AMD Vega Hardware Reviews

Here they actually consider Navi as a "Performance Increase" while they don't really give Vega any real designation, for example Enthusiast or High End for that manner. Which makes me believe that Vega was just entry level Arch. paving way for Navi. Be that true or false remains to be seen. We know that they are going for Globalfoundries 7LP (Leading-Performance) process in 2/H2018. Refresh my memory does 7LP really mean anything regarding performance?

I'm not sure if that will be limited to just Navi though. I wonder if they will test drive 7LP with Vega and Polaris?
I'm pretty sure those are meant to describe all 3 architectures and you're not supposed to try to tie any specific one of them into any specific architecture
 
I'm pretty sure those are meant to describe all 3 architectures and you're not supposed to try to tie any specific one of them into any specific architecture
I can agree with that, thanks!
What specific meaning does 7LP have if used with Navi? And if 7LP was used with Polaris or Vega what that provide an inherent benefit over TSMC 7nm process?
I did a quick search and found this article.
http://techreport.com/news/32110/globalfoundries-fires-up-its-7-nm-leading-performance-forges
In part:
40% performance improvement in the same power envelope or a whopping 60% power reduction within the same frequency range.
 
Last edited:
Yes. GF might be worse than TSMC which is why the higher clocks (only slightly, compared with Polaris) are barely noticeable in the end product and power is such a nightmare...

I think this is a large part of it, both vega and ryzen have this amazing voltage walls where things just take off. You saw the same with Apple on lpe, if vega managed another ~15% clock it would be a very differently received product. Also while we haven't seen much in the way of analysis on it even GP107 can't clock as high as GP104 while being significantly cooler/easier to cool.

But keep the voltage low and 14nm LPP does really well, just look at Raven Ridge, in the HP laptop its able to sustain in long running benchmarks that 2.4ghz all thread base and thus beats intels latest and greatest because it has a higher base for a given TDP. But then on the shorter benchmarks where boosting far over TDP comes into play intel wins quite comfortably because they can clock higher.
 
You saw the same with Apple on lpe, if vega managed another ~15% clock it would be a very differently received product.
Yes, I think Vega at TSMC might just have done that. But hey, GF wafers are "heavily subsidised" by the WSA.

Which makes me wonder, will Polaris for Intel be built at TSMC? Is that why AMD paid so much to change the WSA in a way that meant making GPUs at TSMC wouldn't obliterate their margins? Could GF keep up with the number of wafers AMD would need for that? Anyway, that's for another thread.

Also while we haven't seen much in the way of analysis on it even GP107 can't clock as high as GP104 while being significantly cooler/easier to cool.
Notebook version's base clock is higher than the desktop version's boost clock according to:

https://en.wikipedia.org/wiki/GeForce_10_series
 
Yes, I think Vega at TSMC might just have done that. But hey, GF wafers are "heavily subsidised" by the WSA.

Which makes me wonder, will Polaris for Intel be built at TSMC? Is that why AMD paid so much to change the WSA in a way that meant making GPUs at TSMC wouldn't obliterate their margins? Could GF keep up with the number of wafers AMD would need for that? Anyway, that's for another thread.


Notebook version's base clock is higher than the desktop version's boost clock according to:

https://en.wikipedia.org/wiki/GeForce_10_series

Or by Intel?
 
Which makes me wonder, will Polaris for Intel be built at TSMC? Is that why AMD paid so much to change the WSA in a way that meant making GPUs at TSMC wouldn't obliterate their margins? Could GF keep up with the number of wafers AMD would need for that? Anyway, that's for another thread.
If GF could keep up, mining should have had little effect on prices. That boom went on long enough surely one of the big banks could have special ordered a batch or two at higher cost.

If that deal was for Intel alone they must be needing an awful lot of chips. AMD financials listed 100m payment, 75m share stock warrant, and a per chip fee. That would necessitate billions in revenue without absolutely destroying margins. AMD specifically referenced a one time 335m charge for it.

Or by Intel?
While possible, that would mean Intel providing access to their fab tech to design the chips. The point of integrating dGPU would be that they could be sourced from almost anywhere. As it stands the semi custom for Intel is likely just a headless Polaris or Vega with specified CU count. Along with modified packaging for the bridge. It doesn't seem likely there were any console style tweaks like checkerboarding or ID buffers.
 
The GPU seems to be too big for standard Polaris. Compared to the HBM2 chip (7.75×11.87 mm) it seems to measure 16.61×13.14 mm, so around 218 mm². Published specs are 1536 SPs and 1024bit HBM2 interface. Polaris 10 with 2304 SPs and much larger 256bit GDDR5 interface has 238 mm². The GPU is only 20 mm² smaller than Polaris 10.

According to the dieshot of Polaris 10, removing 12 CUs (768 SPs), removing 256bit interface and adding 1024bit HBM2 interface (Vega 10 dieshot) should result in much smaller die size, around 183 mm². But in fact it's about as big as full Polaris 10 (2304 SPs) with GDDR5 interface replaced by HBM2 interface. It doesn't make sense. Why would AMD add 50 % more (spare?) CUs in a custom GPU?

On the other hand, 218 mm² is almost half of Vega 10. 1792-2048 SPs, Vega architecture and 1024bit HBM2 interface would correspond to the actual size of the GPU. I wouldn't be surprised if the GPU is in fact partially disabled Vega 11.
 
If GF could keep up, mining should have had little effect on prices.
AMD took huge inventory writedown hits previously (Hawaii was a disaster though some of the blame could be laid squarely upon the stupid reference cooler being 100% perfect negative marketing) and that was despite previous mining booms. The 2010/11 mining boom was purely an AMD boom (NVidia was a complete waste of time after the initial proof of concept) the like of which hasn't been seen since.

Each mining boom is over too quickly to "respond". From wafer order to card in consumer hand is about 3 months. Responding is just going to result in stock that needs to be written off. Especially when the mining cards are sold second hand after they've done their job.

It might be argued that the WSA (which didn't apply to Hawaii) makes inventory writedown less of a problem. Soon I'll be accused of leading the WSA cult, where nothing AMD does can result in losing more money, because AMD's already losing so much money on WSA.

That boom went on long enough surely one of the big banks could have special ordered a batch or two at higher cost.
The big mining farms we hear about (10s of thousands of GPUs) amount to a few hundred wafers of a given GPU in total. The effect of mining is mostly disruption to the conventional retail channel, not sheer volume.

To be fair this new shitcoin era is subtly different from previous mining booms, because there's effectively an on-going guarantee of there being a profitable coin to mine, for years. Because each week there's a new coin or 20. We're well past the craze for ETH mining that happened this year (which was a repeat of last year's ETH craze).

If that deal was for Intel alone they must be needing an awful lot of chips. AMD financials listed 100m payment, 75m share stock warrant, and a per chip fee. That would necessitate billions in revenue without absolutely destroying margins. AMD specifically referenced a one time 335m charge for it.
This WSA lasts four years, so some of that money is very much forward looking.

This appears to be the most astute analysis that I've come across

https://wccftech.com/analysis-amd-multi-year-amendment-2016-wafer-supply-agreement/

Designed to confuse onlookers, of course. I think we have to assume that AMD knew about the deal with Intel when drawing up this agreement, and GF didn't. Bear in mind that the WSA had been on a yearly rewrite cadence until 2016, so changing to a four year duration should have been a big clue to GF that AMD was up to something. GF would have already known how good Zen is...

While possible, that would mean Intel providing access to their fab tech to design the chips.
Intel has done this before.

The point of integrating dGPU would be that they could be sourced from almost anywhere.
That's not correct. The integration is because Intel doesn't have the IP to build a competitive chip. If it wants this kind of GPU performance in this form factor it has no choice about this kind of deal. The fab stamping the silicon is just a detail after that.

As it stands the semi custom for Intel is likely just a headless Polaris or Vega with specified CU count. Along with modified packaging for the bridge. It doesn't seem likely there were any console style tweaks like checkerboarding or ID buffers.
Yes it's unlikely Intel wanted anything novel. But don't forget Polaris is not full fat D3D12.x/SM6.x so it's likely Vega stuff was retrofitted for those function points.

It's fun to contemplate Intel building AMD GPUs, but I agree it's unlikely.
 
Each mining boom is over too quickly to "respond". From wafer order to card in consumer hand is about 3 months. Responding is just going to result in stock that needs to be written off. Especially when the mining cards are sold second hand after they've done their job.
This boom has been going on half a year and only starting to slow down though. For a financial sourcing cards directly at a premium, surely some would have approached AMD directly to rush a batch at some additional cost. Would have been more expedient than fighting for retail volume. Even then Polaris would have been relatively early in its life cycle. Getting stuck with product should have been less of a risk. Unload product before miners offload into the second hand market.

That's not correct. The integration is because Intel doesn't have the IP to build a competitive chip. If it wants this kind of GPU performance in this form factor it has no choice about this kind of deal. The fab stamping the silicon is just a detail after that.
I'll agree IP is the big concern, however if Intel we're sharing fab tech and wanting to integrate graphics, they could have done so with a giant monolithic die. No need for the bridge beyond HBM2 stack. Should be less power, better performance from fab tech, and cost wise not all that different from current chips. The trade-offs are interesting to consider.
 
This boom has been going on half a year and only starting to slow down though. For a financial sourcing cards directly at a premium, surely some would have approached AMD directly to rush a batch at some additional cost. Would have been more expedient than fighting for retail volume. Even then Polaris would have been relatively early in its life cycle. Getting stuck with product should have been less of a risk. Unload product before miners offload into the second hand market.
What you're missing is that there are millions of kids out there who have worked out that a graphics card pays for itself in a few months (shorter if living with paroids). Soon these kids are buying 2, 6, 20 cards. The big miners (people who negotiate electricity contracts, or sort out a dedicated renewable power source) are nothing compared to the kids.

I'll agree IP is the big concern, however if Intel we're sharing fab tech and wanting to integrate graphics, they could have done so with a giant monolithic die. No need for the bridge beyond HBM2 stack. Should be less power, better performance from fab tech, and cost wise not all that different from current chips. The trade-offs are interesting to consider.
Maybe that's version 2... Walk before you run.

Anyway I'm not arguing Intel is fabbing these...
 
I've got bored waiting for AiO cooling kit for Vega 56/64 so decided to go ahead with Morpheus II air cooler mod.

Album from the modding process:

Results - amazing!
I can't hear my PC when mining :)
As displayed on one of SS, mining temps are well under 50C GPU and 64C HBM2 with silent fans. Stock card had to run 2700-2900 RPM to keep HBM2 just below 85C throttling point and GPU ran at 66C-69C.

1600MHz real core clock in Superposition 1080 Extreme maintained thorough the test and only at the end HotSpot temps were touching 104C with fans already 100% (1660RPM) but still very quiet.
Stock cooler even at 4400RPM would not keep up and hotspot temp would force cards clock down to 1200MHz area.

Anyone waiting for custom models, this is what they should do out of the box ;)
 
Last edited:
I've got bored waiting for AiO cooling kit for Vega 56/64 so decided to go ahead with Morpheus II air cooler mod.

Album from the modding process:

Results - amazing!
I can't hear my PC when mining :)
As displayed on one of SS, mining temps are well under 50C GPU and 64C HBM2 with silent fans. Stock card had to run 2700-2900 RPM to keep HBM2 just below 85C throttling point and GPU ran at 66C-69C.

1600MHz real core clock in Superposition 1080 Extreme maintained thorough the test and only at the end HotSpot temps were touching 104C with fans already 100% (1660RPM) but still very quiet.
Stock cooler even at 4400RPM would not keep up and hotspot temp would force cards clock down to 1200MHz area.

Anyone waiting for custom models, this is what they should do out of the box ;)
Just for reference what paste did you use and how much did you use?
Usually needs to be a fair bit more than one would use on CPU, sorry if you know but worth mentioning it if others try.
Thanks
 
@Lightman
It's not pretty, but one can't argue with the results... How many slots does that Frankenstein's monster block, tho? Four? :p

Three cheers for you! Pretty epic mod, I have to say. Very brave, I would get weak in the knees attempting something like this.
 
Just for reference what paste did you use and how much did you use?
Usually needs to be a fair bit more than one would use on CPU, sorry if you know but worth mentioning it if others try.
Thanks

I've used Noctua NT-H1 and I did 5 dips on the GPU die and 1 on each HBM2 module. What you see on the pics is before cleaning old stuff.

@Lightman
It's not pretty, but one can't argue with the results... How many slots does that Frankenstein's monster block, tho? Four? :p

Three cheers for you! Pretty epic mod, I have to say. Very brave, I would get weak in the knees attempting something like this.

Cheers!
Pretty it is not, that's why I was waiting so long for Alphacool Euswolf 240 GPX Pro AMD RX Vega M01 kit to do it nicely, but gave up 2 weeks ago and bought Morpheus II. Took me two weeks to find time to do it and now that I've checked Aplhacool website, the kit is shipping!! My luck :D

Anyway, not regretting it at all as performance is there, and my case is not the one to look at, not yet at least, as I want to redo my setup with nice WC and RGB stuff, but that's project for another year. Currently it is an old monster BIG Tower ATX case from 2007 which sits under the desk and only bit of it sticks out and is visible.
Modding was a bit scary, but living with VEGA stock cooler made me take the risk :D. Gaming was fine, but mining while working not so much!

BTW this monster takes 3 slots with slim fans or 3.5 with my fat fans.
 
I've used Noctua NT-H1 and I did 5 dips on the GPU die and 1 on each HBM2 module. What you see on the pics is before cleaning old stuff.



Cheers!
Pretty it is not, that's why I was waiting so long for Alphacool Euswolf 240 GPX Pro AMD RX Vega M01 kit to do it nicely, but gave up 2 weeks ago and bought Morpheus II. Took me two weeks to find time to do it and now that I've checked Aplhacool website, the kit is shipping!! My luck :D

Anyway, not regretting it at all as performance is there, and my case is not the one to look at, not yet at least, as I want to redo my setup with nice WC and RGB stuff, but that's project for another year. Currently it is an old monster BIG Tower ATX case from 2007 which sits under the desk and only bit of it sticks out and is visible.
Modding was a bit scary, but living with VEGA stock cooler made me take the risk :D. Gaming was fine, but mining while working not so much!

BTW this monster takes 3 slots with slim fans or 3.5 with my fat fans.

When reading Alphacool Eiswolf 240 GPX Pro AMD RX Vega M01, I thought you were making things up to poke fun at hardware manufacturers that give ridiculous names to their products. I shudder at the thought of what they might come up with next.
 
Back
Top