(gamersnexus) AMD: "FX is Not EOL" & Why What We Need in a CPU is Changing

[Gikaseixas]

ix the massive performance problems AMD got

Many didn't upgrade to Haswell because their SB and IB cpus have great performance and the existing FX competes well against those two arch's. I have Haswell and Piledriver, i use both daily and i don't see those performance problems you talk about.

I do hope they push the FX line forward and improve it even further. They did with the Bulldozer crap to Piledriver so perhaps we might see another revision.

 

[SlickR12345]

I think the newer FX processors from AMD are great for 2014 and 2015, its just that AMD released them too soon, when no application was using more than 2 cores at best, most were one core applications, so the fact that they released 6 and 8 cores meant very few applications can take advantage of.

Even today we don't see many requiring more than 2 cores and at best 4 cores.

There are few games like Crysis 3 and Battlefield 4 that use more than 4 cores and of course there are few applications that use more than 4 cores, I suspect its only going to be the mid of 2014 where we see more applications using 6 or 8 cores.

I think AMD should have stuck with the Phenom II architecture till now, improve it, shrink it, change things up, increase the frequency and make them competitive and then in early 2014 release the FX version fully ready after 3/4 years of research and development.

They could have also worked more with developers if they hadn't released it so early to make more applications use 6 and 8 cores and when they released the FX processors in 2014 have huge amount of applications utilizing all the cores and they would have been winning or at least on par with Intel.

 

[bgt]

I have Haswell and Piledriver, i use both daily and i don't see those performance problems you talk about

This,

 

[rtsurfer]

What had you all expected the guy to say. AM3+ is dead. But AMD would like to keep selling them to the naive people who still think there is an upgrade path and that the next CPU will magically fix the massive performance problems AMD got. Money is money. However there is not enough money in making a new CPU. And as mrmt says, its a free place to dump broken Opterons.

While the FX brand may not be dead, since AMD can just brand a 2M/4T SR or EX as FX. The AM3+ platform certainly is.

I don't think the issue was if AM3+ is dead or not.

The issue was that if AMD was going to make more FX CPUs aka 4M/8C processors. I don't think it is a bad thing to dump broken Opteron as it seems the standard industry practice to rebrand and sell broken products.

As to AM3+, it is as dead as LGA 1150. Given the frequency at which Intel changes sockets, AMD can be allowed a lit bit of leeway as they are still not as bad as Intel.

 

[NaroonGTX]

Dump broken Opterons to AM3+? How would that work? The FX series used the exact same dies from Opteron to begin with, the only difference being 3 of the 4 HT links being disabled.

I think AMD should have stuck with the Phenom II architecture till now, improve it, shrink it, change things up, increase the frequency and make them competitive and then in early 2014 release the FX version fully ready after 3/4 years of research and development.

This would've been terrible. The K8 uarch hit its wall (yes, what people refer to as K10 was really still K8 -- Phenom and Phenom II were literally K8 in just about every way, just slightly enhanced as usual) and AMD knew it. That's why the first Bulldozer the public got in 2011 was rushed, they needed a new product and decided to launch it and 'fix it' with later revisions instead (PD, SR, etc.)

We already got a tweaked Stars part in the form of Llano, which was faster than Phenom II by about 6~7% clock-for-clock. Bulldozer was never gonna be a majestical part. JF-AMD was the only one saying there would be a jump in IPC, whilst the actual AMD engineers were the ones who told everyone the real deal -- that there would be a slight regression because the focus of BD was higher parallelism-based workloads and not ST performance.

Sadly Kaveri was delayed by a whole year for reasons we don't and probably won't ever know, which also pushed Carrizo back as well. The K15 uarch (Bulldozer family) has a lot of room for improvement in various areas, and SR looks to be the first part that makes a decent leap above the performance that AMD originally offered in 2008~2009.

 

[ShintaiDK]

Dump broken Opterons to AM3+? How would that work? The FX series used the exact same dies from Opteron to begin with, the only difference being 3 of the 4 HT links being disabled.

Bins and leakage.

You can try find the temperature specs for a FX8350 for example. You might notice there isnt any. Other specs missing too. Hence the issue of 140W+ FX8350 125W TDP CPUs.

 

[NaroonGTX]

Oh, that. I remember some people reporting some of their FX octocores had higher actual TDP's than what was reporting on the boxes.

 

[nismotigerwvu]

Dump broken Opterons to AM3+? How would that work? The FX series used the exact same dies from Opteron to begin with, the only difference being 3 of the 4 HT links being disabled.



This would've been terrible. The K8 uarch hit its wall (yes, what people refer to as K10 was really still K8 -- Phenom and Phenom II were literally K8 in just about every way, just slightly enhanced as usual) and AMD knew it. That's why the first Bulldozer the public got in 2011 was rushed, they needed a new product and decided to launch it and 'fix it' with later revisions instead (PD, SR, etc.)

We already got a tweaked Stars part in the form of Llano, which was faster than Phenom II by about 6~7% clock-for-clock. Bulldozer was never gonna be a majestical part. JF-AMD was the only one saying there would be a jump in IPC, whilst the actual AMD engineers were the ones who told everyone the real deal -- that there would be a slight regression because the focus of BD was higher parallelism-based workloads and not ST performance.

Sadly Kaveri was delayed by a whole year for reasons we don't and probably won't ever know, which also pushed Carrizo back as well. The K15 uarch (Bulldozer family) has a lot of room for improvement in various areas, and SR looks to be the first part that makes a decent leap above the performance that AMD originally offered in 2008~2009.

Just playing devil's advocate here, but Llano carried a 6~7% increase in IPC without an L3$ and wattage going to the iGPU (limiting clocks on the CPU). Now, we don't know for sure how much improvement, if any at all, an L3$ would have given or how well a "Phenom III" (Llano without a GPU on die) would have clocked (it doesn't seem like much of a stretch that the die shrink would have allowed for higher clocks), but considering how well the existing Thuban held up next to Bulldozer, I would have bet a Phenom III X4/6/8 would have been a more competitive product than Bulldozer. Granted, hindsight is always 20/20, but in this case they could have kept treading water with the Phenom line and put more time into the new uArch. Definitely a stopgap solution that would have only had about a year on the market though. It just felt really odd that they massaged the stars cores as much as they did only for Llano. Who knows, maybe this was the plan and the initial results came back really poorly and BD was promoted. Anyways, I'm just rambling at this point...

 

[Blitzvogel]

Stars could've been employed for another couple years, no doubt, especially with a move to 256 bit FPU and 28 nm. The lost of IPC due to the module nature of Bulldozer is where the snag was hit. Not sure how much R&D would've been required to make it worth while, but 3/4 core Stars w/ 256 bit FPU + 512 GCN SPs would've been pretty sweet, with a 6 or 8 core FX version on AM3 to succeed Thuban.

Just wondering, on 32 nm how does a single Steamroller module compare to a couple Llano Stars cores size wise?

Perhaps module architecture might have been the way to go later on as it was engineered and matured to the state that we're expecting Excavator to be with Stars carrying things on through 2013/2014.

 

[Spawne32]

with the way their motherboard support is right now (8 core cpu's having nothing but issues with VRM power) and the lack of mATX form factor that support FX 8 core cpu's, i dont see why that graphic could not be true.

 

[NaroonGTX]

Llano was originally supposed to hit clocks beyond 4ghz (which is what Trinity and Richland ended up doing instead), but due to GloFo yield and production issues (their 32nm transition at this point was crap), Llano wound up with the low/moderate clocks the final product had. I was also disappointed with the OC potential of Llano, both on the GPU and the CPU. CPU could routinely only get up to 3.5~3.6ghz stable, and the GPU somewhere a little above 800mhz.

They could've used Phenom II a bit longer, but 'Phenom III' would've been misleading because the performance boost wouldn't have been big enough to warrant such a label.

Just wondering, on 32 nm how does a single Steamroller module compare to a couple Llano Stars cores size wise?

We don't know the size of a SR module yet, and SR is on 28nm, not 32nm. I think AMD may reveal more about the area of the die (and hopefully provide some pretty die shots finally!) at CES 2014.

 

[nismotigerwvu]

I was only referring to it as a "Phenom III" to follow with the naming conventions AMD had already used. The architectural changes between between Agena on 65m and Deneb on 45nm are on par with (or possibly less than) the the changes seen with Llano on 32nm. That said, even with a 6~10% IPC increase (the Llano core tweaks plus whatever a nice big L3 cache could do) these things would have had to clock up quite high to make Intel alter their path.

The truly odd point here is that Deneb/Thuban managed to reach 3.7ghz with the X4 980 and X6 1100t (at greater IPC) on 45nm while only the only mainstream chips from the BD/PD lineage to surpass this mark have been the A10-6800k and 6600K along with the FX-8350, 9370 and 9590 (with the last 2 shipping in only tiny quantities). Considering the hallmark of the new architecture was high clockrate and that they came with a die shrink (which traditionally comes with higher clock ceilings) this just doesn't add up. There's no way for us to know for sure, but it could be due to a poor 32nm node at GF, fundamental flaws with BD/PD or both.

 

[NaroonGTX]

I remember old AMD projections from around 2002 or 2003, when they were mentioning a future architecture that would reach 10ghz base clock. I wish I had some of the stuff they were smoking at the time (to be fair, most of the other things mentioned in said projection list either panned out or were at least realistic.)

 

[JDG1980]

I remember old AMD projections from around 2002 or 2003, when they were mentioning a future architecture that would reach 10ghz base clock. I wish I had some of the stuff they were smoking at the time (to be fair, most of the other things mentioned in said projection list either panned out or were at least realistic.)

Intel had the same 10GHz plans for NetBurst. Of course, that didn't work either - it topped out around 4 GHz. There seems to be some sort of natural "limit" between 4 and 5 GHz, that makes it very difficult to run a CPU any faster than that no matter how simple and streamlined it is.

 

[Homeles]

It's a limit of silicon. And really, silicon can clock much higher, but complex logic makes it unfeasible. Power consumption is also a concern. Intel allegedly had a working 8GHz FPU back in the Netburst era.

SiGe and Ge transistors are on their way in the next few years. III-V semiconductors are on their way as well. Clock speeds should rise substantially, but there are obviously tremendous hurdles.

Here's an interesting PDF: http://www.avsusergroups.org/cmpug_pdfs/CMP2013_5RHill-Sematech.pdf

 

[nemesismk2]

All this means is that i can use my amd fx 8350 and amd fx 4300 for even longer. I was planning on keeping my amd fx 4300 longer because IT'S THE most quiet CPU i have ever owned! Yes it is still using the standard AMD Cooler which it came with the AMD FX 4300. I have always cleaned the dust off the CPU Cooler every so often to keep it working 100% and working quiet.

 

[Ajay]

It's a limit of silicon. And really, silicon can clock much higher, but complex logic makes it unfeasible. Power consumption is also a concern. Intel allegedly had a working 8GHz FPU back in the Netburst era.

SiGe and Ge transistors are on their way in the next few years. III-V semiconductors are on their way as well. Clock speeds should rise substantially, but there are obviously tremendous hurdles.

Here's an interesting PDF: http://www.avsusergroups.org/cmpug_pdfs/CMP2013_5RHill-Sematech.pdf

Well, initially, I know that the P4 had a double pumped ALU. I don't know if it stayed that way, but it would explain why they had a 8 GHz ALU running - they would need it!

Thanks for the paper. Last I read about this was a summary presentation on options for future materials by Kuhn (including Graphene, etc.). Obviously, it was light on details. I do have to say, she is one smart lady! The talent pool at Intel is sick (Intel Fellows are listed -> here <-).

 

[sm625]

If FX8350 or 9370 or 9590 is all you can get at the end of 2015, then that's the same thing as saying it is eol now. It will never have a faster microarchitectural upgrade available.

 

[NTMBK]

If FX8350 or 9370 or 9590 is all you can get at the end of 2015, then that's the same thing as saying it is eol now. It will never have a faster microarchitectural upgrade available.

EoL = no longer being sold at all, and not stocked. For instance, Intel has plenty of Ivy Bridge and Sandy Bridge processors which have not been EoL'd yet, despite the fact that the socket is dead.

 

[Leyawiin]

Semantics.

 

[Spawne32]

If FX8350 or 9370 or 9590 is all you can get at the end of 2015, then that's the same thing as saying it is eol now. It will never have a faster microarchitectural upgrade available.

I think thats exactly whats going to happen, production will wind down over 2014 to the point where its EOL in 2015. If AMD was smart they would continue the FX lineup in the same limited capacity they have been, 2 different 8 core processors, 2 different 6 core processors, and 3 different quad core processors for the desktop lineup, on a new manufacturing process. It doesnt have to have an insane amount of variety, they just need to keep competitive in the small area of the market that enthusiasts reside. Again the problem right now is lack of variety with the chipsets and motherboards. Yeh 990FX is a great chipset for the FX, but its available on so few boards, zero mATX boards, and most of those boards cannot run the 9xxx series and have trouble with the 83xx series.

 

[inf64]

I think thats exactly whats going to happen, production will wind down over 2014 to the point where its EOL in 2015. If AMD was smart they would continue the FX lineup in the same limited capacity they have been, 2 different 8 core processors, 2 different 6 core processors, and 3 different quad core processors for the desktop lineup, on a new manufacturing process.

I put bold on the problematic section of you post. You conclude why we don't have such products on either fm2+ or am3+. Won't comment more than that .

 

[Spawne32]

I put bold on the problematic section of you post. You conclude why we don't have such products on either fm2+ or am3+. Won't comment more than that .

not sure what the problem was with my post, i said if they continue the lineup with a few models of a 28nm product they could continue to compete with intel in this area.

 

[NaroonGTX]

Piledriver on 28nm wouldn't really be anything special, unless you were referring to Steamroller FX, which everyone would agree would be nice to have, though it won't be happening.

 

[Spawne32]

Piledriver on 28nm wouldn't really be anything special, unless you were referring to Steamroller FX, which everyone would agree would be nice to have, though it won't be happening.

of course not, thats why i said, if they were smart, but they arent, and thats why they are suffering lol