AMD Vega 10, Vega 11, Vega 12 and Vega 20 Rumors and Discussion

[Deleted member 13524]

No way AMD doesn't release a $700 Vega. Even if it needs a CLC and a 300W TDP to barely keep up with the 1080 Ti, AMD will do it. Otherwise, they are leaving money on the table via unexploited market segmentation.

AMD has lost quite a bit of popularity in the PC market in the last couple of years, even more in the high-end area.
They'll need to undercut the 1080 Ti in price, and if that means a $600 full-fledged Vega then that's just the price for not competing in that market for almost 2 years.

 

[ImSpartacus]

AMD has lost quite a bit of popularity in the PC market in the last couple of years, even more in the high-end area.
They'll need to undercut the 1080 Ti in price, and if that means a $600 full-fledged Vega then that's just the price for not competing in that market for almost 2 years.

I think AMD can have their cake and eat it too.

AMD can have a presence at $700 and simultaneously be price competitive.

If I were AMD, I'd line things up as follows:

$700 Vega XT

• CLC (and clocks as high as possible)
• Short Fury X-sized PCB (take the compact gpu crown)
• 2 Gbps HBM2 (even if it causes severe supply issues)
• Potentially >250W TDP (no need to jive with OEM cooling standards if you're already violating them with a clc)
• No custom cards (just like Fury X)

$600 Vega XT

• Air cooled (with more reasonable clocks)
• ~1.6 Gbps HBM2 (assuming supply issues w/2 Gbps)
• <=250W TDP (make OEMs happy)
• Custom cards allowed (maybe even save R&D $ with no reference card, a la Fury)

$450 Vega Pro

• Air cooled (with the lowest clocks, relatively)
• 1.6 Gbps HBM2
• <=250W TDP (make OEMs happy)
• Custom cards allowed (maybe even save R&D $ with no reference card, a la Fury)

Do it that way and you get several things going right for AMD:

• Proper price discrimination of AMD fanboys (market segmentation matters, detestable as it might feel)
• A legitimate "halo" card to gain back some mindshare from Nvidia (even if it can't be kept in stock)
• The small form factor desktop performance crown
• Undercutting price on both 1080 Ti and 1080
• Efficient use of HBM2 supply (in case rumors of 2 Gbps shortages are true)
• Potentially only one reference design (R&D is tight at AMD)

 

[Deleted member 2197]

A liquid cooled dual GPU Radeon RX Vega graphics card has been spotted in a brand new Linux driver released by AMD yesterday.

table->Tliquid1Limit = cpu_to_le16(tdp_table->usTemperatureLimitLiquid1);
table->Tliquid2Limit = cpu_to_le16(tdp_table->usTemperatureLimitLiquid2);

http://wccftech.com/amd-dual-gpu-vega-liquid-cooled-graphics-card-spotted-in-linux-drivers/

 

[Deleted member 2197]

I mean looking at performance a 64 rop Polaris with higher clocks would be 1080 performance levels . So unless AMD did nothing but that they should be close to 1080TI performance imo

Now, based on AMD’s performance figure for its Vega 10 based Radeon Instinct MI25 accelerator we know that the company expects Vega to run at above 1.55GHz. The MI25’s single precision compute performance figure is 12.5 TFLOPS and its half precision compute performance is rated at 25 TFLOPS. AMD’s graphics guru Raja Koduri revealed that this is the reason the accelerator was in fact named the MI25.
...
Realistically speaking, if AMD manages to successfully hit its 1.56GHz projection, production ready Vega 10 will run at a clock speed that’s 30% higher than what we’re seeing from the 687F:C1 prototype. Performance doesn’t necessarily scale in perfect linearity with the core clock speed however, as other factors play a major role like memory bandwidth and internal architectural bottlenecks that we know very little about at this point.

According to the JEDEC spec HBM2 should scale all the way to 1000MHz/2Gbps but currently available HBM2 stacks are limited to 800MHz/1.6Gbps. With that in mind, a 30% higher GPU clock speed and a 100MHz boost to the memory would move Vega 10 from its current 1070’ish spot to a more competitively favorable position with the GTX 1080.

http://wccftech.com/amd-vega-10-3dm...c1-device-8gb-700mhz-vram-1200mhz-core-clock/

 

[entity279]

^^

Except the accelerators are passively cooled not because Vega can confortably reach past 1.55 Ghz, but because this is the way to install Gpu accelerators in a rack. The airflow there will be far and above anyone of us has in their cases ( I hope at least.. for neighbours' sake)

 

[Geeforcer]

^^

Except the accelerators are passively cooled not because Vega can confortably reach past 1.55 Ghz, but because this is the way to install Gpu accelerators in a rack. The airflow there will be far and above anyone of us has in their cases ( I hope at least.. for neighbours' sake)

This is a point worth emphasizing, because of the prevalence of "Vega 1.55GHz on AIR, 2GHz consumer cards & Half-Life 3 confirmed!!1" sentiment on certain websites. These accelerators (which includes P100, also passively cooled) simply shift burden of generating sufficient airflow onto rack's cooling system.

 

[3dilettante]

The use of the word passive seems to evoke a different image for consumers versus server racks, although it's true that generally even fanless solutions in a desktop are often relying at least a little on case/PSU fans for decent results. It seems like the level of obligatory air flow could in some way be used to distinguish one from the other.

 

[Ryan Smith]

The use of the word passive seems to evoke a different image for consumers versus server racks, although it's true that generally even fanless solutions in a desktop are often relying at least a little on case/PSU fans for decent results. It seems like the level of obligatory air flow could in some way be used to distinguish one from the other.

TDP/power consumption is a far more useful metric in this respect.

 

[entity279]

The issue for me is that someone from wccftech staff made this confusion.

 

[Anarchist4000]

Or is "consumer segment" as of Samsungs proposal to be interpreted as mainstream/entry-level market as opposed to the high-end cards that were referenced by System Plus Consulting?

That would be my guess. It's also possible the low cost implies not using an interposer, but within the package to decrease form factor. Ideally as a replacement for DIMMs which we haven't really seen yet.

 

[Malo]

The issue for me is that someone from wccftech staff made this confusion.

I doubt it's confusion. They will take anything from anywhere to make a click-bait "article" from it, irrespective of its source or validity.

 

[eastmen]

http://wccftech.com/amd-vega-10-3dm...c1-device-8gb-700mhz-vram-1200mhz-core-clock/

we have no idea what that chip actually is. New features may not be working on it , drivers may not expose new features and so on. I will wait another 2 weeks at this point itsn ot much of a wait

 

[HKS]

TDP/power consumption is a far more useful metric in this respect.

Yes, and according to AMD the TDP on Instinct MI25 is < 300W, so same envelope as Tesla P100 and Tesla V100 (which will ship in Q3)

http://instinct.radeon.com/en-us/about/

 

[xEx]

Yes, and according to AMD the TDP on Instinct MI25 is < 300W, so same envelope as Tesla P100 and Tesla V100 (which will ship in Q3)

http://instinct.radeon.com/en-us/about/

Yes but well probably only the investment in the design of volta was higher than the R&D of AMD. that chip is really a statement from Nvidia kind of the f40 from Ferrari in its days.

 

[Panzer_Penguin]

Yes, and according to AMD the TDP on Instinct MI25 is < 300W, so same envelope as Tesla P100 and Tesla V100 (which will ship in Q3)

http://instinct.radeon.com/en-us/about/

Seems like 300W is just fairly vague envelope. If this image is legitimate then Vega 10 is 12TF @ 225W

https://videocardz.com/65521/amd-vega-10-and-vega-20-slides-revealed

The extra +0.5TF shouldn't need an extra ~75W of power.

 

[CarstenS]

Do not confuse maximum throughput and sustained throughput at maximum load. The gap between those two metrics is becoming the most important metric.

 

[lanek]

Do not confuse maximum throughput and sustained throughput at maximum load. The gap between those two metrics is becoming the most important metric.

thoses Tflops are only calculated by mhz x shaders ... This said AMD have a nice efficiency on computing ... ( i mean computing as not gaming ).. if you have 12Tflops, efficiency is nearly at 11.99Tflops (OpenCL, Directcompute etc. )..

Ofc, theres the question of turbo mode.. is the clock are sustained at max level ? or not ? cooling enter too in question... now.. for be honest, in computing mode, if shaders are used at 100%, other parts of the gpus are idle, so it is not really as in gaming.

GCN 1 was allready absolutely amazing in term of efficiency for this.. 4 Tflops max base mean nearly 4 Tflops of sustained throughput.

 

[Deleted member 13524]

AMD's Computex press conference scheduled for May 31st at 10 a.m.

This is when they announced the RX480 last year, but this year they might have a lot more to talk about:

- Vega
- Two different Raven Ridge SoCs (a small 2-core/3-NCU 4-15W chip and a larger 4-core/11-NCU 15-65W)
- Naples (apparently 4*Summit Ridge in MCM)
- Snowy Owl (apparently 2*Summit Ridge in MCM)

 

[CarstenS]

FWIW, golem.de claims, that the interposer is indeed two separte ones.
https://www.golem.de/news/nvidias-g...mit-des-technisch-moeglichen-1705-127773.html
The relevant passage for your google translate pleasure:
"Der Interposer, auf welchem der GV100 und die vier HDM2-Speicherstapel sitzen, sprengt die Dimensionen der Maske (Reticle), weshalb zwei benötigt werden."
Or, if you take my (deliberately literal) word for it:
"The Interposer on which the GV100 and the four HDM2-stacks sit, blows the dimensions of the reticle so that two are required."
From the grammatical structure, two would refer to reticles. It could be a a word omission though:
-> "…weshalb zwei Belichtungsdurchgänge benötigt werden."
-> "… so that two exposures are required".

 

[CarstenS]

thoses Tflops are only calculated by mhz x shaders ... This said AMD have a nice efficiency on computing ... ( i mean computing as not gaming ).. if you have 12Tflops, efficiency is nearly at 11.99Tflops (OpenCL, Directcompute etc. )..

Ofc, theres the question of turbo mode.. is the clock are sustained at max level ? or not ? cooling enter too in question... now.. for be honest, in computing mode, if shaders are used at 100%, other parts of the gpus are idle, so it is not really as in gaming.

GCN 1 was allready absolutely amazing in term of efficiency for this.. 4 Tflops max base mean nearly 4 Tflops of sustained throughput.

Of course I was refererring to the turbo mode, hence "maximum load".

Another interesting question to ponder would be, when Nvidia (or AMD, for that matter) sees the HPC/AI market as sufficiently large enough for the **100 chips to no longer require graphics specific assets like rasterizers and texture mapping units. GP100 already has not seen much use outside of professional products, mainly Tesla and a bit of Quadro. At least the TMUs are still part of GV100 and since it is partitioned in GPCs, it seems like a safe bet that the rasterizer parts are still on board as well.