Why is AMD losing the next gen race to Nvidia?

ok itsmydamnation, I didn't take the first time you called me a troll personally, as I ignored it, since I find the topic interesting.

In the perfect world yeah they would transition into hyper scale models, but in the real world, it takes time. If at NBC we had hyperscale ready for SB 2012, we wouldn't have had the our network go down. But we see the benefits of doing it, but the red tape, corporate structure, many reasons, and what on that we use just won't work on hyperscale systems yet.

For company's like google, amazon, facebook, they are going hyperscale, but they are burning hundreds millions of dollars doing it, but they are making a business out of it, most companies aren't looked to do that, so they will not be willing to invest that much money in the short term, as they start transitioning off into hyperscale system in future, its a slow growth thing.
 
You need to look at chef, puppet, answible, Azure services bus etc. People aren't spending anywhere near the kind of money your thinking of to do this. The standard web application hierarchic ( presentation, application, db, security/auth) already allows for significant scale. I was talking to an enterprise architect today of a small 400 seat org who was all excited by Azures new tempting abilities he can now spin up an entire environment from a single line of power shell. This isn't future world leading edge stuff, this is now, if your company thinks its leading edge stuff then chances are they are behind their competitors.

I called you a troll because AMD's potential competitive advantages have been spelt out night and day, yet you cant acknowledge anything, sorry but that is a troll. I never pretended that Zen will be everything to everyone, its doesn't have to be, there is a silent majority of server workloads and that is what i was demonstrating to you. But you are saying with certainty that Zen will be nothing to everyone just because of what looks to be like a small IPC deficit.
 
Last edited:
You need to look at chef, puppet, answible, Azure services bus etc. People aren't spending anywhere near the kind of money your thinking of to do this. The standard web application hierarchic ( presentation, application, db, security/auth) already allows for significant scale. I was talking to an enterprise architect today of a small 400 seat org who was all excited by Azures new tempting abilities he can now spin up an entire environment from a single line of power shell. This isn't future world leading edge stuff, this is now, if your company thinks its leading edge stuff then chances are they are behind their competitors.

I called you a troll because AMD's potential competitive advantages have been spelt out night and day, yet you cant acknowledge anything, sorry but that is a troll. I never pretended that Zen will be everything to everyone, its doesn't have to be, there is a silent majority of server workloads and that is what i was demonstrating to you. But you are saying with certainty that Zen will be nothing to everyone just because of what looks to be like a small IPC deficit.

It might not be a "small IPC deficit only" we don't know where Zen will end up, in all likely hood its going to be a BIG IPC deficit, compared to Sky Lake Xeons. Even if they were equal in IPC and over all performance in the various tasks, once a company like Intel has the entire market, AMD is going to have to lift mountains to get in. Price/performance is not the object anymore as the scales of economy will be all on Intel side. AMD needs to balance that out first before they can even think of pushing forward.

I used the battle analogy because it was best suited for this. Why do you think business schools make you read sun tzu's art of war or the book of five rings?

Intel's greatest strength isn't their processors, its the their size and dominance of the market, processors are like the individual soldiers on the battle field. But their size is like the US armed forces. And if US wants to go into to Iraq they will, like they did. If Intel wants to go at AMD they will, it doesn't work too well the other way around though, AMD will have to pick and choose the left overs and then strategically take small chucks out of Intel, which is not going to start for a long time, because businesses although want competition on the processor side, they can't move that fast and they also have to see what is best for them.

Companies like what you just mentioned are doing this for a business, they aren't going to go over to AMD unless AMD has some sort of advantage over Intel that can be used to satisfy a need. I don't have any ideas where Intel can't match AMD for what AMD has shown so far.

Also this takes time too, because their clients have to also be aware of that need and are willing to monetize on that need as well otherwise that need won't be used.

Further look at this, when AMD started losing ground against nV in graphics, look at what they did with their business model, they switched over to consoles, they down sized their desktop capabilities for high end and are refocusing their primary objectives in the short term. This is with them not falling too far behind like what happened in the CPU side. Their CPU division will have to do these kind of tactics and more so because its much further behind right now. They need to find niches where Intel can't compete with them, and the server market is not where its going to be. The needs of the server market as a service to other companies are fairly grounded.
 
Last edited:
I used the battle analogy because it was best suited for this. Why do you think business schools make you read sun tzu's art of war or the book of five rings?

If business is specified as war, then the fundamental tactic to strip the enemy of it's resources (namely 'soldiers' or 'morale'), leads to the same devastating consequences as 'real' war: suffering of the people. So, this entire idea justifies to destroy the life-lines of people who want to realize their own idea on a wast market, just because of ... arbitrarily being defined as the enemy? Exorbitance?
There are lots of people, and lots of ideas, and lots of individualism, and lots of company, and lots of market on this planet. For which higher goal does any company need/want to monopolize creativity, innovation, individualism, prerogative of interpretation? Why does a company need to be the largest, the most powerful, the most successful? I ask, because by implication this makes every other expression (company) unsustainable, or an even more vicious alternative. And that's really really sad.
 
Intel's greatest strength isn't their processors, its the their size and dominance of the market, processors are like the individual soldiers on the battle field.
Processors are the product Intel are selling and if they are suddenly sub par and a competitor can beat them in price, performance and features the market will eventually shift.
Intel is not immune to this as history has proven.
 
yeah that is true, to take away the soldier then the war machine dies lol, that's simple. AMD is not in a position that we know of yet with Zen to do such a thing. Their soldier is Zen, and its going up against Sky Lake Xeons, so its possibly even footing when it comes to the product (soldiers) but this is TBD. What is left? Intel's capabilities as a company and AMD's capabilities as a company. Can we draw lines that AMD can match Intel in every aspect? I see many places AMD can't, easiest to see is resources. Intel can play a game of attrition and block AMD out if it needs to, even in niche server markets, if there are any that AMD can push forth, Intel will try its best to block. Both companies are using 14nm tech so silicon per chip should roughly be the same if die sizes are close and yields are close, but Intel might still have an advantage there as they have a full control over their fabs, they also have a possibly more mature 14nm process.

What are the advantages Zen will bring that Intel can't match in server markets? I can't think of a single place where Zen could be consider a de facto winner (not just performance, feature sets), even a marginally winner would be hard to say.

The rule of three in economics states in any free market, three competitors will arise as being the major competitors in that market, as they will eat up the smaller company shares and an equilibrium is reached. We have seen this in the GPU market and many other tech markets.

The CPU industry went through the same thing, but because of costs of innovation only two companies were able to sustain it in the x86 market as MS had virtual monopoly on the OS side which forced the hand on the CPU side.

As the CPU industry matured, and cost of innovation still rose, and AMD's misstep with architecture left only one company in the x86 market. AMD's refocused work on Zen to break a monopoly in the x86/x64 market is extremely hard. The rule of three actually states, its a loosing affair, better take your money and run.

Now in the rule of three any competitor that has less that 25% of the share of the largest company can't be a competitor. This is a bit on the nose, as they still can be in a niche. But when you look at the experience curve relationships it can be said they are not directly competing. This is where AMD is at right now and this is without the server market. In the server market they are no where.

So what AMD is looking at is a mature market that has been in a virtual monopoly for 10 years coming in with a product that is equal to what is on the market. To sustain something like this, they will need the same amount of resources accessible to them as Intel has or find a niche where they can grow and then directly compete.

The rule of three can be broken (well shifted) if a new competitor has a superior product. The market will recognize this and force the larger competitor's hand to either innovate or die. This is what happened in the past with AMD and the first Opteron in the server market and PC markets.
 
Last edited:
Then that is competing more against Knights Landing rather than traditional Xeons, and KNL has had a big improvement over previous generations.
I was wondering the same myself, but I am not sure how well AMD will do in that space against Intel or IBM-Nvidia/Cray-Nvidia, all with their own memory/interconnect/fabric solutions.
It can, but it doesn't have to. A dual socket board could be two CPUs or a CPU+GPU. Maybe dual GPU, but Vega would be really interesting if that happened. Point being it's the same platform with both options. HPC Zen might have some advantages there and AMD still seems to have higher theoreticals for likely parts. A fury for example already had more compute/bandwidth than KNL as I recall. There is also the rasterization capability if your server actually has to accelerate graphics tasks. IBM/Nvidia are also RISC based as opposed to x86 which could be significant. AMD also has the HSA software stack they were working on along with a bunch of the mobile companies. VMs servicing mobile/tablet clients could have some different needs.

Intel can play a game of attrition and block AMD
Is that why they let go up to 15,000 employees as they refocused from PC to data/mobile markets?

What are the advantages Zen will bring that Intel can't match in server markets? I can't think of a single place where Zen could be consider a de facto winner (not just performance, feature sets), even a marginally winner would be hard to say.
Graphics platforms. They still have stronger graphics performance than Intel and HPC Zen's on the same socket could manage there. If the intent was to render a desktop or some 3D scene and stream it to a mobile from some sort of VM that might be ideal. Unlike Nvidia they still had hardware VMs on the GPUs as opposed to software. Hurray for those programmable ACEs again. Sure they could add GPUs to the system, but the video memory requirements of 100 or more clients on a GPU could be interesting as opposed to working out of system memory with a high speed link. I don't think their advantage is a straight up CPU but a more integrated or versatile processor platform likely involving graphics.
 
Is that why they let go up to 15,000 employees as they refocused from PC to data/mobile markets?

Adapting to the market is a very important aspect of survivability, the rule of three even state this too, once the three main companies become so large that they can't foresee the changes in the market, they can't survive against the others in the market and new market players arise.

Graphics platforms. They still have stronger graphics performance than Intel and HPC Zen's on the same socket could manage there. If the intent was to render a desktop or some 3D scene and stream it to a mobile from some sort of VM that might be ideal. Unlike Nvidia they still had hardware VMs on the GPUs as opposed to software. Hurray for those programmable ACEs again. Sure they could add GPUs to the system, but the video memory requirements of 100 or more clients on a GPU could be interesting as opposed to working out of system memory with a high speed link. I don't think their advantage is a straight up CPU but a more integrated or versatile processor platform likely involving graphics.

Err no doesn't work that way, the need for something like that isn't there, AMD will need to drive the software to create that need. They have trouble just driving games which are made by others.....

If they create that software, they create that need, and then they create that niche.
 
Err no doesn't work that way, the need for something like that isn't there, AMD will need to drive the software to create that need. They have trouble just driving games which are made by others.....
Both AMD and Intel have been plowing away on open source linux drivers. It seems likely they're chasing something. Now either linux gaming is about to become huge, or something is being worked on. I recall reading something about it the other day, just can't remember what it was atm. Had to do with virtualized graphics on linux platforms. Accelerating Citrix, Android, or something. Ironically asynchronous shading is also something that occurs when an abundance of VMs on a server try to access the same GPU.
 
Still need the software..... its not about their products its about products that run on their products.
Even Android runs a linux kernel. There are plenty of products out there that run Android. Google is even in the process of spinning their own OS which could be similar, although I thought it was supposed to be media focused. It could be something as practical as a VM platform to accelerate augmented reality or something. VMs with strong graphics capacity for accelerating AR would definitely be an upcoming market. A market that I could easily see outpacing the typical server demand.
 
That is pretty much what nV's grid is all about, so that won't be a niche unless AMD wants to do other things outside of games, which means software licensing models, etc. Not easy to do. Not to mention an entire server setup and maintenance and upgrading based on the software. Resources ;)
 
That is pretty much what nV's grid is all about, so that won't be a niche unless AMD wants to do other things outside of games, which means software licensing models, etc. Not easy to do. Not to mention an entire server setup and maintenance and upgrading based on the software. Resources ;)
Software licensing of what? They already have the hardware that supports everything. VM software already exists. I'm not aware of any major difficulties firing up a hardware VM on an AMD CPU or executing a program on a GPU. The only issue I recall had to do with resetting the device with multiple VMs present and that may have been addressed already. They already have GPUs that expose multiple virtual interfaces in hardware. The only issue would seem to be getting a program to run on a VM that uses a GPU, and just about all modern software is capable of that. As for NV's Grid, do they have in excess of 500GB of memory on their GPUs yet? Or NVLink working with x86 processors from AMD or Intel?
 
the software and tools that you will be running on the vm, ya know someone has to buy it, you think all those games on nv grid are for free? I'm sure nV has to pay something to the developers/publishers too.

What programs in VM need excess of 500gb of memory dedicated to graphics?, just asking cause lets say I'm using 3dsmax or maya and my model is around 100million polys with textures and the whole shebang, err 64 gigs ram for my system is enough and a 12gb video card is all I need.
 
Software licensing of what? They already have the hardware that supports everything. VM software already exists. I'm not aware of any major difficulties firing up a hardware VM on an AMD CPU or executing a program on a GPU. The only issue I recall had to do with resetting the device with multiple VMs present and that may have been addressed already. They already have GPUs that expose multiple virtual interfaces in hardware. The only issue would seem to be getting a program to run on a VM that uses a GPU, and just about all modern software is capable of that. As for NV's Grid, do they have in excess of 500GB of memory on their GPUs yet? Or NVLink working with x86 processors from AMD or Intel?
What you don't understand or maybe you don't want to see, is that hardware is very small part of the equation, especially on VM market. You need to create tools to manage your hardware resource (supervisor, scheduler, application profiles etc), provide technical support (ie man power) to the VM vendors, build commercial partnership with the market leaders, in brief provide a full solution that can be deployed in real world, not just PowerPoint slides...
During the last 3 years, Nvidia has built strong presence in the VM market, with Microsoft, Citrix, VMware and Nice on software side and all major player on hardware server side. (long) List of GRID suppliers with links to the manufacturer websites:
http://www.nvidia.com/object/grid-certified-servers.html

Nvidia even offers free trial of GRID in their own infrastructure:
http://www.nvidia.com/object/vmware-trygrid.html

All in one, Nvidia has created a win-win cooperation with all the market leaders, built on top of a solid ecosystem. Its a very fast growing income as we can see with each quarterly report.

AMD ? 6 months ago, their virtualization section on their website was a single page. yes a single page with broken links. Fortunately, they have now a partnership with Citrix but its still in the infancy. They are so much behind that it will be difficult to catch up as Nvidia is in full throttle in this market, with so much resources (ie money) invested...
 
the software and tools that you will be running on the vm, ya know someone has to buy it, you think all those games on nv grid are for free? I'm sure nV has to pay something to the developers/publishers too.
So people need to buy software to run on their VMs... what's new? It's not like programs need specially adapted... a machine is a machine. I didn't have any trouble getting Win7 running on my Linux box simultaneously with it's own GPU. The VM even came with a nice app to configure everything. I'd expect the company making the VM to have some better enterprise tools though. I'm sure there could be some uses in partitioning a GPU, but why not just throw the entire thing at any VM attempting to use it?

What programs in VM need excess of 500gb of memory dedicated to graphics?, just asking cause lets say I'm using 3dsmax or maya and my model is around 100million polys with textures and the whole shebang, err 64 gigs ram for my system is enough and a 12gb video card is all I need.
What about 128 instances of your model? Can you fit 128 or more on your card all at once? What about 256? It's not about a single program needing that much but a bunch of them in aggregate. If a single machine is hosting 500 VMs just the desktops will add up. We're going for density here with many of the VMs likely idle a substantial portion of the time.

What you don't understand or maybe you don't want to see, is that hardware is very small part of the equation, especially on VM market. You need to create tools to manage your hardware resource (supervisor, scheduler, application profiles etc), provide technical support (ie man power) to the VM vendors, build commercial partnership with the market leaders, in brief provide a full solution that can be deployed in real world, not just PowerPoint slides...
Making some software packages to configure settings doesn't seem like it would be too difficult. You're arguing the creation of tools that ultimately aren't necessary on a properly virtualized platform. Maybe you want to cap usage, but with all resources properly aggregated together it could be far simpler to just let them run FIFO. While not for every application, that can manage a ton of simple ones. By your own argument here Nvidia has a robust toolset and hardware, yet Citrix is partnering with AMD? Why even bother? Why make tools when you have hardware that can largely do it for you in a more secure manner with minimal overhead? Or is this going to turn into an argument that software based VMs are superior to hardware tagging methods implemented by AMD and Intel?

Both of you seem to be arguing requirements that are only required when virtualization is limited.
 
FWIW, KNL does not support virtualization out of the box.
source: http://ark.intel.com/products/94033/Intel-Xeon-Phi-Processor-7210-16GB-1_30-GHz-64-core

Is that why they let go up to 15,000 employees as they refocused from PC to data/mobile markets?
No, that's one of the idiocracies of public companies. Intels traditional markets are saturated. Hence, they need to find other field to foster growth and appease shareholders. That's necessary in order to avoid loss of stock market value and prevent a hostile buyout.


Did i read that article correctly: No Polaris-based solutions as of yet?
 
Last edited:
So people need to buy software to run on their VMs... what's new? It's not like programs need specially adapted... a machine is a machine. I didn't have any trouble getting Win7 running on my Linux box simultaneously with it's own GPU. The VM even came with a nice app to configure everything. I'd expect the company making the VM to have some better enterprise tools though. I'm sure there could be some uses in partitioning a GPU, but why not just throw the entire thing at any VM attempting to use it?


What about 128 instances of your model? Can you fit 128 or more on your card all at once? What about 256? It's not about a single program needing that much but a bunch of them in aggregate. If a single machine is hosting 500 VMs just the desktops will add up. We're going for density here with many of the VMs likely idle a substantial portion of the time.


You don't seem to understand what we are talking about, Dave's article pretty much sums it up.

The reality is that AMD does not exist in a vacuum, and NVIDIA GRID is still the market leader when it comes to features and market share. NVIDIA supports up to 16 users per GPU just like AMD does and does so while offering great performance. NVIDIA’s software overlay architecture also allows for more flexibility in provisioning which allows NVIDIA to provision any number of users from 1 to 16 while AMD can only provision 2, 4, 8 or 16 users. NVIDIA’s product line is also a bit more differentiated with an MXM model for blades, a two GPU board and a four GPU board with up to 32GB of GDDR5. Additionally, NVIDIA has more hypervisor support with support for ESXI, XenServer and KVM while AMD only supports ESXI. NVIDIA also has an easier time of configuration with profiles and allows for the management of VMs in console.

And AMD will not be in a niche, it already has a market leader that has over all better capabilities for the time being.

AMD has its work cut out for it. The only way for AMD to make a move is to be able to either make something entirely new where nV can't compete, or go into the market with just as much gusto as nV has done so far and battle it out, right now they are at the short end of the stick because they are way behind. Wasn't AMD that laughed at nV when nV stated they were now a software company? What is this took a few years to understand what nV was talking about? Well now they realize it and now they need to spend even more resources than nV has over the past few years just to play catch up.


Why the hell would I need 128 instances of that model?

I know what I need it for even in movies I would never need that for that model. You can't even animate such a dense model.

Also nV started this in 2009, that's 7 years ago.
 
Last edited:
Back
Top