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

Why would AMD and NVidia be involved in this? It's the AIBs which take GPU chips from the IHVs and add all the gubbins required to make a card you can buy. It's also in their interest to offer a short warranty. And, it's also in their interest to sell cards that won't hit the second hand market as gamer cards.

Is it not in AMD's best interest to try to come up with a way to have their latest gaming graphics cards ending up in gamers' hands instead of miners'?

Miners getting all the new cards means marketshare on the actual gaming community stays low, which in turn means development efforts on optimizing for that architecture will also be low.
This is a problem that AMD needs to address by themselves, as AIBs don't tend to be farsighted at all.
 
Why would AMD and NVidia be involved in this? It's the AIBs which take GPU chips from the IHVs and add all the gubbins required to make a card you can buy. It's also in their interest to offer a short warranty. And, it's also in their interest to sell cards that won't hit the second hand market as gamer cards.
Because AMD and NVIDIA could sell them GPUs configured different at cheaper price?
 
Have to ask for a source on this. Went through a dozen Nvidia pages and blogs and nothing official that I've come across states that. Just FP32/64 and Tensor. Emphasis on official and not a sites interpretation.
Maybe this should be done in the Volta thread?
The rest of the posts that kicked this off are over there.
Like I mention and show the performance data is Vec2 FP16 with Caffe2 (this was only recently updated to FP16 training but with cuDNN7), then I quoted one of the senior and experienced CUDA devs at devtalk that also mention cuDNN to date for real world use with certain frameworks does not currently support the Tensor cores, which also ties into the Caffe2 results with Volta 100 vs GP100.

You need to be able to explain this for it to be removed; how would every current traditional HPC implementation out there using cuBLAS and GEMM or designed around cuDNN/TensorRT work with totally different instruction and structure requirement of Tensor mixed precision cores compared to existing Vec2/SGemmEx/HGemm if Nvidia for some crazy reason removed Vec2/HGemm/SGemmEX from V100?
Cheers
 
Last edited:
Interesting tid bits on apple's imac pagefirefox_2017-06-06_20-31-13.png

r9 395m was what was in previous imac,

that puts Vega at around or just above the gtx 1080 in performance (granted clock speeds and what not have to be looked into)

Just keeping in mind that Vega 10 in this Imac pro is replacing a 250 watt TDP m r9 395
 
Last edited:
So 3.5tf x 3.x is...

So you expect 0 gains in per clock performance with Vega?


Hey that is the best case from what Apple is saying so.....

If we believe in their marketing team, then yeah but If its hitting gtx 1080 or just above performance at 250 watts, its quite a bit more efficient than a Fury X, that's 40% more in perf/watt.
 
Hey that is the best case from what Apple is saying so.....

If we believe in their marketing team, then yeah but If its hitting gtx 1080 or just above performance at 250 watts, its quite a bit more efficient than a Fury X, that's 40% more in perf/watt.

It's also something you'd expect from a straight shrink with no architectural gains.

I don't think it's worth reading into marketing, they probably just multiplied the tflops number and called it a day.
 
Just keeping in mind that Vega 10 in this Imac pro is replacing a 250 watt TDP m r9 395
No, it's not.

Desktop Tonga XT (R9 380X) has a TDP of 190W. Notebookcheck claims the M395X has a 125W TDP.

No iMac has ever carried a 250W GPU. It's a bit ridiculous to suggest that, to be honest. If they could fit 250W GPUs in those AiOs, why would they even bother using only mobile versions of GPUs?


Hey that is the best case from what Apple is saying so.....
No, it's not.

"The Radeon Pro Vega is over three times faster than any previous iMac GPU"


Having 3X better performance than the M395X is the worst case scenario. And since that text doesn't mention which version of Vega they're comparing to, obvious comparison would be for the lowest-end Vega 56.
Otherwise the wording would be "up to", not "over".
 
Last edited by a moderator:
The GPUBoss lists both M395X Mac and M395X TDPs at 250W which are obviously both wrong. We know that PC-version of M395X has 125W TDP.
If we believe the GPUBoss specs are correct, Mac-edition would have ~25% higher GPU and ~9% higher memclocks - you're really suggesting that those would get the TDP from 125 to 250W?

Marketing never shows the lowest they possible can do, they always show the best they can do. Also they even try to lie if its possible, so in this case, I would expect no different, unless Apple marketing got religion as of late.
You're right that marketing never shows the lowest they possible can do. You however completely ignore that "over 3x" doesn't limit how much over it is, it could be anything between 3.0....01x to 3.9....9x before we can be certain they'd use 4x instead.
 
The GPUBoss lists both M395X Mac and M395X TDPs at 250W which are obviously both wrong. We know that PC-version of M395X has 125W TDP.
If we believe the GPUBoss specs are correct, Mac-edition would have ~25% higher GPU and ~9% higher memclocks - you're really suggesting that those would get the TDP from 125 to 250W?


yes when Hawaii and Granada clocks go up, power consumption goes up at a much higher rate than the frequency increase.

The frequency of the r9 395x mac edition is 900 mhz, which is 100 mhz less than a r290x, the r290x has a TDP of 300 watts.

And we have seen AIO's that use full cards so the form factor of this Imac can't even use that as an excuse.

You're right that marketing never shows the lowest they possible can do. You however completely ignore that "over 3x" doesn't limit how much over it is, it could be anything between 3.0....01x to 3.9....9x before we can be certain they'd use 4x instead.

I always take the lowest factor in marketer's case it can be 3.01x, if this is compute performance, which in a mac pro's case it probably is, that doesn't bode well for gaming. If they stated average, then we can presume some will be a factor higher and a factor lower, but they didn't.

Has AMD shown anything more than that so far? All their games they have shown have been around gtx 1080 performance or just above.
 
Last edited:
You're right that marketing never shows the lowest they possible can do. You however completely ignore that "over 3x" doesn't limit how much over it is, it could be anything between 3.0....01x to 3.9....9x before we can be certain they'd use 4x instead.

There seems to have been a cleanup so that bit got lost, but I think it's important to mention that one of the options in the iMac Pro is a cut-down Vega with 54 CUs.

The "over 3x faster" text doesn't say it applies to the faster (and probably very expensive) full Vega with 16GB HBM2, only that it applies to "Vega" in general.

So not only is the "over 3x" statement a "lower limit", it's probably a lower limit applied to the cut-down Vega 54 and not the fully enabled version.

If apple wanted to declare an upper limit they would've written "up to X times faster", and then we should assume the model in question would be the fully-enabled 16GB version. Declaring "over X timer faster" means the opposite.


Isn't the R9 m395X Tonga based?
Yes. It's the exact same chip as the desktop R9 380X graphics card, which has a 190W TDP and has higher core and memory clocks.
You (or anyone else) can take your own conclusions about this side topic from this fact.
 
Last edited by a moderator:
Isn't the R9 m395X Tonga based?

Tonga XT has a 190W TDP as a desktop board, to suggest that mobile Tonga would ever have a 250W TDP is ridiculous.
if that is case then I'm incorrect

But if that is the case as well, my performance extrapolation is way off to the north, and Vega will perform like x3 of tonga? Thats not good at all!

Tonga is what 1050 performance levels? That puts x3 or x4 performance above 1070 at x3 and at x4 above 1080.
 
if that is case then I'm incorrect

But if that is the case as well, my performance extrapolation is way off to the north, and Vega will perform like x3 of tonga? Thats not good at all!

Tonga is what 1050 performance levels? That puts x3 or x4 performance above 1070 at x3 and at x4 above 1080.
The Mac-versions are clocked around 20% lower than Vega Frontier Edition and probably RX's too, though.

But, for comparisons sake, you're estimating Tonga performance far too low.
Going by TPU reviews, 380X (Tonga XT's desktop version) is around 7-8 % faster than GTX 1050 Ti factory-OC (msi gaming x), which makes it over 30 % faster than 1050 (also factory OC version by MSI). 380X is clocked at 970 MHz vs ~900 MHz of M395X, but both use full chip.

edit: PCGH tests show how vulnerable these numbers are based on tested software and all
 
edit: PCGH tests show how vulnerable these numbers are based on tested software and all
Since we're testing with "guaranteed/typcial" boost clocks, those numbers might not be too far off of each other. Pascal chips tend to boost quite a bit, especially the GP107 with their lowly advertised clock rates.
 
Last edited:
Back
Top