AMD delays the delay -Kaveri- is its last competitive market gone?

[mrmt]

All boils down to Steamroller CPU efficency

Steamroller main improvements aren't really refinements of the architecture, but more of the same brute force approach that made Bulldozer what it is. For example, they are going for a bigger decoder and for a bigger L1 cache, both of which increases die size area and power consumption.

So we might expect a bigger chip. If it is bigger and is not sold for a bigger ASP, then AMD will be in a worse position than they are now.

 

[ShintaiDK]

Right now Steamroller is gone from roadmaps.

 

[Abwx]

Steamroller main improvements aren't really refinements of the architecture, but more of the same brute force approach that made Bulldozer what it is. For example, they are going for a bigger decoder and for a bigger L1 cache, both of which increases die size area and power consumption.

So they will implement a bigger decoder and cache L1 that will just increase die area and power comsumption with no other effect ?..

 

[mrmt]

So they will implement a bigger decoder and cache L1 that will just increase die area and power comsumption with no other effect ?..

It will increase performance but at the expense of die area and power consumption, areas where AMD is already with problems.

 

[AtenRa]

Steamroller main improvements aren't really refinements of the architecture, but more of the same brute force approach that made Bulldozer what it is. For example, they are going for a bigger decoder and for a bigger L1 cache, both of which increases die size area and power consumption.

So we might expect a bigger chip. If it is bigger and is not sold for a bigger ASP, then AMD will be in a worse position than they are now.

Steamroller will not be at 32nm SOI, most probable at 20nm, so no problems with die sizes and ASPs.

 

[Abwx]

It will increase performance but at the expense of die area and power consumption

http://www.hardware.fr/focus/71/amd-devoile-steamroller-jaguar.html

 

[boxleitnerb]

20nm...2015 if I had to guess. Glofo don't get their act together. I doubt they will catch up to Intel in the next 5-10 years.

 

[mrmt]

Steamroller will not be at 32nm SOI, most probable at 20nm, so no problems with die sizes and ASPs.

That's really simple, no, just shrink the chip and the area problem is gone. When will Steamroller arrive and what will be its main competitor, and at what node?

 

[mrmt]

20nm...2015 if I had to guess. Glofo don't get their act together. I doubt they will catch up to Intel in the next 5-10 years.

At a projected average 200 million dollars cash burn per quarter, their cash will get them through 5 quarters at most. They don't have 5 to ten years.

 

[boxleitnerb]

They can compete on a larger process node, that is not that much of a problem. But two process nodes...that's too tough. And I see exactly that happening, because I don't trust Global Foundries with the ability to stick to their roadmap.

 

[mrmt]

They can compete on a larger process node, that is not that much of a problem. But two process nodes...that's too tough. And I see exactly that happening, because I don't trust Global Foundries with the ability to stick to their roadmap.

Can they? In servers they are in the same process node as of Intel and their server business went kaput. Their server chips are competitive with Nehalen, a 45nm chip. For desktops their offers aren't competitive with Sandy Bridge, a 32nm architecture at the same node, just with Westmere, a three years old technology (albeit 32nm too). Only in APUs they offers are competitive with Intel's, only because they went ballistic on the GPU part of the chip, which resulted in a chip almost twice the size of Intel low cost chips, which is what is effectively killing their balance sheet. For everybody waiting GNC, remember that GNC also increases die size area compared to Cayman, so some of the shrinkage budget will have to be spent here.

AMD fundamental problem is not the process lag, but a very inefficient CPU architecture that can only compete by throwing at the problem a lot of more hardware than the competition. Steamroller does not change that, it is more of the same brute force approach. More decoder, bigger cache, more frequency.

edit: Don't you see how AMD approach with bulldozer mirrors Intel strategy with Netburst? "Moar, Moar, Moar!!!!"

 

[boxleitnerb]

When - if - their Fusion strategy pays off, they could have a chance. I wouldn't necessarily compare their strategy with Netburst. I've seen presentation slides where they could achieve as much as +30% ops/cycle with Steamroller which in the best case means +30% IPC:
http://www.xbitlabs.com/news/cpu/di...Steamroller_Third_Gen_Bulldozer_Incoming.html

If they pulled that off, then they might have a chance. But that strongly depends on when Steamroller actually comes, on what node and what Intel did in the mean time.

 

[mrmt]

When - if - their Fusion strategy pays off, they could have a chance.

Fusion won't pay off.

I wouldn't necessarily compare their strategy with Netburst. I've seen presentation slides where they could achieve as much as +30% ops/cycle with Steamroller which in the best case means +30% IPC

AMD market is usually very incompetent, but the fact that the web is full of people that are a little more educated than the average consumer and those same people still buy AMD fairy tales is a remarkable success.

 

[boxleitnerb]

Personally, I'm sceptical too, but I wouldn't completely rule out a surprise. It's not like their engineers are bad, but everything else about that company is, especially marketing and management which are dragging the design teams down.

 

[Idontcare]

Steamroller will not be at 32nm SOI, most probable at 20nm, so no problems with die sizes and ASPs.

If AMD did that then that would mean they are continuing to do exactly what they said they don't want to be doing (i.e. the stuff on the lefthand side of the slide):

I doubt they went to all the trouble of assembling the slide above, making analysts aware of it, only to then turn their back to it and slog straight ahead into pursuing the development of high performance products like steamroller that require adopting bleeding edge technology nodes.

That would be pretty dumb of them to stand up and say "we know what the problem here, but we are going to keep doing it for years and years to come!"

 

[mrmt]

Personally, I'm sceptical too, but I wouldn't completely rule out a surprise. It's not like their engineers are bad, but everything else about that company is, especially marketing and management which are dragging the design teams down.

You should rule out. When did we see 30% jumps in IPC in the same uarch? That's right, never. Unless AMD is using a lot of fairy dust, there is no way they can extract 30% more performance from the same uarch by simply adding a decoder and enlarging L1 cache.

 

[AtenRa]

For desktops their offers aren't competitive with Sandy Bridge, a 32nm architecture at the same node.

FX8350 vs 2600K


Adobe Photoshop CS4 - Retouch Artists Speed Test
Time in Seconds - Lower is Better

FX8350 = 13.3
2600K = 11.3

DivX 6.8.5 Encode (Xmpeg 5.0.3)
Time in Seconds - Lower is Better

FX8350 = 31.4
2600K = 29.1

x264 HD Encode Test - 1st pass - x264 0.59.819
Frames Per Second - Higher is Better

FX8350 = 90.4
2600K = 106.4

x264 HD Encode Test - 2nd pass - x264 0.59.819
Frames Per Second - Higher is Better

FX8350 = 44
2600K = 36.3

Windows Media Encoder 9 x64
Time in Seconds - Lower is Better

FX8350 = 25
2600K = 20

Cinebench R10 - Multi-Threaded Benchmark
Score in CBMarks - Higher is Better

FX8350 = 23437
2600K = 22875

POV-Ray 3.7 beta 23 - SMP Benchmark
CPU Score in PPS - Higher is Better

FX8350 = 5008
2600K = 4875

Par2 - Multi-Threaded par2cmdline 0.4
Time to Recover 57.2MB from 707.5MB Archive in Seconds - Lower is Better

FX8350 = 16.1
2600K = 17.3

Blender 2.48a Character Render
Time in Seconds - Lower is Better

FX8350 = 44.6
2600K = 40.1

Microsoft Excel 2007 SP1 - Monte Carlo Simulation
Time in Seconds - Lower is Better

FX8350 = 12.6
2600K = 11.1

Sorenson Squeeze Pro 5 - Flash Video Creation
Time in Seconds - Lower is Better

FX8350 = 77.4
2600K = 72.7

WinRAR 3.8 Compression - 300MB Archive
Time in Seconds - Lower is Better

FX8350 = 74.9
2600K = 59.6

Cinebench 11.5 - Multi-Threaded
Score in CBMarks - Higher is Better

FX8350 = 6.89
2600K = 6.86

x264 HD Benchmark - 1st pass - v3.03
Frames per Second - Higher is Better

FX8350 = 89.6
2600K = 94.9

x264 HD Benchmark - 2nd pass - v3.03
Frames per Second - Higher is Better

FX8350 = 41.9
2600K = 36

7-zip Benchmark
32MB Dictionary - Total MIPS - Higher is Better

FX8350 = 23407
2600K = 19744


FX8350 is faster at 7 out of 16
2600K is faster at 9 out of 16.

note: FX8350 is faster in x264 benchmark (overall) counting both first and second pass. That would make the score 7 - 7 each.

I believe anyone could see that at the same 32nm, AMD Piledriver is very competitive against Intel's 32nm Sandybridge Core i7. If AMD had 32nm and 22nm at the same time(within 6 months) as Intel, things would be completely different.

 

[boxleitnerb]

It's still losing badly in lighly threaded apps and most games - not everything is multithreaded well - and uses quite a bit more power. A little different, yes. Much? No.

 

[mrmt]

I believe anyone could see that at the same 32nm, AMD Piledriver is very competitive against Intel's 32nm Sandybridge Core i7. If AMD had 32nm and 22nm at the same time(within 6 months) as Intel, things would be completely different.

Oh dear, what are you trying to prove here? Putting a 346 mm^2 125W server chip against a 216mm^2 95W mainstream chip with an IGP? Are you really serious?

Why didn't you take the 2700k or even an overclocked mainstream SNB for instance? Why don't you bring here gaming performance and less multi-threaded benchmarks, which are a reality for a lot of people? And why don't you bring SNB-EP, which is where you can find the true limits of SNB?

You just proved my point in other message: AMD needs to burn more power and throw more hardware to get the same performance of Intel *past* products.

 

[frozentundra123456]

It's still losing badly in lighly threaded apps and most games - not everything is multithreaded well - and uses quite a bit more power. A little different, yes. Much? No.

Yea, in gaming of the six games in anand bench which has data for the 8350 and 2500K, it ties one and loses the rest by a few percent to 30 or 40 percent. More power for lower performance, at least in gaming. Since that is the only thing I do that is CPU intensive, intel all the way for me. Only place AMD makes sense is if you do a lot of work with the heavily threaded apps in which it does well.

 

[AtenRa]

It's still losing badly in lighly threaded apps and most games - not everything is multithreaded well - and uses quite a bit more power. A little different, yes. Much? No.

You dont buy the Core i7 unless you want Multithreaded performance, same goes for the FX8350. In games at 1080p they are almost the same in the majority of the new games(yes yes SC2 is an exception).

 

[JumpingJack]

FX8350 vs 2600K


Adobe Photoshop CS4 - Retouch Artists Speed Test
Time in Seconds - Lower is Better

FX8350 = 13.3
2600K = 11.3

DivX 6.8.5 Encode (Xmpeg 5.0.3)
Time in Seconds - Lower is Better

FX8350 = 31.4
2600K = 29.1

......



note: FX8350 is faster in x264 benchmark (overall) counting both first and second pass. That would make the score 7 - 7 each.

I believe anyone could see that at the same 32nm, AMD Piledriver is very competitive against Intel's 32nm Sandybridge Core i7. If AMD had 32nm and 22nm at the same time(within 6 months) as Intel, things would be completely different.

Now, overclock each to the best stable and 24x7 usable OC and rerun these tests. Who would win?

People's point is that AMD's flagship is topped to the hilt on die space, power, and usable clock and you can barely call it competitive. If used by enthusiast who do enjoy squeezing more via an OC, the 2600k is hands down the better processor. AMD even acknowledges this fact as reflected in pricing.

 

[boxleitnerb]

You dont buy the Core i7 unless you want Multithreaded performance, same goes for the FX8350. In games at 1080p they are almost the same in the majority of the new games(yes yes SC2 is an exception).

If you only need multithreading performance, you can go for the 8350 if you don't mind the higher power consumption and if you don't care for lightly threaded performance. Unfortunately, there is still alot of software out there that is not multithreaded. If you run a mix of these applications, the 3770K is the better choice.
And no, they are not. You have proven often enough that you have absolutely no clue about benchmarking games. Let it rest.

 

[cytg111]

If you only need multithreading performance, you can go for the 8350 if you don't mind the higher power consumption.
And no, they are not. You have proven often enough that you have absolutely no clue about benchmarking games. Let it rest.

I cant find fault in his reasoning?

 

[AtenRa]

Oh dear, what are you trying to prove here? Putting a 346 mm^2 125W server chip against a 216mm^2 95W mainstream chip with an IGP? Are you really serious?

2600K has 8MB of cache and FX8350 has 16MB of cache and as a consumer i dont care even if the die size is the size of Crete Island.