VR-Zone: FX-9590 is the final legacy of the FX line

[Silverforce11]

Seems AMD should rename itself to ATI...

LOL, sounds about right. Awesome GPUs, awesome APUs.. fail CPU for multiple generations already with no sign of improvement and now they've given up.

They may as well give up, because they arent going to compete with Intel's architecture AND node advantage on the CPU front.

What they really are more notebook design wins with their APUs.

 

[Idontcare]

Seems complicated.

White collar crime always is.

 

[BallaTheFeared]

White collar crime always is.

Haha 😛

 

[beginner99]

Yes everyone who wants to know already knows the "secret" formula for Coca-Cola. And they can manufacture it to their heart's delight, but they can't sell it (which includes giving it away) because they don't have a license from Coca-Cola to sell their IP.

Off topic:

This is in fact wrong. Anyone is free to sell a completely 100% correct copy of coke. The formula for Coke is held as a trade secret and hence no patent protection exists. The advantage being that there is no fixed expiry date on the secret. Patents have an expiration date and after that anyone can use the patents for free. Patents are important in areas were reverse engineering is trivial like with pharmaceuticals. (I know you get that but just as a general comment).

No you can argue analyzing coke is trivial but in fact that is a lot harder than one would think. The first thing is that you can add stuff that has no effect on taste but on the analytical results. Then smelling (yes, most of the "taste" actually comes from smelling) is very sensitive and miniscule amounts of a flavor can have huge effects. Just think about strawberry flavor. Personally I think its pretty bad and has not much in common with a freshly picked strawberry.

But still a competent analytical lab can do this. However they will still miss the most important part of the "secret formula". And that is simply the brand name "Coca-Cola". Which they can't use as it is a trademark and hence the actual IP of Coke. Trademarks do not expire.

 

[Vesku]

I assumed he was talking about the trademark and copyright IP in terms of Coca-Cola. Knowing you are drinking an official Coca-cola cola actually does make up part of the "tastes great, I'd never enjoy a Pepsi this much" experience for many people.

 

[MisterMac]

I assumed he was talking about the trademark and copyright IP in terms of Coca-Cola. Knowing you are drinking an official Coca-cola cola actually does make up part of the "tastes great, I'd never enjoy a Pepsi this much" experience for many people.

It's a mindset thing.

Id still wager if someone said " hey - we've got the same tastes as one of the largest soda brands in the world - test us blindly".
And it was cheap as fuck people would bait.

Assuming they cracked the whole smell\taste thing to a 99% percentile.

 

[NTMBK]

I bet Phenom is coming back.

In brand name? Perhaps, although it never gained as much traction or reputation as Athlon.

The core itself? Not a chance. Llano (improved Stars core on 32nm) was worse than Trinity (Piledriver on 32nm).

 

[bgt]

It's a biblical expression meaning the return of Jesus. In common english usage it usually means something so amazing that it's godlike.

:thumbsup:

 

[mrmt]

The core itself? Not a chance. Llano (improved Stars core on 32nm) was worse than Trinity (Piledriver on 32nm).

To be fair with the Athlon legacy AMD milked the same core from A64 (2003) until Phenom II X6, excepting for bolting new instructions. And Only after 7 years consuming a huge piece of AMD R&D pie derpdozer could bet that ancient core at the same node.

I don't know whether it would be enough to beat Intel, but I don't think it's too far fetched that had AMD picked K8 and not derpdozer they would be in a better position and better financial shape than they are now. With derpdozer they threw away the chance of their corporate lives in order to pursue a pipe dream.

 

[Charles Kozierok]

Originally Intel actually gave AMD the designs, they had to because AMD was actually a foundry for Intel. You have to give the foundry the designs (the masks). The foundry doesn't have to analyze them, but they easily and readily can if they want to.

This was done at IBM's behest -- they didn't want to be reliant on a single company to source CPUs for the PC. But AMD's reverse engineering of Intel products went back well before the x86 series -- I believe they reverse engineered the 8080 back in the 70s well before the PC came around.

To AMD's credit, they successfully transformed themselves from merely copying Intel's designs to making their own, in a few cases even besting Intel at their own game.

 

[NTMBK]

To be fair with the Athlon legacy AMD milked the same core from A64 (2003) until Phenom II X6, excepting for bolting new instructions. And Only after 7 years consuming a huge piece of AMD R&D pie derpdozer could bet that ancient core at the same node.

I don't know whether it would be enough to beat Intel, but I don't think it's too far fetched that had AMD picked K8 and not derpdozer they would be in a better position and better financial shape than they are now. With derpdozer they threw away the chance of their corporate lives in order to pursue a pipe dream.

Yeah, it's certainly not inconceivable that if all the R&D money sunk into Bulldozer had instead gone towards improving the Stars core further, AMD would have a better product today. But from their current situation, I doubt we will ever see a resurrection of the Stars core.

 

[NostaSeronx]

@NTMBK, What do you think will come after 2-way CMT?

 

[NTMBK]

@NTMBK, What do you think will come after 2-way CMT?

Not sure, really. Its not inconceivable that AMD could go for SMT on top of their CMT, but that's starting to get into SUN levels of thread silliness. They may well stick with 2-way CMT for a while, and just keep beefing it up.

Of course, there is the long term goal of somehow turning the CPU and GPU into truly a single device, but I'm not sure how close to that they are. Maybe Intel will get there first- putting 512-bit AVX3.2 into Skylake seems like a pretty big step in that direction to me.

 

[NostaSeronx]

I'm betting my money on 8-way CMT.

== Post-Dozer Family Module ==
-- 8-way Fetch --
-- 16-way Decode --
8 Integer Cores(4-way Execution/4-way Address Generation)
Ultra FlexFPU(4 * 16 * 64b)
64KB L1d per core
512KB L1i per module
12MB L2 per module​

 

[sushiwarrior]

I'm betting my money on 8-way CMT.

== Post-Dozer Family Module ==
-- 8-way Fetch --
-- 16-way Decode --
8 Integer Cores(4-way Execution/4-way Address Generation)
Ultra FlexFPU(4 * 16 * 64b)
64KB L1d per core
512KB L1i per module
12MB L2 per module​

omg no way, y not 64-way CMT. just make it the size of a small city block, just use one whole wafer for a single CPU

 

[NostaSeronx]

omg no way, y not 64-way CMT. just make it the size of a small city block, just use one whole wafer for a single CPU

DEC Alpha, that is why. AMD chip architectures after 00h have been always designed/based around DEC Alpha chips.

Bulldozer => EV6
Steamroller => EV7
Post-Bulldozer => EV8

 

[NTMBK]

I'm betting my money on 8-way CMT.

== Post-Dozer Family Module ==
-- 8-way Fetch --
-- 16-way Decode --
8 Integer Cores(4-way Execution/4-way Address Generation)
Ultra FlexFPU(4 * 16 * 64b)
64KB L1d per core
512KB L1i per module
12MB L2 per module​

I doubt we'll see 4*16*64b FPU- more likely to get a 16*32b/8*64b, to support Intel's AVX-512. (Although I guess 16*64b would support all the way up to AVX-1024.)

EDIT: Actually, I'd frankly expect more numerous, smaller vector units, if Bulldozer is anything to go by. Able to gang together multiple units to execute the wider instructions, but able to execute separate smaller instructions simultaneously (if the decode/issue bandwidth is there).

 

[NostaSeronx]

I doubt we'll see 4*16*64b FPU- more likely to get a 16*32b/8*64b, to support Intel's AVX-512. (Although I guess 16*64b would support all the way up to AVX-1024.)

Bulldozer => 4 * 64b
Steamroller => 8 * 64b
Post Bulldozer => 4 * (16 * 64b)

:whiste:, the idea of the Bulldozer family is get to ISAEs before Intel implements them, through XOP.

XOP1 => 256-bit Integer and Floating Point <-- AVX1 gen
XOP2 => 512-bit Integer and Floating Point <-- AVX2 gen
XOP3 => 1024-bit Integer and Floating Point <-- AVX3 gen

XOP doesn't share instructions with SSE, while AVX shares instructions with SSE.

XOP - eXtended OPeration: multimedia and vectorization extensions beyond AVX

----
Just to get back on topic, FX-9590 is the last FX product to be on AM3+ before EOL. FM2+ and/or G34(GC34/GC36) will carry on the FX names.

Warsaw => FX Moniker
Kaveri => AX(A10) Moniker

 

[Slomo4shO]

You're missing the point

this requires lots of engineering, verification, testing, designing, etc. TONS of extra costs vs. adding a new bin

Am I? Does Kavari and its successors not require engineering, verification, testing, designing, etc.? The Jaguar cores in the PS4 and XBox One already utilized technology to merge multiple cores with a powerful GPU. These chips utilize low TDP cores that are based on mobile platform but is it too difficult to imagine that similar research and development isn't already in progress in the desktop lineup?

 

[Exophase]

Am I? Does Kavari and its successors not require engineering, verification, testing, designing, etc.? The Jaguar cores in the PS4 and XBox One already utilized technology to merge multiple cores with a powerful GPU. These chips utilize low TDP cores that are based on mobile platform but is it too difficult to imagine that similar research and development isn't already in progress in the desktop lineup?

Of course all those things require those resources to complete. Those things also have a viable market, and are stretching the limits of what AMD can fund. There's no money to make a specialty chip that will sell in far lower quantities. An APU in particular needs to be viable in laptops and this one isn't.

Consider Intel's enthusiast line of desktop CPUs with 6 cores. These are just salvaged server parts with some cores and cache fused off. They also come out much later than the mainstream parts using the same uarchs because they're waiting on all the extra validation the server variants need. Intel could be making native 6-core parts that are only for the enthusiast market, they would come out a lot sooner and probably cost less per chip. But they don't, because the market is too small to justify the specific development of this product. And this is Intel we're talking about, who already makes far more different CPU-oriented dies than AMD does.

Arguably, even AMD's FX line would be difficult to support without cross-over into the server market, which your high-end APU would never manage.

AMD could sell PS4 chips, but they're probably not totally suitable for PCs without some modifications, and their licensing arrangement with Sony would have to allow this (I doubt it does).

 

[mrmt]

Arguably, even AMD's FX line would be difficult to support without cross-over into the server market, which your high-end APU would never manage.

Actually Bulldozer is much better suited for desktop environment than for servers environment. Just look at what happened to AMD share in desktops since Bulldozer (it decreased, and so did the share on the internal product mix) vs what happened to AMD share in servers (it plunged, and made AMD the laughing stock of that market despite the ridiculously low prices of their server chips). It's the desktop segment that sustains the server line up, not the other way around.

Ed: FX 9590 is more expensive than any 8C AMD server processor.

 

[Exophase]

Actually Bulldozer is much better suited for desktop environment than for servers environment. Just look at what happened to AMD share in desktops since Bulldozer (it decreased, and so did the share on the internal product mix) vs what happened to AMD share in servers (it plunged, and made AMD the laughing stock of that market despite the ridiculously low prices of their server chips). It's the desktop segment that sustains the server line up, not the other way around.

Ed: FX 9590 is more expensive than any 8C AMD server processor.

Bulldozer, with its focus on large core counts and large L2 cache, is easily a better fit with servers than desktops. How it did in those respective markets doesn't really change this. And it doesn't say anything about what AMD's expectations were.

But I think you missed the actual point I was making. It wasn't about whether Bulldozer was made for servers or desktops. I was saying that AMD may not have made it (it being Zambezi and successors, the 8-core non-APU chips, not the CPU uarchs) if they couldn't sell it in both desktops and servers.

 

[mrmt]

But I think you missed the actual point I was making. It wasn't about whether Bulldozer was made for servers or desktops. I was saying that AMD may not have made it (it being Zambezi and successors, the 8-core non-APU chips, not the CPU uarchs) if they couldn't sell it in both desktops and servers.

I don't think it was a case of being able to sell a product in both markets or else they would have no product, it was a case of cross subsidy, much like Nvidia used to do with their Quadro/Tesla line. There was a subordination of priorities at AMD, and servers took precedence to consumers. It is clear that since the beginning the desktop segment was supposed to subsidize the server segment through higher die costs and desktop volumes before Bulldozer were high enough to warrant a die of its own, but can you say the same about servers?

 

[Slomo4shO]

AMD could sell PS4 chips, but they're probably not totally suitable for PCs without some modifications, and their licensing arrangement with Sony would have to allow this (I doubt it does).

A derivative of this chip would be a better option than reconfiguring the Jaguar APU for the desktop. The arguement is simple: a 6, 8, or more core APU with a solid GPU that is capable of hybrid crossfire would enhance AMDs position in the mainstream market while encouraging enthusiast to consider AMD for gaming without actually directly competing with Intel on the performance front(AMD as expressed that it has no interest in competing with Intel on performance).

 

[sushiwarrior]

A derivative of this chip would be a better option than reconfiguring the Jaguar APU for the desktop. The arguement is simple: a 6, 8, or more core APU with a solid GPU that is capable of hybrid crossfire would enhance AMDs position in the mainstream market while encouraging enthusiast to consider AMD for gaming without actually directly competing with Intel on the performance front(AMD as expressed that it has no interest in competing with Intel on performance).

The counter argument is simpler: AMD can't afford to make just a 6 or 8 core CPU, and therefore AMD probably also cannot afford to make a 6 or 8 core APU with a very large graphics component. I think you underestimate how hard it is to change ANYTHING about a design. It's not as simple as "snip that wire, take out that block from the design", it is thousands of manhours of work designing, verifying, constructing, testing etc. for even the most minor alteration to the design.