AMD ''Interlagos'' Bulldozer Benchmarks Leaked

Page 3 - Seeking answers? Join the Tom's Hardware community: where nearly two million members share solutions and discuss the latest tech.
Toggle sidebar Toggle sidebar
Status
Not open for further replies.


Yeah, I built it about 6 months ago and I'm pretty happy with it. My only regret is that it's not a i5-2500K and SLIed GTX 560's 😛 The P55 is quite a nice platform, albeit slightly lacking on PCIe capability. It's kind of a shame that Intel had to replace it so quickly. The i5-750/760s were right in the sweet spot, no Hyper-threading which games don't use to increase cost or heat, but with 4 cores, plenty of room to overclock and a reasonable price tag. The i5-2500K seems to be following in it's footsteps.
 
jprahman writes:
> only regret is that it's not a i5-2500K and SLIed GTX 560's 😛 ...

😀😀


> ... The P55
> is quite a nice platform, albeit slightly lacking on PCIe capability.

I've done a lot of tests with the Asrock P55 Deluxe boards I obtained
(excellent slot layout btw, far better than most, room for two cards
SLI/CF with an extra empty slot between, plus a 3rd slot for other
stuff - mine has a 3041E SAS card), the situation re PCIe bw is nowhere
near as critical as many would have one believe. I have a friend who
has an i7 930 system with two Sonic 460 Platinums (slightly lower core
clock), my P55 beats his setup everytime (his ASUS mbd cost 267 UKP, my
P55 cost 75 UKP). Sadly, availability of boards like this is often
patchy. It's not made anymore.

Asrock did make a nice combo board (P67 Transformer) that has a P67
chipset (so full PCIe speed) with socket 1156 to take the older
i3/i5/i7s, but it only has one PCIe slot (doh!). Nice idea, poorly
realised. If it had been given the same layout as the P55 Deluxe it
would have been a good staging upgrade for those who'd already invested
in a P55 and wanted to retain their CPU.


> It's kind of a shame that Intel had to replace it so quickly. ...

Indeed. But who knows, perhaps it was "too good"? Look at how often the
860/870 kept beating the 920 in 'normal' tests because of its better
turbo, and of course the lower power consumption, etc. My tests prove
X58 isn't necessary even for 2 gfx cards to achieve good performance,
but lots of people bought X58 convinced it was essential for more than
one GPU. Having 3+ GPUs is where X58 makes sense, or those who need
more than 4 cores, or other high bw requirements. Even then though,
I get 700MB/sec with just 4 SAS disks on my 75-quid P55 board. 😀 Not
tried with more than 4 yet.


> ... The
> i5-750/760s were right in the sweet spot, no Hyper-threading which
> games don't use ...

Yup, I've not found any 3D tests that usefully gain from HT so far. Turning
it off dropped temps by 10C and allowed my 870 to reach 4444 with ease, as
opposed to 4270 with it on. I'm sure it could go a lot higher, but don't
want to risk it. With the 750/760 not having HT, they're perfect for gaming.
I'm building a 760/GTX460 (Sonic) for my brother, using an Asrock P55 Extrme
(same slot layout as the Deluxe, so he can add a 2nd card later without
worrying about cooling issues).


> ... to increase cost or heat, but with 4 cores, plenty of
> room to overclock and a reasonable price tag. The i5-2500K seems to be
> following in it's footsteps.

Certainly looks like it!


Anyway, back to AMD. I just hope they can get back in the game. A few
years ago when I was looking for an upgrade, it just happened to be the
very week AMD slashed prices by 50% on many top-end Athlon64 X2 models.
The 6000+ was suddenly only 156, massively less then an equivalent Intel
at that point, so the 6000+ was a bargain by comparison. It worked really
well. A lot has happened since then of course, but it's just sad now that
AMD simply has no real performance option available for the desktop; kinda
shocking that the 1100T often fails to beat even the old i7 870;
big price difference of course right now, but SB takes care of that (the
2500K costs 10 UKP less but is crazy amounts quicker). Indeed, when the
750 came out I was amazed to see it matching or beating the Phenom2 965 BE.

Hmm, maybe what AMD is planning is to use the server market to gain an
edge which will then give it a revenue pool to pull back in the desktop
stakes later. Presumably Opteron is where the money is for them atm, yes?
I expect the many-cores approach will be very attractive to a wide
variety of HPC/Enterprise customers, what with virtualisation being all
the rage, etc.

Ian.

PS. Looking forward to seeing what the EVGA P55 FTW can do. 8)

 
kinda
shocking that the 1100T often fails to beat even the old i7 870;
Click to expand...
I've noticed that too. The problem is that the K10 architecture just can't stand up to Nehalem's much higher IPC, even with 50% more cores. Sandy Bridge widened the gap even further. I have to agree that for the past two years Intel has pretty much dominated the high end desktop market, with AMD being able to offer good enough performance at a low price. Unfortunately the performance gap has widened with Sandy Bridge and while Intel is still more expensive the performance disparity is becoming harder to justify even in the light of lower prices on AMD processors.


It certainly does seem as if Bulldozer is somewhat more focused on the server market. I mean looking at the architecture as a whole it looks to be more focused on packing more cores on one chip, rather than trying to squeeze maximum performance out of each core, ala Sandy Bridge. For example, each pair of cores share a common floating point unit, showing that the designers, while not completely neglecting floating point capabilities, certainly want to focus on integer performance more. To me at least in a server setting Bulldozer looks quite attractive, and I would seriously consider getting a Bulldozer based server over a Xeon because of higher core counts and presumably lower cost. Obviously we won't know exactly how Bulldozer stands up to Intel's Xeon platform until June.


 
@jprahman

and how many desktop applications you know requires 256bit FLOP, i think 2 x 128bit FLOP per cycle is more then enough, for raw integer crunching the fusion approach makes more sense seeing as a stream processor can be used to crunch integer and there is odds on over 100 stream processor per GPU
 
I don't think that having both cores use the same floating point unit will cause any problems for 99% of applications, I point it out as an example of how AMD is trying to maximize the core count by changing the fundamental architecture of x86 CPU cores. Overall it seems like a major goal of Bulldozer is high core counts.
 
Seems like a lot of people are running wild with this data from some engineering sample. I am curious, historically, how accurate a picture do engineering samples play in all this? I know the core count is way up but I was still thinking the clock speed would be a bit higher. Though being a server part I grant that they need to fit things in very specific TDP envelopes. The desktop version will probably be at least 3.5GHz though.
 
It's worth noting that AMD's FP design can process either 2x128-bit SSE ops or 1x256 AVX op per clock. I've done a bit of checking and Sandy Bridge can apparently do either 1x128 or 1x256 (and power-gates the upper 128). So, from what I can see, AMD is gambling that AVX will have a slow uptake but have offered an (reduced) ability for their processors to handle it just in case.

Any setup is only as good as the front end so I'm hoping that AMD have gotten that right.
 
I just want to make a note for everyone saying that it isn't doing as well as the high end xeon.....this is just the engineering model that is only at 1.8ghz.......the final product to hit the market will be clocked arround 3.4ghz!!!!! DO THE MATH!!!! If the full market version of the xeon is doing only 13 some seconds better at its full speed, and this is only at 1.8ghz....what do you think it will be doing at 3.4ghz, and then again when the turbo core kicks in, and yes the server model has the turbo core as well...all the things going into the chip are posted on AMD's website. Also right now, the performance of the I7 compared to the xeon is better, and this is why they are comparing it to those models, also the primary architecture of the desktop version is the exact same, but will have lower L3's, and maybe lower L2's, so this lower clock speed just might be a touch less for overall performance for the multithreaded processes.
 


True, but that 2x128 SSE or 1x256 AVX operation per clock figure is for every PAIR of Bulldozer cores, while the Sandy Bridge 1x128 SSE or 1x256 AVX operations per clock figure is per SINGLE core. So Sandy Bridge, theoretically has twice the AVX capability and the same SSE capability, per core, although AVX isn't widely used, so this likely won't be a big deal for a while. Obviously the final benchmarks will tell the full story, and so all of this is just theory.
 
Yes but it's also worth pointing out that "every PAIR of Bulldozer cores" is a module, whereas "is per SINGLE core" is exactly that. It's only Bulldozer's aggressive implementation of SMT, and AMD's subsequent classification of both integer units as "cores", that really differentiates itself from Intel's HyperThreading. Remember that HyperThreading doesn't duplicate the execution resources of a core yet a Bulldozer module does, albeit at the cost of a small increase in die size. HyperThreading also results in a small increase in die size, obviously with a different implementation.

Had AMD not classed the integer cores in a Bulldozer module as separate entities within themselves, I think many people would've been happier to treat them as a 1 core, 2 thread setup. For FP, it still is, I suppose.

I didn't state it at the time, but my comparison was between a Bulldozer module and a Sandy Bridge core. Seeing as though Orochi will be a four-module CPU, it would, to me, be comparable to a four-core Sandy Bridge - both can handle 8 threads at once. Technically, a single module Bulldozer offering wouldn't really have a direct competitor as AMD have already said a Bulldozer "core" would be weaker than a single Stars core, however it'd still likely outpace anything like Brazos or Atom.
 
[citation][nom]silverblue[/nom]Yes but it's also worth pointing out that "every PAIR of Bulldozer cores" is a module, whereas "is per SINGLE core" is exactly that. It's only Bulldozer's aggressive implementation of SMT, and AMD's subsequent classification of both integer units as "cores", that really differentiates itself from Intel's HyperThreading. Remember that HyperThreading doesn't duplicate the execution resources of a core yet a Bulldozer module does, albeit at the cost of a small increase in die size. HyperThreading also results in a small increase in die size, obviously with a different implementation. Had AMD not classed the integer cores in a Bulldozer module as separate entities within themselves, I think many people would've been happier to treat them as a 1 core, 2 thread setup. For FP, it still is, I suppose.I didn't state it at the time, but my comparison was between a Bulldozer module and a Sandy Bridge core. Seeing as though Orochi will be a four-module CPU, it would, to me, be comparable to a four-core Sandy Bridge - both can handle 8 threads at once. Technically, a single module Bulldozer offering wouldn't really have a direct competitor as AMD have already said a Bulldozer "core" would be weaker than a single Stars core, however it'd still likely outpace anything like Brazos or Atom.[/citation]
Well, there was a time when a "core" had no floating point resources at all, and I don't think you would not still class that as a core. Therefore, it is the integer execution units that define a core (and are all that are used in a majority of workloads), therefore it is fair to define a BD module as two cores. It has two int cores, so two cores. Also, servers barely use fp, so it is definitely not 1 module = 1 core there – it is very much two cores in that situation/workload.
 
Depends on the usage. If you're only using SSE, then yes. If you are using a lot of AVX, then no, but the only workloads that would use that would be graphics (2D, 3D (not real-time rendering), video incl. playback) audio production and scientific stuff. Video playback isn't a problem. Plenty of video encoding is done with the GPU. 2D graphics wouldn't use it that much, and the speed difference wouldn't be much. Anything else is actually more workstation-type workloads, so desktop workstation rather than normal desktop usage. You want my opinion? BD's 1 module = 2 cores will probably hold up very well with all but workstation workloads. That is my opinion, and it could be very wrong, and if it is, please tell me.
 
You're probably right, but we won't really know for a while. In any case, you sound more qualified to answer that than I am; I'm just interested in its potential performance and don't really possess much more than that.
 
I agree that this comparison is really stupid. However, I don't think that bulldozer will be a wimp in single threaded performance. Remember these are 1.8 GHz processors, so naturally the single threaded performance will not be as good. When only a few threads are running, each bulldozer module performs as a very fast single CPU. When more threads are running, each module becomes a dual core CPU that operates at about 160% of the relative performance. So both single-threaded and multi-threaded performance should be decent.
 
I think what the first person meant by xeon vs opteron bulldozer benchmarks was that xeon X series for instance work on dual QPI which is the undeniable strength and flag of xeon X series; therefore I agree that I would prefer to see benchmarks against Xeon series rather than mass market.
We all know that clock speed is important but clock speed and cache do not suffice in dual socket war between the two giants.
Note: I am only a noob...do not come murder me but after reading pffff 50+ websites on Xeon/SBe/Ivy and AMD opteron comes this understanding.
 
Status
Not open for further replies.

TRENDING THREADS

Latest posts

Moderators online

Share this page

COMPANY
RESOURCES
FOLLOW
Top Bottom