[Videocardz rumour] Vega pushed forward to October

[railven]

yeah I guess if they want to forget it altogether and go custom only like that. I would think (with HardOCP's article stating Falcon Northwest boss wanting high quality blower) that AMD wouldn't want to miss out on those boutique builders buying a large quantity and those guys require a reference blower.

Well I'd assume AMD would contract someone else. But that's me personally. If the go CLC, they go CLC.

AMD should have a proper founders edition with watercooling built in. Not this wannabe crap from nvidia. That would actually allow it to OC better than regular.

Just like the Fury X (sorry low hanging fruit )

 

[MrTeal]

If that so, GP100 might get delayed then too... Both for Christmas and AMD Zen (no one cares Kabylake, the Godavari of Intel)...

Why would that affect Zen or GP100? P100 isn't even shipping in OEM systems until Q1 2017 according to JHH, let alone retail Tesla modules or a GPU based on the die. Seems completely unrelated.

 

[Snarf Snarf]

In a perfect world, this is the big die Vega coming out just in time for X-mas splurge on this and an HDR monitor to replace my 980ti and Qnix QHD monitor.

 

[raghu78]

Why do you even need 4096 bit?

Small Vega using improved compression should be able to get by with 2 stacks of HBM2 giving 410-512 GB/s and 8GB memory. A much smaller interposer and also saving a bit on memory controller die space. Should keep production costs down a fair amount plus less assembly defects as only 3 components vs 5. Also allows twice the production output early in the cycle when HBM2 will be in short supply.

Small Vega might very well use just 2 HBM2 stacks. I am talking about the big die Vega which also will serve as AMD's next gen Firepro flagship. That would surely be 4 HBM2 stacks. 16 GB Radeon. 32 GB firepro using 8 Hi HBM2 stacks.

 

[el etro]

Too early to speculate on Vega. Without basic Polaris data is all too early.

 

[maddie]

Small Vega might very well use just 2 HBM2 stacks. I am talking about the big die Vega which also will serve as AMD's next gen Firepro flagship. That would surely be 4 HBM2 stacks. 16 GB Radeon. 32 GB firepro using 8 Hi HBM2 stacks.

Everybody likes it big.

Seriously though, If Vega is coming earlier than thought, we should expect the smaller one first, especially if it can use 2 HBM2 stacks. Supplies will be tight early in the HBM2 ramp.

 

[RussianSensation]

Are you serious? A person with a $200- $300 budget should wait on Vega?

RS showed that only 15% of total discrete sales were above $300 and 85% was $300 and below. This is the present reality.

For 2015, it was actually 11.8% [5.9M/50M] but I estimated an increase to 15% by now given the trends towards more expensive cards.

"The total number of AIBs sold in 2015 was 50 million compared to 44 million in 2014. Of these, about 5.9 million are enthusiast level AIBs (add-in boards / discreet graphics cards) shipped in 2015 compared to 2.9 million in 2014." ~ Source

We could see >=$300 dGPU segment experiencing growth in 2016 due to some highly anticipated games coming out, DX12 games more or less making older GPUs obsolete forcing upgrades and record high pent-up demand from gamers using older gen cards (i.e., 70% of NV users are on pre-Maxwell cards).

Just my guess. As cool as it was, it hurt Fury X at launch due to pump issues, it added cost to the overall build which was most likely needed to keep the card cool and quiet.

The pump issues were unfortunate but were later fixed by CM. The reason the AIO CLC was 'required' is due to the card's small 7.5" footprint. How would you cool 280W with a blower/open air cooler? It's not required if you extend the PCB -- in fact Sapphire Fury Tri-X is more or less as quiet at max load @ max overclocked than a Titan blower 980Ti is at idle. So no, it's not "required" to have AIO CLC to cool a 280-300W 10.5-12 inch card, but sure a 7.5" card requires it.

Unfortunately for AMD, using the far superior cooling system brought too much negative publicity due to poor launch execution. With a properly working AIO CLC, most people would be amazed at how well it works on 250-300W cards, or even 500W cards.

Agree (though harder to fit multiple CLC radiators in multi-card boutique rigs)

FWIW - I hope they keep doing the reference CLC. Its one of the things that attracted me to the Fury X in the first place and its a good differentiator for AMD

After using an AIO CLC on the R9 295X2 and seeing the improved larger rad doing even better on Radeon Pro Duo, I am sold. If prices are similar and the card is well built, I'd always go AIO CLC > open air cooling > blower.

Asus Strix 980Ti SLI = 85C @ 58% fan speed
Radeon Pro Duo = 55C @ 24% fan speed (almost all of it is exhausted out of the case)
https://www.youtube.com/watch?v=CMsoQwbxELo

If the case can accommodate it, it's impossible to compete with 2-4 cards with AIO CLCs exhausting 95% of 250-300W of heat out of the case, while operating quietly. To outperform that, a full custom water loop costing hundreds if not $1000+ is required, and of course you don't get any warranty with it.

As far as 90%+ of the gaming market goes, I cannot think of many cases that won't fit a single 120mm rad. AMD or NV, AIO CLC is simply amazing compared to any open air cooled card, with the ONLY exception being MSI Lightning 980Ti.

Moving-up Vega is GREAT news. AMD is taking Pascal's release seriously and doesn't want to abandon the high-end for 9 months. Interesting enough, this same timeframe is probably when we will see NV launch the 1060/1060Ti. Different approaches for sure, and it will be interesting to see what bears more fruit.

Regardless, AMD has the opportunity to release something better than Pascal 1.0 and trump some enthusiasm for NV HBM parts in 2017. Great move. Also could give current 980Ti owners a more beneficial upgrade path vs the 1080/1070 that targets more 7xx and 6xx users IMHO.

If true, I am concerned that they are going to sacrifice some performance. Why would all of a sudden Vega move up 2-3 months? Hopefully it has more to do with the volume of HBM2 and yields rather than AMD releasing a cut-down Vega 11 or worse Vega 11 with 15-20% lower GPU clocks on top just so that they have something on the table.

Then again, this 14nm node could be very long, which means AMD/NV will love to sell us 3-4 flagships on the same node over the next 3-4 years.

 

[C@mM!]

With Polaris not having any AIB partners showing cards at Computex, and Vega being moved up, something seems fishy.

Polaris dies being bought for XB1.5\PS4.5 and other OEM's is the first thing that comes to mind.

 

[eRacer]

For 2015, it was actually 11.8% [5.9M/50M] but I estimated an increase to 15% by now given the trends towards more expensive cards.

"The total number of AIBs sold in 2015 was 50 million compared to 44 million in 2014. Of these, about 5.9 million are enthusiast level AIBs (add-in boards / discreet graphics cards) shipped in 2015 compared to 2.9 million in 2014." ~ Source

The quote is wrong based on the graph and the mention of declining overall AIB sales in 2015. The 44 million units actually applies to 2015, while 50 million should be 2014.

 

[JDG1980]

With Polaris not having any AIB partners showing cards at Computex, and Vega being moved up, something seems fishy.

Polaris dies being bought for XB1.5\PS4.5 and other OEM's is the first thing that comes to mind.

Current word is that AMD is planning to launch Polaris not at Computex, but at a separate show in Macau.

PS4K won't be using off-the-shelf Polaris chips, but a custom SoC containing the Polaris IP. Besides, that won't be released until later. I do think that Apple will be soaking up a lot of the early Polaris production for products to be released at WWDC in mid-June.

 

[tential]

Current word is that AMD is planning to launch Polaris not at Computex, but at a separate show in Macau.

PS4K won't be using off-the-shelf Polaris chips, but a custom SoC containing the Polaris IP. Besides, that won't be released until later. I do think that Apple will be soaking up a lot of the early Polaris production for products to be released at WWDC in mid-June.

What a surprise, the show starts 1 day before the 1080 launches.

 

[RussianSensation]

The quote is wrong based on the graph and the mention of declining overall AIB sales in 2015. The 44 million units actually applies to 2015, while 50 million should be 2014.

Either way, the point still stands. Using your numbers:

5.9M/44M = 13.4%.

It's reasonable to estimate that 85% of dGPU sales are $300 and below as of now. It took $330 970 1.5 years to reach 5.10% market share. The vast majority of PC gamers worldwide (unfortunately) still use 1920x1200 60Hz or lower. 1070/1080 level of performance is a wast of $$ for such low end monitors. If most people had 1080p 120-200Hz and Skylake i7 @ 4.8Ghz, it would make sense.

NV even pushed hard to have HDMI 2.0, DP1.4. These cards are aimed dead on at 2560x1080 and above users, especially 2560x1440 and 4K users. Mainstream gamers who buy $300 dGPU or less are less likely to have those types of monitors. The reason why AMD is prioritizing $300 and below market segment is that's where 85% of the sales are and market share gains to be had.

 

[flopper]

question will be if amd talk vega later this month alongside Polaris?

 

[Glo.]

We know that Vega is supposed to triple performance per watt compared to previous generations of GPUs, yes?
We know that Vega is supposed to have 4096 GCN cores and HBM2, yes?

Fiji has similar core count. Now, lets look at Fury X efficiency.

275W of TDP, 8.6 TFLOPs of compute power. That gives roughly 31.2 GFLOPs/watt.
So if the slide that AMD presented is true, we are looking on GPU that gives around 90GFLOPs/watt.

4096 GCN cores, clocked at 1500 MHz with 140W TDP. 12 288 GFLOPs. 87.7GFLOPs/Watt.

I do not want to hype it up, I just wanted to give a little perspective to understand what is happening.

 

[Ancalagon44]

We know that Vega is supposed to triple performance per watt compared to previous generations of GPUs, yes?
We know that Vega is supposed to have 4096 GCN cores and HBM2, yes?

Fiji has similar core count. Now, lets look at Fury X efficiency.

275W of TDP, 8.6 TFLOPs of compute power. That gives roughly 31.2 GFLOPs/watt.
So if the slide that AMD presented is true, we are looking on GPU that gives around 90GFLOPs/watt.

4096 GCN cores, clocked at 1500 MHz with 140W TDP. 12 288 GFLOPs. 87.7GFLOPs/Watt.

I do not want to hype it up, I just wanted to give a little perspective to understand what is happening.

2.5X performance per watt, not triple.

 

[Glo.]

That is for Polaris. Vega is higher than Polaris on this slide:
https://www.extremetech.com/wp-content/uploads/2016/03/AMDGPU.jpg

.

 

[ShintaiDK]

You assume it is compared to a Fury X. But it could be its not even compared to a Fiji GPU.

Anyway, this October rumour is based on nothing but a single forum post at beyond3d.

 

[Ancalagon44]

Fair enough - didn't realize the Y axis of that graph was performance per watt.

Performance per watt will be dependent on the memory technology used though. If they make a Vega that uses GDDR5X instead of HBM2, it won't be as efficient.

 

[airfathaaaaa]

You assume it is compared to a Fury X. But it could be its not even compared to a Fiji GPU.

Anyway, this October rumour is based on nothing but a single forum post at beyond3d.

well the early slides (not the new ones) had polaris lines from down 300 till up the fury lines..
and vega was starting from fury lines and above

 

[beginner99]

Fair enough - didn't realize the Y axis of that graph was performance per watt.

Performance per watt will be dependent on the memory technology used though. If they make a Vega that uses GDDR5X instead of HBM2, it won't be as efficient.

GDDR5x would make more sense for small Vega than 2-stack HBM2. It's cheaper and you get similar bandwidth.

 

[Ancalagon44]

GDDR5x would make more sense for small Vega than 2-stack HBM2. It's cheaper and you get similar bandwidth.

My guess:

Big Vega Pro: HBM2
Big Vega non Pro: HBM2
Small Vega Pro: GDDR5X
Small Vega non Pro: GDDR5X
Big Polaris Pro: GDDR5X
Big Polaris non Pro: GDDR5
Small Polaris: GDDR5

 

[cyclohexane]

Hell yes! competition working the way it should be.

 

[Krteq]

Anyway, this October rumour is based on nothing but a single forum post at beyond3d.

Welp, no

There is a slide from eSilicon presentation for 14nm LPP HBM2 integration - test chip tepe out at March 2016 - Final Models availability at September 2016

//And there are some other ESes circling the world in Zauba with higher price then Polaris ones.

 

[maddie]

GDDR5x would make more sense for small Vega than 2-stack HBM2. It's cheaper and you get similar bandwidth.

My guess:

Big Vega Pro: HBM2
Big Vega non Pro: HBM2
Small Vega Pro: GDDR5X
Small Vega non Pro: GDDR5X
Big Polaris Pro: GDDR5X
Big Polaris non Pro: GDDR5
Small Polaris: GDDR5

The thing is the improved perf/W compared to Polaris. I doubt we will see many architectural improvements so soon to account for this, implying HBM2.

Small Vega can use 2 stacks of HBM2 and still have the bandwidth of FuryX. More than enough. Also allows a small interposer and lower assembly costs. 4 components vs 6.

Chicken and egg. For HBM2 based products to drop in price, you have to increase production and vice versa.

 

[Krteq]

Deleted