know. But Vega 11 on GDDR5(x), as in smaller ANOTHER chip, would be a very welcomed thing for AMD.Yes Vega 10 can scale down.
But its BIG, it needs HBM2, it needs inerposer. A smaller chip (much cheaper to produce) on a widespread cheap memory is certainly something that AMD needs...

What worries me is that at least NOW, it seems, that polaris has more performance for mm2/hz(up to 1300mhz or so).So that could be a reason to can it, or maybe rework it.