ExtremeTech: AMD Bulldozer FX pricing revealed: a lot cheaper than Sandy Bridge

[BlueBlazer]

Giorgio Tsoukalos bad hair day (and tan).

 

[LOL_Wut_Axel]

It's not 100% certainty, but all the rumors are pointing that way. It's inconceivable to me to imagine that AMD has an 8 core CPU with Nehalem like IPC, superior power consumption, and greater overclocking that SB, and all that for $266. Consider me "pessimistic" because I was here during Phenom I launch and the same hype followed it too.

This. If it were $320-350 for the 8-core and lower-than-3.5GHz clock speed, then I'd believe it. For $265, no way.

 

[inf64]

IMHO that could be IPC issues (or perhaps bugs). If you look at the Cinebench scores (both R10 and R11.5), it performs somewhere around 4.7 to 5.2 multi-CPU speedup given 4 modules (or 4 real cores). That's greater than 4 real cores (additional 18% to 30% performance per core, ala SMT). The IPC in Bulldozer may not be clear cut (due to longer pipelines), just like this example >> Pentium 4 EE 3.46GHz vs Athlon 64 3400+. Thus one of the main reasons the need for higher clock speed. :hmm:

8C Zambezi has 4 FLeXFP units which work in SMT mode (internally handling 2 threads via SMT). So speed up of around ~4.5-5x is expected. The problem however is the performance of one FlexFP. The brand new 8C 8150 can't beat Thuban. This is going against even what AMD stated for HPC workloads when Interlagos is in question. They stated: 35% more performance versus top MC we have today. Top Interlagos will be 2.3Ghz at launch. Correct for clock speed and you are left with : 1.35x2.5/2.3=1.46 speedup versus MC at same clock(Turbo won't kick in FPU heavy workloads). This is spec fp rate.Note that BD has peak flops that are the same for all 3 ISA targets : legacy SSE,AVX 128 and AVX256 .This is since you have fixed amount of SIMD pipelines that handle both AVX 128 and legacy SIMD,while AVX 256 is done via 2 units(so same peak flops again).
Now we know that 8150 will be running at 3.6Ghz in Cinebech,right? That is 3.6/3.3=1.09 or 9% faster than 1100T. All combined : 1.46x1.09xThuban's C11.5 score = 9.41 pts. This is huge difference to what leaks and guys who allegedly have seen NDA'd docs say -they say around 6pts for 8150 which is practically the same as Thuban.

 

[Vesku]

If those "leaks" showing 6ish Cinebench 11.5 for the 8150 are remotely true then AMD would be better off leaving BD gen 1 as server only and launching for desktop with gen 2. However, I don't think Cray would have accepted shipment of Interlagos if it performed as you suggest inf64. Perhaps programs need a targeted compile to see the full gains from MagnyCours to Interlagos.

 

[BlueBlazer]

8C Zambezi has 4 FLeXFP units which work in SMT mode (internally handling 2 threads via SMT). So speed up of around ~4.5-5x is expected. The problem however is the performance of one FlexFP. The brand new 8C 8150 can't beat Thuban. This is going against even what AMD stated for HPC workloads when Interlagos is in question. They stated: 35% more performance versus top MC we have today.

The figure is for servers and HPC. Quote from my post here....

I think that will depend on application tested. Remember David Kanter's analysis? It might vary, whether its server or desktop type applications. In servers, specific compiler optimizations can used for tuning. However in desktop, applications are already pre-compiled (fixed). Thus would be interesting to see how this turns out. :hmm:

..... and here.....

You may notice he keeps insisting about being a server guy, and in servers you can use compiler tuning to optimize (for Bulldozer pipleline architecture) and to enable new features in Bulldozer (such as AES, AVX and FMA4) that older generation Magny-Cours doesn't have, to claim increased performance. However on the desktop, the situation may be different. The programs are pre-compiled (fixed) and "run-as-it-is". You can also check JF-AMD's FAQ especially these lines....

Final OS optimizaitons
Final drivers
An app compiled with the latest flags

Perhaps you can start to see the difference... :hmm:

Top Interlagos will be 2.3Ghz at launch. Correct for clock speed and you are left with : 1.35x2.5/2.3=1.46 speedup versus MC at same clock(Turbo won't kick in FPU heavy workloads). This is spec fp rate.Note that BD has peak flops that are the same for all 3 ISA targets : legacy SSE,AVX 128 and AVX256 .This is since you have fixed amount of SIMD pipelines that handle both AVX 128 and legacy SIMD,while AVX 256 is done via 2 units(so same peak flops again).
Now we know that 8150 will be running at 3.6Ghz in Cinebech,right? That is 3.6/3.3=1.09 or 9% faster than 1100T. All combined : 1.46x1.09xThuban's C11.5 score = 9.41 pts. This is huge difference to what leaks and guys who allegedly have seen NDA'd docs say -they say around 6pts for 8150 which is practically the same as Thuban.

That may not be accurate because its possible Bulldozer's performance behaviour changes when the clocks gets higher. Quote from my post here...

The scores seems to get lower as the clock frequency increases (when compared, taking into account the number of cores and clock speed). :\

You calculate yourself the slow down, taking into account the actual clock speeds shown in the SiSoftware results (2.0GHz and 2.75GHz). :hmm:

 

[inf64]

AMD actually said 35% better than top MC today. This implies 10% better "per core" and per clock than Magny Cours (2.5Ghz MC vs 2.3Ghz Interlagos). 10% is not bad ,it's good actually. AVX/FMA benefits come on top of this,via recompiled software.

edit:
@Blueblazer
"possible Bulldozer's performance behaviour changes when the clocks gets higher."
Are you serious here? This won't happen unless the multi clock domains are run asynchronously inside the module and one FlexFP then runs at half the clock or something along those lines. I seriously doubt that.

 

[BlueBlazer]

@Blueblazer
"possible Bulldozer's performance behaviour changes when the clocks gets higher."
Are you serious here? This won't happen unless the multi clock domains are run asynchronously inside the module and one FlexFP then runs at half the clock or something along those lines. I seriously doubt that.

I do speculate that perhaps some sections/parts of the module as well as the rest of the CPU may be running at limited clock speed to keep the TDP in check. And running each section with different clock speeds and communicating asynchronously does have its ups and downs. This will depend on the engineering to minimize latency and lost/missed clock cycles. Setting the proper frequency, timing and clock phases at which these domains run together improves performance. Setting them improperly can decrease performance. :hmm:

 

[garagisti]

http://semiaccurate.com/2011/09/13/amd-sets-clock-speed-world-record-with-bulldozer/

 

[BlueBlazer]

http://semiaccurate.com/2011/09/13/amd-sets-clock-speed-world-record-with-bulldozer/

Redundant post, by the way.

 

[Khato]

@Blueblazer
"possible Bulldozer's performance behaviour changes when the clocks gets higher."
Are you serious here? This won't happen unless the multi clock domains are run asynchronously inside the module and one FlexFP then runs at half the clock or something along those lines. I seriously doubt that.

We already know from one of AMD's previous Bulldozer presentations that there are multiple clock domains running at different frequencies within a module - they did a demonstration in early June where they showed AMD OverDrive of all 8 cores with different frequencies. So we know that the reported frequency is tied to the integer core. What we don't know is what frequency the rest of the functional logic runs at - is it a static frequency? Is it the same as the fastest integer core in the module? Thus far, I've seen zero evidence either way.

 

[garagisti]

Redundant post, by the way.

You didn't read it. It gives info specifically on the fact that 8.3 ghz achieved by Intel chip was a cache-less one. Here you have 2 cores with 8 mb of L3 hitting that speed. Significant difference this. Even if you point out to 7 ghz by 990X, you can't discount another ghz on top of it, will you now?

 

[Arkaign]

You didn't read it. It gives info specifically on the fact that 8.3 ghz achieved by Intel chip was a cache-less one. Here you have 2 cores with 8 mb of L3 hitting that speed. Significant difference this. Even if you point out to 7 ghz by 990X, you can't discount another ghz on top of it, will you now?

Who cares? This isn't relevant, neither is the 7ghz 990x. The 8.3Ghz Celeron is ridiculous because it's what, 8+ year old tech or even older (netburst). Mainboards and memory back in the netburst days were old school as well. Add to that the BD 'modules' aren't precisely two cores anyway. With 32nm and a focus on clockspeed getting high overclocks would be what you'd expect.

If Intel overclocked an Atom to 20ghz, would that impress you? How about if AMD made a 40ghz 386? They would still suck.

 

[ransomlist]

This is marketing- we have yet to see real proof it translates to average joe/enthusiast gains in performance and we probably wont until said groups have retail chips in hand.

 

[BlueBlazer]

You didn't read it. It gives info specifically on the fact that 8.3 ghz achieved by Intel chip was a cache-less one. Here you have 2 cores with 8 mb of L3 hitting that speed.

Do you realize that 8MB of L3 cache means half of its 16MB L3 cache is disabled? And as mentioned in other thread, Bulldozer "cores" are not true cores. The fact is that only 1 "module" is enabled. That's why you do not see 1-core, 3-core, 5-core, 7-core SKUs. Most of us know about that Celeron already, long ago.

Significant difference this. Even if you point out to 7 ghz by 990X, you can't discount another ghz on top of it, will you now?

That's 7GHz with all 6 cores running, not with cores disabled. You are also starting to derail this thread. :hmm:

 

[RobertPters77]

Glad I dumped all of my amd stock months ago.

 

[Vesku]

Do you realize that 8MB of L3 cache means half of its 16MB L3 cache is disabled? And as mentioned in other thread, Bulldozer "cores" are not true cores. The fact is that only 1 "module" is enabled. That's why you do not see 1-core, 3-core, 5-core, 7-core SKUs. Most of us know about that Celeron already, long ago.

That's 7GHz with all 6 cores running, not with cores disabled. You are also starting to derail this thread. :hmm:

It's 8MB L3 + 2MB L2 x 1 module = 10MB total cache of 16MB of the full 4 module.

 

[Bearach]

Do you realize that 8MB of L3 cache means half of its 16MB L3 cache is disabled?

Codename: Zambezi
Process Technology: 32 nm SOI
Cores: 8
Core Speed: 3.6
Turbo Speed: 4.2
L2 Cache: 8MB
L3 Cache: 8MB
Black Edition: Yes
Memory Support: Dual Channel DDR3-1866
Socket: AM3+
TDP: 125 watt

I thought this was the specifications of the 8150. Could you link me this new information that shows they now have doubled the L3 cache?

EDIT: Looks like Vesku brought it up before me.

 

[Arachnotronic]

Benchmarks. NOW!

 

[Vesku]

If only we were getting release grade benchmarks, the desire to pour over tech site reviews is building up! Hopefully Anandtech or similarly reputable site compares the 4, 6, 8 starter clocks, 8 max clocks all in one go within a few days of NDA lift.

 

[choliscott]

I'm trying to remember, but wasn't the NDA lift on the 6 cores 1 or 2 days before release day.

If only we were getting release grade benchmarks, the desire to pour over tech site reviews is building up! Hopefully Anandtech or similarly reputable site compares the 4, 6, 8 starter clocks, 8 max clocks all in one go within a few days of NDA lift.

 

[BlueBlazer]

I thought this was the specifications of the 8150. Could you link me this new information that shows they now have doubled the L3 cache?

My bad (should be "cache" instead of "L3 cache"), and Vesku is correct. Must have Interlagos in my mind.

If only we were getting release grade benchmarks, the desire to pour over tech site reviews is building up! Hopefully Anandtech or similarly reputable site compares the 4, 6, 8 starter clocks, 8 max clocks all in one go within a few days of NDA lift.

Lack of previews at the moment. However it seems systems are coming >> iBUYPOWER Launches World's Fastest AMD FX-8150 "Bulldozer" System

 

[psolord]

This. If it were $320-350 for the 8-core and lower-than-3.5GHz clock speed, then I'd believe it. For $265, no way.

The 5870 cost 200$ less than the GTX 480 when it first came out, while delivering 85% of the performance. What was so bad about it?

Now that AMD has shaped up, the 6000 series is priced completely adjacent to Nvidia's prices though. Ergo, when you are the underdog, you produce a very decent product and shove the price down your opponent's throat. What's so improbable about it?

 

[BlueBlazer]

The 5870 cost 200$ less than the GTX 480 when it first came out, while delivering 85% of the performance. What was so bad about it?

Now that AMD has shaped up, the 6000 series is priced completely adjacent to Nvidia's prices though. Ergo, when you are the underdog, you produce a very decent product and shove the price down your opponent's throat. What's so improbable about it?

That's the GPU market and AMD's profit here is razor thin (majority of AMD's profit still stem from CPU market). Just look at NVIDIA's higher profits due to charging higher prices. Sometimes too much under-cutting have negative impacts. Its not a question of being an underdog, but its a question of business sense. :hmm:

 

[Anarchist420]

I always expected BD to be a flop. By May, it was already 4 months later than SB, I had read that its TDP was 125W, and the L3 cache is clocked ridiculously slow which is what held the PII's performance back to a large degree.

I expect AMD to close their doors by years end and they deserve to for making lousy products and for being a product of the state (protected by antitrust and IP laws). Same goes for Microsoft, although their products aren't quite as lousy.

 

[Anarchist420]

The 5870 cost 200$ less than the GTX 480 when it first came out, while delivering 85% of the performance. What was so bad about it?

Now that AMD has shaped up, the 6000 series is priced completely adjacent to Nvidia's prices though. Ergo, when you are the underdog, you produce a very decent product and shove the price down your opponent's throat. What's so improbable about it?

You're only taking performance into account. Nvidia's IQ, extra features, and in my opinion, their drivers are superior. Besides, people should know generally not to get the first nvidia product of its generation, mainly because of all the revisions. While the 8800GTX was kind of an exception, the Geforce FX 5900 ultra was better than the 5800Ultra. I agree that the 480GTX was a rip off, but then so was AMD's offering regardless of how inexpensive it was. Once nvidia gets their GL frame rates under control (if they haven't already), there will be no reason to buy AMD.