AMD - Bulldozer Chitchat

[DanceswithUnix]

With anything above 4 cylinder, the power strokes overlap so you don't get the same pulsating delivery of power you get with 4's (or less) so should be significantly smoother. They're not inherently balanced like a 6 but vibrate less than a comparable 4, especially at low-medium revs. But if you're flooring an engine, vibration wouldn't be top of my list.

Big fan of flat-4 myself.

My first Alfa was a 33 with that revvy boxer 4.

Only 5 pot I have driven is a diesel, but for me if you want smooth you want at least 6 cylinders, as many as you can stuff in the engine bay. If you want economy then get 3 or 4 cylinders. If you want 5, just stop sitting on the fence and make your mind up

Getting a tad off topic here, BD slow news day

[Golden Dragoon]

Seems that way
Sooooo nice weather we are having right now

[kalniel]

What engines do bulldozers have?

I remember the jet powered JCB, that was cool.

(see what I did there? )

[JF-AMD]

*shrug* I'll happily give you that it's not as good as the Xeon. But, since you've accepted that they're not incredibly slow, to also call then "rubbish" at virtualisation seems unfair too. They're actually perfectly acceptable at virtualisation. The thing is the Xeon's are better than acceptable, at about the same price. That makes the Opterons poor value, but it doesn't make them "rubbish".

I think we've all accepted that Bulldozer failed to acheive it's target of making AMD properly competitive again, but branding it "rubbish" and "useless" seems unnecessarily harsh. An FX4100 has more than enough power to meet the computing needs of 99% of computer users. It's not necessarily the best choice, but that makes it neither "rubbish" nor "useless".

As you say, it looks like Trinity might be a decent improvement, and if it is that bodes well for the future Piledriver-based server CPUs. It's rare for a new architecture to be a massive improvement over the incumbent (think Netburst P4!) and all the signs are that AMD can address the deficits in bulldozer and make it competitive in the future. How effectively they do that only time will tell, but they have a decent track record of architecture tweaks in the recent past (i.e. Phenom I/II) and given that bulldozer already excels in a (very!) small number of workloads (something Phenom I didn't) they may yet produce something special.

I will take the counter to your "at about the same price" argument.

Opteron 6282 is ~$1000USD
Xeon E5-2690 is ~$2000USD

Unfortunately, it is not 100% faster.

I had one of the guys on my team do a pricing build out to see what the net difference was at the platform level. Building a Dell PowerEdge R720 with top bin procs, 32GB of memory and the basic configurations was ~$10,000USD. The exact same configuration on a PowerEdge R715 (AMD-based) was ~$5700USD.

So, while it is faster, would you pay almost 2X more for your server if you were not getting 2X the performance?

In reality, most customers see performance as being "good enough" on almost all of today's servers. Most customers do not need the performance and are far more interested in their budget. Look at top bin processor sales, less than 5% are top bin, 95% are down the stack. For Westmere, almost 70% of their sales were in the bottom 3 SKUs. Performance is a red herring in most cases.

[badass]

I will take the counter to your "at about the same price" argument.

Opteron 6282 is ~$1000USD
Xeon E5-2690 is ~$2000USD

Unfortunately, it is not 100% faster.

I had one of the guys on my team do a pricing build out to see what the net difference was at the platform level. Building a Dell PowerEdge R720 with top bin procs, 32GB of memory and the basic configurations was ~$10,000USD. The exact same configuration on a PowerEdge R715 (AMD-based) was ~$5700USD.

So, while it is faster, would you pay almost 2X more for your server if you were not getting 2X the performance?

In reality, most customers see performance as being "good enough" on almost all of today's servers. Most customers do not need the performance and are far more interested in their budget. Look at top bin processor sales, less than 5% are top bin, 95% are down the stack. For Westmere, almost 70% of their sales were in the bottom 3 SKUs. Performance is a red herring in most cases.

How about comparing the prices of a similar performing Intel system to the AMD system instead of picking the much faster Intel system and then pointing out it's much more expensive.
When have you ever got twice as much performance for twice the price at the high end of IT ever?

Seriously, it's like you're comparing Intel apples to AMD oranges then saying the AMD oranges are more citric acidy so you should buy the AMD oranges, despite the fact that Intel also sells oranges and for less money.

If the AMD system is good enough then a cheaper Intel system with similar processing power is also good enough.
I've just been on the Dell configurator and compared bare R720's and R715's.

Trying to compare Apples to apples as much as possible, I came up with the following configurations:
R715:
32GB UDIMMs (cheaper than RDIMMS)
2xOpteron 6274 - the 6276 is an extra £190 and the 6272 is £140 cheaper.
No HDD's, cheapest RAID controller etc.
£2614

R720
32GB UDIMMs (RDIMMS not available)
2xXeon 2620 - 2x2630 is £360 more
No HDD's, cheapest RAID controller etc.
£2356

The faster Intel system is cheaper. I can make it more expensive, but then it just keeps getting faster.
So, when you compare like for like, AMD opteron CPU's are now also very bad value as well as being poor performers.

As I said before, the only sane reason to go for the Opteron system is to add it to a cluster that already uses AMD CPU's.

In case you haven't guessed, I really take issue with this kind of straw man argument. It makes you look like you're deliberately being misleading.

[JF-AMD]

Can you show me the performance of the E5-2620? I think there are only 2 fair comparisons that you really should do:

1. Same performance, what is the price
2. Same price, what is the performance

My comparison is based on someone calling out intel performance being "so much better" but not taking price into consideration. They used their fastest processor (E5-2690) for their benchmarks, so chose our top bin to compare. Ultimately that isn't how customers do this. Either there is a performance level that they need to hit for the app (not as common) and they want to know the price to hit that level, or (more common) we have XX in our budget, what can I get for that.

unfortunately there is no SPEC benchmarks for the broader stack of processors.

[CAT-THE-FIFTH]

Some interesting news about Kaveri:

http://www.nordichardware.com/news/6...n-hd-7750.html

http://h9.abload.de/img/kaveri_slidetualp.png

The GPU section seems to be based on Cape Verde Pro.

[watercooled]

Very interesting! Unified memory should also help with migrating code to the GPU section/another step towards blurring the line between the two. I'm sure AMD have considered another memory channel, we'll have to wait and see if they've found a way around it.

[scaryjim]

... we'll have to wait and see if they've found a way around it.

Desktop DDR5 / QDR memory? One has to assume not, since there's not even the vaguest move away from DDR3 at the minute, but it'd make more sense if they're going to push the GPU section up to a genuine mid-range configuration. Or they could stick one of their famous hybrid memory controllers in it, so it'll run in existing boards with DDR3 and in new boards with whatever the next desktop memory standard ends up being...

[watercooled]

Well DDR4 probably isn't far off, mass production in H2 2012 IIRC.

[DanceswithUnix]

Sideport FTW

Or dram stacked onto the CPU die for mega bandwidth.

[badass]

Can you show me the performance of the E5-2620? I think there are only 2 fair comparisons that you really should do:

1. Same performance, what is the price
2. Same price, what is the performance

Which is exactly my argument.
Same performance, Xeon is cheaper
Same price, Xeon is faster.

This was demonstrated on my previous post. The Xeon E5-2620 is faster than the top bin Opteron and cheaper than the mid range one.

My comparison is based on someone calling out intel performance being "so much better" but not taking price into consideration. They used their fastest processor (E5-2690) for their benchmarks, so chose our top bin to compare. Ultimately that isn't how customers do this. Either there is a performance level that they need to hit for the app (not as common) and they want to know the price to hit that level, or (more common) we have XX in our budget, what can I get for that.

unfortunately there is no SPEC benchmarks for the broader stack of processors.

The comparisons are linked by me earlier in this thread. They are on Anandtech.

[CAT-THE-FIFTH]

The comparisons are linked by me earlier in this thread. They are on Anandtech.

Actually not entirely true. They used a simulated E5-2630 with 33% more L3 cache - both me and scaryjim pointed this out to you. Being objective,I cannot say whether the 100MHZ clockspeed decrease they implemented will compensate for the increase in L3 cache.

[badass]

Very interesting! Unified memory should also help with migrating code to the GPU section/another step towards blurring the line between the two. I'm sure AMD have considered another memory channel, we'll have to wait and see if they've found a way around it.

It may well also make coding a lot easier. It will eliminate having to copy/move data from the CPU memory to the GPU memory. I can see that resulting in a tangible performance improvement in a bandwidth constrained environment.

Desktop DDR5 / QDR memory? One has to assume not, since there's not even the vaguest move away from DDR3 at the minute, but it'd make more sense if they're going to push the GPU section up to a genuine mid-range configuration. Or they could stick one of their famous hybrid memory controllers in it, so it'll run in existing boards with DDR3 and in new boards with whatever the next desktop memory standard ends up being...

The tech from GDDR5 will be incorporated into DDR4 for desktops. IIRC DDR3 for desktops is more like GDDR4 than GDDR3

Well DDR4 probably isn't far off, mass production in H2 2012 IIRC.

Hopefully. However it'll probably be 2014 or later before price parity with DDR3.
However, if memory remains stupidly cheap, even if DDR4 is more expensive, it'll be worth it if bandwidth still constrains the APU.

Sideport FTW

Or dram stacked onto the CPU die for mega bandwidth.

That could really, really boost performance. IIRC the Xbox 360 has some kind of memory either on package or on die and it effectively gives them AA without any performance penalty.
The right size and right bandwidth/latency of on package memory once again could really boost performance if it's bandwidth constrained.
Perhaps this won't be necessary with unified memory and/or DDR4

Just done a quick check. Kaveri looks like it's socket FM2 http://www.fudzilla.com/home/item/26...orts-ddr3-2133
Looks like no DDR4 but if it doesn't need the bandwidth of DDR4 then it's pretty moot.
I'm just glad it's socket FM2. From my point if view, if Trinity isn't fast enough to be able to take anything I can throw at it in 5/6 years time, I can anticipate that an APU only upgrade for better performance will be possible later.

[CAT-THE-FIFTH]

Haswell is rumoured to have on-die stacked RAM for the GPU and it seems that AMD might be also working on something similar too.

SA,leaked this picture last year:

http://semiaccurate.com/2011/10/27/a...-gpu-pictured/



Some more information:

http://semiaccurate.com/forums/showp...8&postcount=26

[watercooled]

The tech from GDDR5 will be incorporated into DDR4 for desktops. IIRC DDR3 for desktops is more like GDDR4 than GDDR3

Yeah, DDR3 != GDDR3, DDr4 != GDDR4, they're very separate technologies, sharing just the number, despite that they're often incorrectly used interchangeably.

That could really, really boost performance. IIRC the Xbox 360 has some kind of memory either on package or on die and it effectively gives them AA without any performance penalty.
The right size and right bandwidth/latency of on package memory once again could really boost performance if it's bandwidth constrained.
Perhaps this won't be necessary with unified memory and/or DDR4

Yeah 360 uses 10MiB of very fast eDRAM. It would be really interesting if they attempted something like that, or some sort of on-die local store. Very interesting time in the industry.

Edit: beaten to it by CAT. How do you find this stuff? xD