Judge Allows Lawsuit Over AMD's FX Processors to Continue

[joeblowsmynose]

... trim ...

I feel like AMD been doing these "tempting" stunts that mislead many AMD fans.

"Hey look AMD Ryzen has more cores than Intel 8th Gen so we perform better" ~ Underperfroms
"Hey look AMD FX 8350 has 8 Cores" ~ Performs like 4 Cores/8 Threads
"Hey look we at 7nm" ~ Intel 14nm stills beat them

Well, they got caught.

Huh? What the heck are you on about ...


1) Ryzen at launch performed better than even Intel 8 core processors at heavily threaded workloads ... Intel didn't even have anything on a mainstream platform that could come even close -- it was obliterating Intel's paltry four cores - the best they had on a non HEDT platform ... obliterated the 7700k (Intels flagship mainstream CPU at the time) in anything over four threads. Like not even close ... Underperforms?

Did you mean it wasn't miles better than Intel at everything? Because for what I use my PC for, Intel had nothing but four cores -- sure it could run superpi better, if that's your thing ... If you are gaming, you have bottlenecked GPU anyway so there's less than 2-5% difference at max. 5% slower at lightly threaded tasks vs +200% faster at heavily threaded tasks. It took almost two years for Intel to finally have an 8 core in its mainstream platform ... before that, they had nothing to compete in threaded tasks unless you were willing to get ripped off and Pay $800 for a HEDT chip and spend $400 on a new mobo. If getting ripped off is your thing, by all means ...

2) The FX8350 did have 8 cores and it performed rather poorly, yes. Where's the contradiction? Intel couldn't make 64 bit either and claimed it was "impossible" to have both 32 and 64 bit on the same chip. AMD did it, and until recently, all Intel processors had AMD 64 bit licensed tech "inside". What does any of this mean? Nothing. lol. Just like your point.

3) "7nm ... Intel still beats them"? Huh? Where's these new 7nm CPUs that Intel is "still" beating - you do realize they've not been launched or reviewed right? The only one I saw was at the CES demo where it smacked down the Intel flagship 9900k using 2/3 the power doing it ... with latest rumours saying the AMD chip was sandbagged ...


You must be sniffing glue or something ...

A "Stunt" is like paying 3rd parties to publish fake gaming benchmark reviews while the real reviewers are gagged under NDA until after launch; or, demoing a "28 core 5ghz" processor then claiming it was a real product launching by end of 2018 ... then being forced to apologize for misleading people after being called out for it ... or something like claiming that your new RTX GPUs are 600% faster than last generation when reality is they are only 15% faster, but that's ok because you paid publications to promote the products before reviews ("Just buy it!") -- those are actual stunts where the bastards were "caught" in the act ...

 

[MasterMadBones]

I wonder if this lawsuit would have even been filed if the performance was any better.

The reason for Bulldozer's single-core performance deficit compared to Intel or even previous-generation AMD CPUs is not just the shared resources. It had a relatively narrow integer pipeline to begin with, and a 6-way instead of 4-way dispatch would've almost nullified the (small) scaling issues. If AMD had fixed these issues (along with some other minor flaws), Bulldozer could've been very close to Nehalem or even Sandy Bridge in integer performance while maintaining the same philosophy.

Floating point performance is a different story. Bulldozer's FPU was 128 bits wide and SMT-enabled to help improve multi-threaded performance, but only for single-precision and double-precision operations. See it as a kind of multithreaded SIMD implementation. This obviously meant that 128-bit instructions could only be performed one at a time and blocked any other floating point operations from being issued as well, and 256-bit AVX instructions, although they were supported, took more cycles to complete. The FPU itself would have offered decent to excellent performance for its time, had it not been hampered by the limited dispatch and branch prediction resources. Only a 512-bit wide FPU, which was monstrous for its time (and still is), would've been able to work on any two threads simultaneously. Even today, Intel CPUs spread their AVX-512 instructions over two cores.

From a performance point of view, had Bulldozer not had limited (but still shared) resources for its cores to work with, you could only criticize AMD for not making a separate FPU for each core. But, as has been mentioned a couple of times in this thread, if we say that a CPU core must have its own FPU, where does that leave the 80386, or indeed Coffee Lake when processing AVX-512 instructions? In fact, it's not too hard for me to imagine that, somewhere out there, there is a niche CPU architecture that has no integrated FPU. Arduinos don't have an FPU after all...

A little food for thought, unrelated to the actual topic here, what if AMD did actually make a 6-way dispatch unit for Bulldozer and made the integer pipelines wider? What would an SMT implementation look like? If each module could operate on 3 or 4 threads simultaneously, with the shared front-end, the CPU would have unprecedented flexibility in which resources to use for each thread. We might be looking at an SMT implementation that delivers up to 50% more performance when doubling the thread count as opposed to 30% for conventional SMT. Obviously, AMD has had some very good reasons not to do any of this. Such a chip would've been far too complex to design and too expensive to produce as well. And remember how much power the FX line used?

 

[InvalidError]

A little food for thought, unrelated to the actual topic here, what if AMD did actually make a 6-way dispatch unit for Bulldozer and made the integer pipelines wider?

Then the more complex execution resource management, instruction scheduling, data routing across the core, etc. would have increased propagation times and forced AMD to either add pipeline stages which increase latency while making instruction scheduling that much more complex (need to manage more resources through more pipeline stages and it takes more clock ticks for dependencies to get resolved) and inefficient to maintain clock frequency or reduce clock frequencies.

You can't just make the architecture arbitrarily deeper and wider. At some point, complexity kills clocks and net throughput goes down instead of up. Zen and Core are the products of AMD and Intel coming to similar conclusion after their Buildozer and Netburst disasters.

 

[MasterMadBones]

You can't just make the architecture arbitrarily deeper and wider. At some point, complexity kills clocks and net throughput goes down instead of up. Zen and Core are the products of AMD and Intel coming to similar conclusion after their Buildozer and Netburst disasters.

High IPC usually goes at the expense of high clock speeds, no surprise there. We simply don't know whether it would've been worth it from a performance standpoint. We know that AMD has always had a limited R&D budget, so Bulldozer could've been significantly better. In fact it did get better, just look at the massive improvements AMD has actually made up to Excavator (by splitting parts of the frontend again so you're probably right). Yes, that took several years, but even now AMD has admitted that Zen and Zen 2 are just part of the road to Zen 3 and beyond. They just hit the nail on the head this time (and they got lucky - the expected 40% IPC improvement wouldn't have made AMD much more significant than they were).

I'm not sure it's fair to compare Bulldozer to NetBurst. AMD actually willingly made the compromise of increasing core count at the expense of floating point performance in particular. They expected the market to shift from single-threaded CPU-bound workloads to multi-threaded GPU-bound workloads, which is where the APUs and AMD's investment in HSA comes in. Unfortunately for them, that didn't really happen. We're seeing some signs of it now, but only in data centers, where AMD was poorly represented. They more or less succeeded with the architecture, but the market for their CPUs was small and Sandy Bridge was the nail in Bulldozer's coffin.

NetBurst on the other hand, was an attempt to maximize clock rates at the expense of IPC. Intel hoped that forthcoming process nodes would help them achieve that, but the results were disappointing. As a result, Pentium 4 CPUs never reached the desired clock speeds, and barely improved over Pentium 3 while using more power and running hotter. Later revisions took performance beyond P3, but they ran even hotter.

 

[InvalidError]

I'm not sure it's fair to compare Bulldozer to NetBurst.

I think it is absolutely fair: Intel had the P3 which had a short pipeline and relatively high IPC, then traded IPC for higher clock frequencies on Netburst, then traded clock for IPC back with Core/Core2. AMD's Athlon had a relatively short pipeline, AMD lenghtened the pipeline achieve higher clocks with the dozers, shortened the pipeline at the expense of clock frequency with Zen.

A large chunk of Intel's IPC gains since first-gen Core comes from being able to make the execution back-end incrementally wider with more mature higher density processes. Zen was similarly made possible largely thanks to the step from 28nm that AMD had been stuck on for ages to 16nm.

 

[jimmysmitty]

I think it is absolutely fair: Intel had the P3 which had a short pipeline and relatively high IPC, then traded IPC for higher clock frequencies on Netburst, then traded clock for IPC back with Core/Core2. AMD's Athlon had a relatively short pipeline, AMD lenghtened the pipeline achieve higher clocks with the dozers, shortened the pipeline at the expense of clock frequency with Zen.

A large chunk of Intel's IPC gains since first-gen Core comes from being able to make the execution back-end incrementally wider with more mature higher density processes. Zen was similarly made possible largely thanks to the step from 28nm that AMD had been stuck on for ages to 16nm.

You mean 14nm?

 

[InvalidError]

You mean 14nm?

Yes, never remember exactly which fab got contracted to make what and which one calls that step of its fab progression 16nm instead of 14nm. Doesn't matter all that much given how vague the definitions have become and how much variation there is even within the same 'node' depending on process maturity and design priorities.

 

[jimmysmitty]

Yes, never remember exactly which fab got contracted to make what and which one calls that step of its fab progression 16nm instead of 14nm. Doesn't matter all that much given how vague the definitions have become and how much variation there is even within the same 'node' depending on process maturity and design priorities.

Agreed on that. Its not quite 14nm but thats the node they are using currently which helped them a lot since GloFlo basically dropped the ball. Big time.

 

[digitalgriffin]

Well they did mislead the consumer!

Well then you can add any Intel system to that list that invoked Hyper Threading via a scheduler.

All AMD was create a single FP unit between two integer cores. If they both needed the FP, then one of the pipes stalled.

Intels Hyper Threading works the same way. There's a fixed number of ALU's and partial stages in the pipeline that are divided out like resources. If a resource is occupied, the pipeline stage for the other thread stalls. Now Intel does admittingly create a lot of ALU's and other resources so this stall is minimal. HOWEVER it can still more than happen. So a cross lawsuit is more than fair.

 

[digitalgriffin]

I can't see them losing, and all this judge has said is it's not being thrown out... yet.

It's not like the inner-working were a trade secret. The marketing didn't fit the product accurately, but it's not like they purposely concealed the true nature of the shared resources.... That's what I think it'll boil down to.

However, if they were to lose, it's their own fault.
With no legal definition of a "core", it's likely to fall onto industry standard acceptance.
Prior to FX, there was never any ambiguity in marketing regarding core count. Even when HT was new, there were rumors/misunderstandings in the space about being double the "cores", but both Intel & AMD never marketed as such.
FX is where AMD decided to change the accepted definition to meet their marketing plans for an "8c" CPU.

I don't think a loss would be beneficial for the industry, all it would do is deplete AMDs current bank balance.
They're not typically in it to maximize profits, but want market share. If their hand was forced through payouts, they may need to increase the profit margin on new products to cover it.

To me a core is anything with a pre-fetch, scheduler, and pipeline, and independent ALU units. If you have two complete pipelines, you have two cores. Everything else is just resources that are shared between cores. AMD never lied and made if very clear the FP unit is shared between the two cores.

 

[digitalgriffin]

Agreed on that. Its not quite 14nm but thats the node they are using currently which helped them a lot since GloFlo basically dropped the ball. Big time.

They rate it by smallest feature, and not by things like fin pitch. Intel rates them by all features crucial to an IC. So even though I'm an AMD fan, Intels 14nm is far more advanced than TSMC's or GloFo's process. This is one reason AMD chips on a 12nm TSMC node can't scale as high as Intel's I series. That said, Intel better move their butt. The days of minor research investment for big profits ARE OVER. Get over it Intel. Other kids have balls to play with now. TSMC's 7nm will blow away anything Intel has now and Intel's 10nm will barely hold it at bay.

 

[jimmysmitty]

Well then you can add any Intel system to that list that invoked Hyper Threading via a scheduler.

All AMD was create a single FP unit between two integer cores. If they both needed the FP, then one of the pipes stalled.

Intels Hyper Threading works the same way. There's a fixed number of ALU's and partial stages in the pipeline that are divided out like resources. If a resource is occupied, the pipeline stage for the other thread stalls. Now Intel does admittingly create a lot of ALU's and other resources so this stall is minimal. HOWEVER it can still more than happen. So a cross lawsuit is more than fair.

Not quite. Intel specified that the processors were X Cores and Y threads. They differentiated them. Intel never claimed that a Pentium 4 or Core CPU with Hyperthreading were dual cores or 8 core CPUs.

AMD was claiming that the FX 8XXX series was 8 core CPUs and while they explained the ideas behind it their marketing was what stated they were 8 core CPUs. AMDs marketing has had issues before, so I am not blaming AMD as a whole as marketing teams can easily ruin a product with marketing.

To me a core is anything with a pre-fetch, scheduler, and pipeline, and independent ALU units. If you have two complete pipelines, you have two cores. Everything else is just resources that are shared between cores. AMD never lied and made if very clear the FP unit is shared between the two cores.

I don't disagree but there is a lot of FP work these days and it has become a integral part of the CPU. AMD even went back to a design similar to the uArch pre Bulldozer, K8.

They rate it by smallest feature, and not by things like fin pitch. Intel rates them by all features crucial to an IC. So even though I'm an AMD fan, Intels 14nm is far more advanced than TSMC's or GloFo's process. This is one reason AMD chips on a 12nm TSMC node can't scale as high as Intel's I series. That said, Intel better move their butt. The days of minor research investment for big profits ARE OVER. Get over it Intel. Other kids have balls to play with now. TSMC's 7nm will blow away anything Intel has now and Intel's 10nm will barely hold it at bay.

From what I have read, TSMCs 12nm is just a 14nm+. There was no change to increase density or shrink size.

And everything I have been reading still show Intels 10nm to be more dense than TSMCs 7nm, of course that is until they actually launch it and we can see it for ourselves.

Process is not even the big thing now. I would say that transistor stacking and the new ideas like AMDs chiplet design or Intels Forevos where they can put multiple chipelets and use TSV and a high speed, low latency bus to connect said chips. It also allows them to do use the process tech that works best for each part, so they can have the CPU be top of the line and the northbridge parts/IMC be a larger node.

We will see how it all plays out but Forevos looks very interesting if it can be applied to consumer parts and not just specialty products.

 

[InvalidError]

That said, Intel better move their butt. The days of minor research investment for big profits ARE OVER.

AMD's large jump forward with Zen is only catching up with Intel's lead on architecture, not something that AMD will be able to repeat. The days of rapid per-core performance scaling are over for all mature CPUs designed for single-threaded performance, not just Intel's. Their IPCs are all stuck near the practical limits of what instruction-level parallelism can be extracted from typical code. Most future performance gains have to come from having more cores and finding ways to write more efficient heavily threaded code for games and mainstream software. There isn't a whole lot more left that AMD and Intel can do about this regardless of how much more R&D money they might throw into desktop architecture, which is why neither is designing chips specifically for desktop anymore. AMD used a single die designed for servers for its entire non-APU lineup while Intel re-packaged ringbus clusters from its LCC/HCC server chips into a 1S die for mainstream.

Both AMD and Intel are focusing heavily on datacenter first. That's where the big money and future sales are, customers that have seemingly infinite budget to upgrade to the most complex CPUs money can buy at the time. Mainstream desktop PCs will likely reach the end of the road for practical scaling at 16 cores and ~10 years upgrade cycles will probably become the norm. If in doubt, just look at how many people here have stuck with their SB/IB/Haswell builds for 6+ years already. Many of us are getting close to that 10 years mark, it can only get worse as performance gains aside from core count bumps continue to slow down.

 

[TerryLaze]

If in doubt, just look at how many people here have stuck with their SB/IB/Haswell builds for 6+ years already. Many of us are getting close to that 10 years mark, it can only get worse as performance gains aside from core count bumps continue to slow down.

To be fair this is mostly due to people only playing consoles games and those are designed for laptop CPUs so there is not much to gain there,also GPUs have been comparatively weak in previous years hiding a lot of performance.
Now with the concept of consoles slowly dying and mowing towards more PC like and cloud based solutions this might change, but probably not they are still going to try and keep CPU usage as low as possible to maximize profits.

 

[InvalidError]

To be fair this is mostly due to people only playing consoles games and those are designed for laptop CPUs

With laptops typically being ~1GHz and one core/thread tier behind desktop CPUs, good luck playing most recent games on most newer laptops even if we ignore the IGP bottleneck. An i5-2500 or better on the other hand can still play the majority of new PC games quite decently well, 50+ FPS in most cases.

 

[TerryLaze]

With laptops typically being ~1GHz and one core/thread tier behind desktop CPUs, good luck playing most recent games on most newer laptops even if we ignore the IGP bottleneck. An i5-2500 or better on the other hand can still play the majority of new PC games quite decently well, 50+ FPS in most cases.

The what?
I guess you didn't understand what I meant, all the games are made for consoles,those have laptop CPUs from 2013 clocked at about 1.6Ghz.
https://en.wikipedia.org/wiki/PlayStation_4_technical_specifications
That's why an i5-2600 is still capable of running them just fine, they are made to run on much much much worse cores then that.

 

[jimmysmitty]

The what?
I guess you didn't understand what I meant, all the games are made for consoles,those have laptop CPUs from 2013 clocked at about 1.6Ghz.
https://en.wikipedia.org/wiki/PlayStation_4_technical_specifications
That's why an i5-2600 is still capable of running them just fine, they are made to run on much much much worse cores then that.

Or its because the developers don't want to put the time in to optimize their games for more than 4 cores?

Consoles have no bearing on games for PC unless the developer is lazy and just does a simple port. Most ports have to be recoded to utilize the PC hardware since its not just AMD and a lot of console games that are written for those consoles also come to PC with nVidia exclusive features not possible on consoles.

On top of that, quad core is still the most commone CPU even with 8 cores being vastly more common. Sales are what drives them and being able to make more sales will push them to optimize to the majority, not the minority.

 

[InvalidError]

That's why an i5-2600 is still capable of running them just fine, they are made to run on much much much worse cores then that.

PC-exclusive game developers still need to target the sort of PCs most PC gamers have regardless of what is happening on consoles if they want to have a market to sell to and for most of the past 10 years, that meant 4C4T at best before losing most of the potential market. 4C4T-4C8T still represents a significant chunk of the current budget segment in the i3-9xxx and Ryzen 2200G/2400G, so developers and publishers still can't ignore those either. It'll be many more years before games that require more than 4C4T for decent performance become the norm. Though between the publishers' infinite greed, governments cracking down on some of their monetizing schemes, a growing number of people getting wary of publishers launching games in their minimum viable product form so they can sell stuff that should have been in at launch afterward and also prioritizing endless engagement over actual fun, AAA gaming may collapse before 4C4T becomes clearly inadequate.

Aside from losing a large chunk of potential market from increasing CPU requirements, adding more threads for different things and making everything work properly requires considerably more work and coordination. It isn't something most developers want to do more than absolutely necessary.

 

[TerryLaze]

Or its because the developers don't want to put the time in to optimize their games for more than 4 cores?

Consoles have no bearing on games for PC unless the developer is lazy and just does a simple port. Most ports have to be recoded to utilize the PC hardware since its not just AMD and a lot of console games that are written for those consoles also come to PC with nVidia exclusive features not possible on consoles.

On top of that, quad core is still the most commone CPU even with 8 cores being vastly more common. Sales are what drives them and being able to make more sales will push them to optimize to the majority, not the minority.

As you can read from the wiki page
The central processing unit (CPU) consists of two x86-64 quad-core modules for a total of eight cores
and if you followed the ryzen in the news at all you know how badly games react to having threads run on cores that are on different CPUs(or CCXs) that's why games don't scale well or even at all beyond 4 cores they are written to run on mostly 4 cores because that's what's available.
"Most ports have to be recoded to utilize the PC hardware since its not just AMD"
x64 is x64 it doesn't have to be intel or amd specific the last time this was a thing was when windows 98 was still around.
"also come to PC with nVidia exclusive features not possible on consoles"
GPU is removed from CPU the GPU requirements don't change how the CPU runs the game in any way,anything graphical is just a module they can stick on to the game code.

 

[TerryLaze]

PC-exclusive game developers still need to target the sort of PCs most PC gamers have regardless of what is happening on consoles if they want to have a market to sell to and for most of the past 10 years, that meant 4C4T at best before losing most of the potential market.

Scaling is a real thing that exists....
There are plenty of game engines that allow you to make a game that will run more or less things depending on how many cores you have,just look at ashes of the singularity.
You can still run it on a dual core but you get more if you run it on 6 or 8 cores.
There are no games anymore that will push the envelope it's all about nice graphics now no matter how bad the game is.

 

[InvalidError]

Scaling is a real thing that exists....

I wrote "It'll be many more years before games that require more than 4C4T for decent performance become the norm."

Scaling may exist but it won't be common any time soon. Most games will remain fundamentally limited by the lowest common denominator they are developed for to keep the experience generally consistent across the board. I doubt developers want the added burden of having to manage substantially different game feel and experience based on CPU scaling on top of the base game itself, not to mention complications in multi-player games if the scaling applies to relevant game mechanics instead of mere visuals.