Should AMD have focused on a Steamroller high TDP product line instead of Vishera?

[galego]

You mean like yours, that a PS4 is better than a high end gaming PC?

This thread is about Centurion and Steamroller. My post was about Centurion. Stop derailing the thread.

I do not think the existence of the piledriver based centurion cpu negates the possibility to make the same in steamroller.

Or rather:

I had think of Centurion only as a bridge between current 8000 series and Steamroller. But your idea of a perpetual enthusiast line for each gen is interesting.

 

[FwFred]

If AMD was smart they would keep this massive niche TDP product segment and introduce some crazy SR based FX with 6 modules and 4.5+Ghz next year. For retail channel with massive price. It would be funny to watch the river of tears as thing thing would probably cream Haswell-E 8C in a lot of benchmarks .

If Intel took that exact Haswell-E die and released a 220W CPU, how would the comparison look? Six core Ivy Bridge-E would be sufficient I would think.

 

[Sweepr]

If Intel took that exact Haswell-E die and released a 220W CPU, how would the comparison look? Six core Ivy Bridge-E would be sufficient I would think.

The comparison would look really bad for AMD. Haswell-E's server brother Haswell-EP will bring up to 14 cores at the usual 70-150W power envelope, they could easily launch one of these as a desktop EE chip if they ever need to (they wont )... and we're not talking about imaginary 220W chips here.

 

[IntelUser2000]

I'm sure everyone would be pretty excited to see a 200+ watt TDP Haswell E(xtra)EE edition released!

It wouldn't work.

Because for one launch the increase in TDP by 50%, you'd get a big gain.

The next time you keep that ridiculous TDP the same, the performance gain would be much less, meaning you'd need 50% gain on top of that already-ridiculous TDP, otherwise you'd have people complaining not-enough-gains again. That's basically how we got from mainstream processors using single digit watts to 100-130 watts in less than a decade.

And based on computers and hardware, that would happen in only 2 years, which is nothing in non-computer time.

 

[AtenRa]

It wouldn't work.

Because for one launch the increase in TDP by 50%, you'd get a big gain.

The next time you keep that ridiculous TDP the same, the performance gain would be much less, meaning you'd need 50% gain on top of that already-ridiculous TDP, otherwise you'd have people complaining not-enough-gains again. That's basically how we got from mainstream processors using single digit watts to 100-130 watts in less than a decade.

And based on computers and hardware, that would happen in only 2 years, which is nothing in non-computer time.

Not if you add a couple of cores and raising the frequency every two years with a new node

 

[bunnyfubbles]

Biggest fuss was created by people who actually won't buy this thing (and they couldn't even if they wanted since it's not going to retail channel). First it's the "TDP" problem, then it's the price problem, after that it's "fake cores", then it's fake "Ghz marketing". I wonder what's next?

If AMD was smart they would keep this massive niche TDP product segment and introduce some crazy SR based FX with 6 modules and 4.5+Ghz next year. For retail channel with massive price. It would be funny to watch the river of tears as thing thing would probably cream Haswell-E 8C in a lot of benchmarks .

while I would very much welcome such a product as its multithread performance would be pretty good and it would likely continue with AMD's aggressive pricing (as long as such a part were under $400, say $300-350, I'd have one), there's still a couple things you are kind of running wild with:

1. TDP on such a part would be very high before going to 4.5GHz

2. it still wouldn't be able to fully compete with a 130W TDP 6 core Ivy-E let alone an 8 core Haswell-E (the 4 module 8350 doesn't "cream" the 4 core i7s in any benchmark let alone a lot of them, there's no way a 6 module part would "cream" the 6 core i7s), although such a part might encourage Intel to ramp up their Haswell-E release timetable and/or release a consumer 8 core Ivy-E

3. there's nothing "smart" about a TDP arms race, its a move mostly out of desperation, as Intel could easily follow suit and put out a higher TDP part and blow away any such hypothetical AMD part. Or just simply lower prices.

 

[IntelUser2000]

Not if you add a couple of cores and raising the frequency every two years with a new node

You don't get it. Say you are getting 10% gains with new node or new architecture by keeping the TDP identical at 150W.

Then a 220W TDP nets you 30% gains(which includes the architecture or node advantage) rather than 10%. Then next generation/node, you'd need a 220W part to get you 10% gain. If you want 30% gain, you'd need 330W chip. If you want 30% after that, >450W and so on.

Oh, don't compare with video cards. Video cards use 300W including the high-clock GDDR5. Chip TDP is probably similar to the 220W AMD chip.

 

[AtenRa]

while I would very much welcome such a product as its multithread performance would be pretty good and it would likely continue with AMD's aggressive pricing (as long as such a part were under $400, say $300-350, I'd have one), there's still a couple things you are kind of running wild with:

1. TDP on such a part would be very high before going to 4.5GHz

2. it still wouldn't be able to fully compete with a 130W TDP 6 core Ivy-E let alone an 8 core Haswell-E (the 4 module 8350 doesn't "cream" the 4 core i7s in any benchmark let alone a lot of them, there's no way a 6 module part would "cream" the 6 core i7s), although such a part might encourage Intel to ramp up their Haswell-E release timetable and/or release a consumer 8 core Ivy-E

3. there's nothing "smart" about a TDP arms race, its a move mostly out of desperation, as Intel could easily follow suit and put out a higher TDP part and blow away any such hypothetical AMD part. Or just simply lower prices.

1. SR is at 28nm, it should have better electrical characteristics.

2. SR will have higher IPC, just put 15% on top of FX8350 and you beat 3770K almost on every MT test.

 

[AtenRa]

You don't get it. Say you are getting 10% gains with new node or new architecture by keeping the TDP identical at 150W.

Then a 220W TDP nets you 30% gains(which includes the architecture or node advantage) rather than 10%. Then next generation/node, you'd need a 220W part to get you 10% gain. If you want 30% gain, you'd need 330W chip. If you want 30% after that, >450W and so on.

Oh, don't compare with video cards. Video cards use 300W including the high-clock GDDR5. Chip TDP is probably similar to the 220W AMD chip.

Having a new node will alow you to install one or two more cores keeping the same die size and power consumption at the same frequency easily. You may not gain a lot of IPC but you will get way higher MT performance.

 

[galego]

while I would very much welcome such a product as its multithread performance would be pretty good and it would likely continue with AMD's aggressive pricing (as long as such a part were under $400, say $300-350, I'd have one), there's still a couple things you are kind of running wild with:

1. TDP on such a part would be very high before going to 4.5GHz

2. it still wouldn't be able to fully compete with a 130W TDP 6 core Ivy-E let alone an 8 core Haswell-E (the 4 module 8350 doesn't "cream" the 4 core i7s in any benchmark let alone a lot of them, there's no way a 6 module part would "cream" the 6 core i7s), although such a part might encourage Intel to ramp up their Haswell-E release timetable and/or release a consumer 8 core Ivy-E

Never say never. I already gave some benchmark where the 4M FX-9590 is faster that the 4C i7-4770k and is faster, or just matches, an 6C i7 CPU

http://forums.anandtech.com/showpost.php?p=35146061&postcount=501

The 9590 is faster than the 6C i7 on a game as crysis 3, and matches it in x264 HD-Benchmark:

i7-4770k: 16.4
FX-9590: 18.0-18.9
i7-3930K: 19.2

It is possible to find some few benchmarks where the old 8350 at stock is faster than the new 4770k

http://openbenchmarking.org/embed.php?i=1306034-UT-INTELCORE31&sha=443c58e&p=2

http://openbenchmarking.org/embed.php?i=1306034-UT-INTELCORE31&sha=0a735b7&p=2

http://openbenchmarking.org/embed.php?i=1306034-UT-INTELCORE31&sha=d1d6430&p=2

An it is possible to find some benchmark where the old 8350 at stock is faster than the 6C i7-3960X.


Therefore his idea of a 6M Steamroller based FX at 4.5+Ghz beating Haswell-E 8C is entirely possible.

 

[guskline]

galego, how are you making predictions about a FX 9500? I haven't seen you run any tests? Are you interpolating from a chart?

Check my post about an ACTUAL FX 8350 OC'd to 4.7 Ghz (mine actually) and see real results for my rig. I used the 4.7 clock because most postings show the fx 9570 WITH A BASE OF 4.7gHZ.

BTW out of curiousity where is the benchmark where a stock 8350 is faster than an I7 3960X?

 

[sm625]

It should have been an APU. An A10-6800K with 2048 shaders. No one needs 200W of cpu power. Lots of people can easily use 200W of gpu.

 

[boxleitnerb]

The FX isn't faster in Crysis 3. Current game version is 1.4 which changes things quite alot vs 1.0:
http://www.pcgameshardware.de/Crysis-3-PC-235317/Tests/Crysis-3-CPU-Test-1068140/

 

[SlowSpyder]

It should have been an APU. An A10-6800K with 2048 shaders. No one needs 200W of cpu power. Lots of people can easily use 200W of gpu.

Memory bandwidth would make such a part dead on arrival.

 

[Sweepr]

The FX isn't faster in Crysis 3. Current game version is 1.4 which changes things quite alot vs 1.0:
http://www.pcgameshardware.de/Crysis-3-PC-235317/Tests/Crysis-3-CPU-Test-1068140/

Crysis 3

Core i7-3770K
1.0: 56,7 Avg-Fps
1.3: 74,9 Avg-Fps

FX-8350
1.0: 58,6 Avg-Fps
1.3: 57,4 Avg-Fps

* They tested Welcome to the Jungle level, the only level FX8350 managed to match/beat a quad-core Ivy Bridge in Crysis 3 version 1.0.

Wow, seems like there were a lot of AMD-specific optimizations at launch, its a Gaming Evolved title after all. Patch 1.3 puts things ''back to normal'', i7 4C/8T easily outperforming 8-core (twice the die size) FX8350.

 

[Gikaseixas]

so AMD actually gives lower FPS from ver 1.0 to 1.3 and Intel improved by almost 20 FPS? weird

 

[galego]

galego, how are you making predictions about a FX 9500? I haven't seen you run any tests? Are you interpolating from a chart?

Check my post about an ACTUAL FX 8350 OC'd to 4.7 Ghz (mine actually) and see real results for my rig. I used the 4.7 clock because most postings show the fx 9570 WITH A BASE OF 4.7gHZ.

BTW out of curiousity where is the benchmark where a stock 8350 is faster than an I7 3960X?

I am reading the benchmarks from the graphs given. Look the score of the FX-9590 in the graph.

Stock 8350 vs i7 3960X

http://openbenchmarking.org/embed.php?i=1305170-UT-LLVMCLANG75&sha=b3c948c&p=2

http://openbenchmarking.org/embed.php?i=1305170-UT-LLVMCLANG75&sha=35221b6&p=2

http://openbenchmarking.org/embed.php?i=1305170-UT-LLVMCLANG75&sha=82ca41f&p=2

The FX isn't faster in Crysis 3. Current game version is 1.4 which changes things quite alot vs 1.0:
http://www.pcgameshardware.de/Crysis-3-PC-235317/Tests/Crysis-3-CPU-Test-1068140/

Yes, they have released patches that downgrade the performance of FX chips. LOL There are FX users complaining on forums because last patches drop 10-20 FPS. LOL

 

[boxleitnerb]

No, the performance change of the FX CPUs is miniscule. I believe proper testing more than some people on some forums. The major change was that the SMT problem was fixed and that the respective Intel CPUs perform much better now.

It would be good if you would read the info in the links and think before you actually post

 

[galego]

No, the performance change of the FX CPUs is miniscule. I believe proper testing more than some people on some forums. The major change was that the SMT problem was fixed and that the respective Intel CPUs perform much better now.

It would be good if you would read the info in the links and think before you actually post

The link shows a small drop in performance for the FX chips in the demo Jungle.

But if you read my post, I wrote about how gamers are complaining in the crysis forums about the entire game. This is the link

http://www.mycrysis.com/forums/viewtopic.php?f=58&t=63842

 

[boxleitnerb]

I did! And I contest the validity of this complaint (one single complaint).

 

[Remobz]

Still sticking with Intel for now. Very interesting read in this thread all the same.

 

[galego]

I did! And I contest the validity of this complaint (one single complaint).

More than one gamer is complaining by massive drops of performance of up to -60 FPS

http://www.videogamer.com/pc/crysis_3/news/crysis_3_patch_1_3_is_causing_pc_performance_issues.html

http://www.eurogamer.pt/articles/2013-04-04-patch-de-crysis-3-causa-problemas-na-performance-do-pc

 

[bunnyfubbles]

Never say never. I already gave some benchmark where the 4M FX-9590 is faster that the 4C i7-4770k and is faster, or just matches, an 6C i7 CPU

http://forums.anandtech.com/showpost.php?p=35146061&postcount=501

The 9590 is faster than the 6C i7 on a game as crysis 3, and matches it in x264 HD-Benchmark:

i7-4770k: 16.4
FX-9590: 18.0-18.9
i7-3930K: 19.2

It is possible to find some few benchmarks where the old 8350 at stock is faster than the new 4770k

http://openbenchmarking.org/embed.php?i=1306034-UT-INTELCORE31&sha=443c58e&p=2

http://openbenchmarking.org/embed.php?i=1306034-UT-INTELCORE31&sha=0a735b7&p=2

http://openbenchmarking.org/embed.php?i=1306034-UT-INTELCORE31&sha=d1d6430&p=2

An it is possible to find some benchmark where the old 8350 at stock is faster than the 6C i7-3960X.


Therefore his idea of a 6M Steamroller based FX at 4.5+Ghz beating Haswell-E 8C is entirely possible.

wait, in which one of those benchmarks does intel get creamed? the point of my post was to address the ludicrous notion of a fantasy product when

1. while current AMD parts can compete with intel in a few select benchmarks, the 4 module AMD part still loses to the 4 core intel in most of them, multithread included

and

2. that a 6 module AMD (50% more) part would not just somehow beat an 8 core (100% more) intel, but would cream it...

I'm sorry, but that's just delusional

I'd really love for AMD to put out these parts though, even if I don't expect miracles from them like you guys do.

 

[inf64]

6 module part would have drastically improved core (SR), not just 6 PD modules.
AMD states 30% higher ops/cycle throughput so 50% more cores(or threads) with 30% higher throughput (20% higher "IPC" x 10% higher MT efficiency due to split decoding stage) would equal 1.95x or practically 2x the performance of 8350 if the new 6M SR based FX would run at ~4Ghz @ 28nm.

So there you go, if ^^ above would to happen it would cream it. No doubt about it.

 

[Exophase]

6 module part would have drastically improved core (SR), not just 6 PD modules.
AMD states 30% higher ops/cycle throughput so 50% more cores(or threads) with 30% higher throughput (20% higher "IPC" x 10% higher MT efficiency due to split decoding stage) would equal 1.95x or practically 2x the performance of 8350 if the new 6M SR based FX would run at ~4Ghz @ 28nm.

So there you go, if ^^ above would to happen it would cream it. No doubt about it.

Maybe I'm reading this wrong, but this seems contingent on the idea that 6M Steamroller will use as much power as 4M Piledriver at the same clock speed. If that's what you expect I think you're overestimating the performance improvements of GF's 28nm process by a lot.

I expect the transition to be somewhat like going from 32nm to 28nm on Samsung's process (given that they're both common platform). Which isn't that big of a deal - Anand guesstimated 15% lower power consumption at the same clock speed, which seems about right with what you can see in Exynos 5250 vs 5410 and Samsung's promotional numbers. In AMD's case you have to factor in moving from PD-SOI to bulk, which I expect to hurt power consumption more than it helps. And I expect an SD module implemented on the same process to use more power/MHz than a PD module - I don't believe AMD got all of these performance improvements from free. So I just don't see a 50% power reduction at the same clocks happening. I'd more readily the power/module to be about the same for the same clocks.

Now if this 4GHz part is supposed to be one of these 220W monsters then it's a different story.. but who knows what thermal envelope Intel will release their 8C Haswell-Es in. Nevermind the 12C parts.

And if SR really did deliver 2x the peak throughput vs 4M Vishera how would that make it cream 8C Haswell-E if that also clocks about the same as the 4C one? 4M Vishera doesn't cream 4C Haswell, and on average 6M SR will only scale worse because it needs 12 fully loaded threads to achieve that throughput.

Personally, I don't think AMD will even do a 6M Steamroller part, at least not one that isn't an MCM. The demand in the desktop space is going to be too low, and I really can't see AMD doing a server-only die now - they never have in the past and their server department looks like it's shifting focus from their high end parts.