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

So, looking at this. It appears that Vega Frontier Edition is going to be aiming at Titan X (P) as its competition and not Quadros, otherwise it'd be in the FirePro line of GPUs, yes?

So it appears to be going after the prosumer/light professional market. This then makes me wonder if its price will be coming in around the 1k USD mark?

This also makes me wonder if it'll be shipping with consumer drivers or professional drivers? Or will it have the ability to install either depending on the buyer's requirements? IE - the drivers they used for the benchmarks, were they from the FirePro line of drivers or from the consumer Catalyst drivers? Don't have time to look to see if it was documented what drivers were used.

Regards,
SB
 
From AMD's official blog: "But because this graphics card is optimized for professional use cases (and priced accordingly), if gaming is your primary reason for buying a GPU, I’d suggest waiting just a little while longer for the lower-priced, gaming-optimized Radeon RX Vega graphics card. You’ll be glad you did."

Damn, I should have waited to post a few more minutes. :p

Regards,
SB
 
From AMD's official blog: "But because this graphics card is optimized for professional use cases (and priced accordingly), if gaming is your primary reason for buying a GPU, I’d suggest waiting just a little while longer for the lower-priced, gaming-optimized Radeon RX Vega graphics card. You’ll be glad you did."

Do you think that could mean higher clocks for the RX Vega?
 
So, looking at this. It appears that Vega Frontier Edition is going to be aiming at Titan X (P) as its competition and not Quadros, otherwise it'd be in the FirePro line of GPUs, yes?

So it appears to be going after the prosumer/light professional market. This then makes me wonder if its price will be coming in around the 1k USD mark?

This also makes me wonder if it'll be shipping with consumer drivers or professional drivers? Or will it have the ability to install either depending on the buyer's requirements? IE - the drivers they used for the benchmarks, were they from the FirePro line of drivers or from the consumer Catalyst drivers? Don't have time to look to see if it was documented what drivers were used.

Regards,
SB

Im not really sure they aimed at TitanXP ( Titan ), Titan are consumers gpus, and prosumer in name ( was different before but titanXP is just the full GP102 )..... i think Frontier is really only for entreprise, professional, workstation.. but they separate a bit the line as they have, Instinct MI on a side (accelerator), PRO on the other ( but on Polaris). This one is more on the Pro line, bringing video output to MI25.. hence the blue scheme. I will classify this one in the Firepro / Radeon Pro lineup".

Well maybe i will see it differently in some week.
 
Do you think that could mean higher clocks for the RX Vega?

It's doubtful.

Note that Vega FE has two 8-pin power connectors, so I figure its TDP is right on 300W.

AMD-Radeon-Vega-Frontier-Edition-Liquid-Cooling.png


AMD-Radeon-Frontier-Edition-Graphics-Card_2.png


The blog Ryan referenced is here:

http://pro.radeon.com/en-us/vega-frontier-edition/

My guess is that we might see a consumer card with a >300W TDP and, therefore, higher clocks. However, I doubt that the clocks would be significantly higher.

I think the thing to hope for is fast memory. In every release of a product with HBM2, we've never seen 2 Gbps offerings (GP100 used 1.4 Gbps, GV100 will use 1.75 Gbps, etc). It looks like AMD was only able to get to 1.88 Gbps. That's probably close enough to extract good performance, but I hope GP102-competing consumer options are at least as fast in that department.
 
Do you think that could mean higher clocks for the RX Vega?

I don't think a RX Vega reference design with a blower cooler will be clocking its core any higher. Why would they be conservative with the clocks on a (supposedly higher-binned) top-end professional card?
What AMD could launch is a top-end RX Vega version with an AiO watercooler and 8+8 pins that reach those 1600MHz we saw in Compubench.

OTOH, we might see the cards with 4-Hi HBM2 stacks going for their rated 2Gbps, reaching 512GB/s total bandwidth. And I doubt we'll see any consumer RX Vega (top-end AiO version included) with 8-Hi stacks.
 
It's doubtful.

Note that Vega FE has two 8-pin power connectors, so I figure its TDP is right on 300W.

The blog Ryan referenced is here:

http://pro.radeon.com/en-us/vega-frontier-edition/

My guess is that we might see a consumer card with a >300W TDP and, therefore, higher clocks. However, I doubt that the clocks would be significantly higher.

I think the thing to hope for is fast memory. In every release of a product with HBM2, we've never seen 2 Gbps offerings (GP100 used 1.4 Gbps, GV100 will use 1.75 Gbps, etc). It looks like AMD was only able to get to 1.88 Gbps. That's probably close enough to extract good performance, but I hope GP102-competing consumer options are at least as fast in that department.
As pointed out earlier, while the renders indeed have 2x8pin, the card Koduri presented at the event had 8+6pin, so could be either one in the final product
 
I believe I've read that there's a mistake on the Vega FE render and the actual card only features one 8-pin connector + one 6-pin connector, indicating a < 300W power draw.

Edit: thanks Kaotik, you were faster! :)
 
As pointed out earlier, while the renders indeed have 2x8pin, the card Koduri presented at the event had 8+6pin, so could be either one in the final product

Do you have a screencap of this or are you asking me to get your proof for you?

Furthermore, even if this is true, I trust the renders. AMD wouldn't put 2x8-pin front & center if that wasn't the final product.
 
Do you have a screencap of this or are you asking me to get your proof for you?

Furthermore, even if this is true, I trust the renders. AMD wouldn't put 2x8-pin front & center if that wasn't the final product.

Lazy artist who have just use an array modifier or just was use the ES sample board as reference ?
In reality i cant tell what is the final design of it, nor the real tdp.
1h11m10 sec.

Lw47bjr.png



But well we will know enough soon .
 
From AMD's official blog: "But because this graphics card is optimized for professional use cases (and priced accordingly), if gaming is your primary reason for buying a GPU, I’d suggest waiting just a little while longer for the lower-priced, gaming-optimized Radeon RX Vega graphics card. You’ll be glad you did."

Ryan,
was it made clear which Frontier (AIO or blower) was used for the testing and benchmark?
I appreciate Raja held up the blower model, but I could not get any clear indicators if that was also used for the benchmarks and peformance comparisons.
I guess it is also too early to tell what difference (if any are notable) between the two models beyond TDP.
Thanks
 
Lazy artist who have just use an array modifier or just was use the ES sample board as reference ?
In reality i cant tell what is the final design of it, nor the real tdp.
1h11m10 sec.

Lw47bjr.png



But well we will know enough soon .

I agree. That is very clearly a 6+8-pin setup. I appreciate the screencap.

I'm still a little hesitant to believe that they screwed up the render.

Remember, this is the same company that has the peace of mind to tape up their power connectors before opening up the case to Linus for his CES video.

AMD-Vega-10-GPU-Radeon-Vega-Graphics-Card_12-768x432.jpg


I feel like AMD understands the conclusions that people take away from visible power connectors. If I were a betting man, I'd bet that Raja's physical card used an ES board without the debug extension. Without the tape, the ES board is a 6+8-pin (released months after the CES tease).

AMD-Radeon-Vega-GPU-1.jpg
 
Maybe a stupid question but, why don't have always 8+8, even if the card not need it ? Like, a 8 pin connector really cost more than a 6 pin ? I doubt it...
 
But the big surprise going by the numbers; Frontier TDP-TBP/TFLOPs relatively is nearly a match for the 1080ti.
1080ti: 10.6 TFLOPs, 1582MHz, 250W.
OC to 1950MHz 290W (accurately measured by PCPer with oscilloscope) gives around 13 TFLOPs

So from the context of efficiency and going by the numbers presented by AMD Frontier has a perf/w pretty similar to the Pascal 1080ti with peak FP32 and performance equalised; one is GDDR5X and other HBM2 for allowances but still.
However I must say if using the Quadro P6000 or Tesla P40 due to being the full core GV102 they are more efficient with just over 12TFLOPs FP32 at 250W.
So yeah will move some depending upon comparing to enthusiast consumer or professional and HPC models, but the gap is closing relative to Pascal; caveat being potentially mixed precision performance and FP16/Int8 for training and inferencing specifically for Deep Learning frameworks/apps.

I doubt reviewers will see this I assume top Vega for awhile so difficult to say if the AMD numbers measure up in terms of those 3 parameters, but with Fiji they were actually within TDP and reasonably conservative.

Cheers

Well they are comparing Vega FE (presumably the watercooled one) to Tesla P100 in a presumably custom workload we don't know much about. Tesla P100 is rated for 9.3tflops (the pcie version, which they undoubtedly used) and is a 300W double precision card that is passively cooled.
upload_2017-5-17_14-35-7.png

They could have simply used a Quadro P100 for comparisons, which would also have been using Quadro drivers, I suspect that would have ruined the Specview performance chart.
upload_2017-5-17_14-38-21.png
 
Well, lots of PSUs are short on power connectors.

Most PSUs have 6+2 gc power connectors, no ? If you have a 6 power connector avaiblable, I guess you have a 8 too ... I get the "only one 6/8 pin", but the difference between 8+6 and 8+8, not so much.... But sorry it's off topic.
 
Maybe a stupid question but, why don't have always 8+8, even if the card not need it ? Like, a 8 pin connector really cost more than a 6 pin ? I doubt it...

Most new PSUs have at least two 6+2-pin PEG connectors (even something like the oft-discounted ~$30-50 CX450M has 2x 2+6-pin connectors), but I think the slow transition is an artifact of two things.
  • PSUs can have a long lifespan and older models aren't as plentiful in the PEG connector department.
  • OEM machines are notoriously skimpy and PSUs are often affected by this.
    • This is why connector-less <75W cards are so important - it's a drop in replacement to any machine with a 16x PCIe slot.
Long term, I'm no EE, but I bet that there's additional power equipment necessary to deal with the current that those beefier/extra connectors can try to draw. Therefore, the cost to allow 2 8-pin connectors on a 150W card is more than just the physical connectors. That's pure speculation though.
 
Has there ever been a case where a professional card was clocked higher than its consumer equivalent?
This is a good point, however there is always the chance the peak throughput is based on a clock speed that is not sustained by the cards, there could be more aggressive voltage frequency regulation on vega, not to menion the fact that clockspeeds could be close to "the wall" much like Fiji was.

having said all that, still sustained clock speeds of 1500mhz are still a massive improvement and bodes well looking forward
 
Back
Top