The AMD Execution Thread [2007 - 2017]

Status
Not open for further replies.
Has this slide been made public before?

LNMMh8c.png
 
I understand why Intel keeps AMD around, but it behooves me why NVIDIA hasn't gone in for the kill in the last couple years (one could argue that they have, but they could have priced AMD completely out of the market with Maxwell pretty easily and never did). Perhaps they are more concerned with antitrust than I thought. Now if I were in charge at NV I would understand the importance of having at least 1 competitor in the market, but from what I've heard the good folks at NVIDIA could care less if AMD dies in a fire.
 
I understand why Intel keeps AMD around, but it behooves me why NVIDIA hasn't gone in for the kill in the last couple years (one could argue that they have, but they could have priced AMD completely out of the market with Maxwell pretty easily and never did). Perhaps they are more concerned with antitrust than I thought. Now if I were in charge at NV I would understand the importance of having at least 1 competitor in the market, but from what I've heard the good folks at NVIDIA could care less if AMD dies in a fire.

Stuff like this amazes me, its "economics 101"

1. In modern world current fin year shareholder return/value is the most important thing to maintaining a share price
2. lots of people in a company especially the higher you go have a large amount of shares in said company
3. A big decrease in operating margins, revenue and gross profit will have an effect on shareprice and create a pain in the arse of the exco to explain to shareholders.
4. If they wish to maintain gross profit then thats cuts to opex and capex ( driver devs, hardware devs, new facilities, new equipment etc)
4. Consumer GPU is bigger then everything else put together for NV, having that go backwards in that area again is a pain in the arse, they could play funny buggers with numbers to shift costs into other units but then if they dont perform they get extra pressure and scrutiny from shareholders etc.

So why exactly is someone who makes high level decision for a company going to put themselves through all that, whats the gain?
 
To be the only provider of discreet GPUs?
It seems they've determined it just isn't worth the bother. Besides, AMD is slowly killing itself anyway; NV can just continue business as usual, overcharging for their hardware, and it won't make any real difference.

Yesterday I saw an editorial on Nordic Hardware that claimed the GF1080 is outselling the RX480 10:1 in Sweden, despite costing twice as much. AMD really, REALLY needs a big break - or more likely, a strong buyout partner...
 
It seems they've determined it just isn't worth the bother. Besides, AMD is slowly killing itself anyway; NV can just continue business as usual, overcharging for their hardware, and it won't make any real difference.

Yesterday I saw an editorial on Nordic Hardware that claimed the GF1080 is outselling the RX480 10:1 in Sweden, despite costing twice as much. AMD really, REALLY needs a big break - or more likely, a strong buyout partner...
Yeah, I guess there's no point in killing off your competition when they seem perfectly capable of suiciding while you continue laughing to the bank. I just wish AMD was competent enough to pose at least a minor threat.
 
It seems they've determined it just isn't worth the bother. Besides, AMD is slowly killing itself anyway; NV can just continue business as usual, overcharging for their hardware, and it won't make any real difference.

Yesterday I saw an editorial on Nordic Hardware that claimed the GF1080 is outselling the RX480 10:1 in Sweden, despite costing twice as much. AMD really, REALLY needs a big break - or more likely, a strong buyout partner...


The only way AMD would get a break is if nV messes up, don't see that happening at least not this round. Just have to see what happens with Navi and Volta. AMD might be able to do something with Vega if they use HBM2 and nV uses GDDR5x with the gp102.......
 
AMD might be able to do something with Vega if they use HBM2 and nV uses GDDR5x with the gp102.......
If I was to speculate, based purely on gut feeling, I'd say that Vega with HBM2 won't be out until late 1st, early 2nd Q next year. But surprise me, AMD! :) If it is decently within GP102 performance-wise, I'm sure I'll buy one because it will no doubt be hellishly better value from a price/performance standpoint. Possibly even buying two, if they are tiny and come with water coolers like the Fury X.
 
To be the only provider of discreet GPUs?
That is not a gain, it has no hard numbers and you dont get magic extra money the day AMD goes bust, what you would have is probably a very slow cycle of increasing margins slowly to get back to where they are now. Revenue Now always wins.
 
Do we know what Vega is? I mean, do we know whether it's just Polaris+HBM or a new micro-architecture?
We don't know, but rumors say it's first with GFX IP 9.x. Rumors started from LinkedIn around same time the 232mm^2 Polaris leak came, and as we later found out, at least the 232mm^2 leak was real.
 
mmmh ?
missed that rumor.
Is it supposedly a completely different arc from GCN or still an evolution ?
Complete D3D12 feature level support would be nice ;)
 
We don't know, but rumors say it's first with GFX IP 9.x. Rumors started from LinkedIn around same time the 232mm^2 Polaris leak came, and as we later found out, at least the 232mm^2 leak was real.


I still have my doubts, I don't see it being that much further away from what they have now. What would be the use of a next gen next gen lol. Simple put, why would AMD try a next gen big part, and not use the same IP for the small part? Usually with new processes, and what not they would do it with the small part first.... As if anything goes wrong the need for a sustained lower power and higher performance would be a risk.

The IP blocks the only thing I can think of is they are using semi custom blocks form TSMC, but I don't think its going to be that much different from an architectural point of view.
 
I still have my doubts, I don't see it being that much further away from what they have now. What would be the use of a next gen next gen lol. Simple put, why would AMD try a next gen big part, and not use the same IP for the small part? Usually with new processes, and what not they would do it with the small part first.... As if anything goes wrong the need for a sustained lower power and higher performance would be a risk.

The IP blocks the only thing I can think of is they are using semi custom blocks form TSMC, but I don't think its going to be that much different from an architectural point of view.
mmmh ?
missed that rumor.
Is it supposedly a completely different arc from GCN or still an evolution ?
Complete D3D12 feature level support would be nice ;)
Forgot to update here, it's now more than just a rumour, AMD's OpenCL drivers have already support for Greenland (HPC APU), Vega 10, Vega 11 and Raven1X (Zen APU), and confirm they're GFX IP 9.x, while Tonga, Fiji and Polaris are 8.x

Going from past "big number changes", at minimum it means ISA changes
 
Have we actually seen any details on Polaris hardware beyond it has shaders inside? There are a lot of details that still seem to be missing for one reason or another. At this point I'm wondering if Polaris was largely just a node shrink with some of the new features (prefetch) software based and backported to fiji/tonga.
 
Have we actually seen any details on Polaris hardware beyond it has shaders inside? There are a lot of details that still seem to be missing for one reason or another. At this point I'm wondering if Polaris was largely just a node shrink with some of the new features (prefetch) software based and backported to fiji/tonga.

Well Tonga/Fiji, the only difference really from Hawaii, is the programmability of their HWS and color compression, if I'm not mistaken. So the IP numbers really don't tell us much. So given the time necessary to really revamp their architecture to lower power consumption and increase frequency with lower voltages. I just don't see that happening with Vega, just not enough time to make those kind of fundamental changes.
 
Well Tonga/Fiji, the only difference really from Hawaii, is the programmability of their HWS and color compression, if I'm not mistaken. So the IP numbers really don't tell us much. So given the time necessary to really revamp their architecture to lower power consumption and increase frequency with lower voltages. I just don't see that happening with Vega, just not enough time to make those kind of fundamental changes.

There were ISA changes, particularly necessary with the addition of scalar memory writes. Internal tweaks to items like the geometry engines and the introduction of power-saving features like adaptive clocking for the GCN3 generation.
However, I am seeing claims that while AMD did disclose some of those power management features, they weren't enabled until Polaris. It does seem plausible, since AMD has repeatedly announced physical architectural features and then quietly failing to get them working starting from Jaguar and onward.

AMD has also changed the criteria it uses for IP numbering and moved some items that used to be externally visible out of the driver-level tables. This means AMD has gone from some externally unknown methodology to another externally unknown one.

It's possible Polaris has refinements or additional power management features that it has failed to make workable. Possibly, if the mobile focus of Polaris is something AMD actually meant, some items might have been less useful at the clocks the desktop GPU was forced to.

One item I'm curious about is a graph that was made of AMD's voltage/clock curve for an ASIC of average ASIC quality.
For all but the >1 GHz range, Polaris appears to sit at roughly .8V, regardless of how low clocks get.
There would be a nominal voltage where further drops would lead to more leakage, something that leads to relying on clock throttling or even intra-frame duty cycling to do more power gating.
 
Have we actually seen any details on Polaris hardware beyond it has shaders inside? There are a lot of details that still seem to be missing for one reason or another. At this point I'm wondering if Polaris was largely just a node shrink with some of the new features (prefetch) software based and backported to fiji/tonga.
The GCN4 ISA is unchanged from GCN3. However it's not a pure shrink, as there are new things on Polaris like the larger instruction buffer within the CUs themselves, and of course new features outside of the CUs like updated color compression and the primitive discard accelerator.
 
Status
Not open for further replies.
Back
Top