Discussion Forums > Technology
Your view on AMD's Bulldozer
kureshii:
A couple more points that popped into mind (and more bored web-browsing at work).
Re-reading the FX-8150 reviews that matter again, I think Anand sums up AMD's position pretty well:
--- Quote from: http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/11 ---The good news is AMD has a very aggressive roadmap ahead of itself; here's hoping it will be able to execute against it. We all need AMD to succeed. We've seen what happens without a strong AMD as a competitor. We get processors that are artificially limited and severe restrictions on overclocking, particularly at the value end of the segment. We're denied choice simply because there's no other alternative. I don't believe Bulldozer is a strong enough alternative to force Intel back into an ultra competitive mode, but we absolutely need it to be that. I have faith that AMD can pull it off, but there's still a lot of progress that needs to be made. AMD can't simply rely on its GPU architecture superiority to sell APUs; it needs to ramp on the x86 side as well—more specifically, AMD needs better single threaded performance. Bulldozer didn't deliver that, and I'm worried that Piledriver alone won't be enough. But if AMD can stick to a yearly cadence and execute well with each iteration, there's hope. It's no longer a question of whether AMD will return to the days of the Athlon 64, it simply must. Otherwise you can kiss choice goodbye.
--- End quote ---
I hope AMD has enough improvements in store for Bulldozer to boost its performance as much, if not more. Few processors make it off the production line with all their intended features in place. Not even Sandy Bridge, with its smooth and timely release (Cougar Point's hiccups notwithstanding), got all its intended features; the rest will come with Ivy Bridge. AMD doesn't have the aggressive tick-tock release schedule of Intel, and instead releases a new, major architecture every few years, following it up with minor revisions/improvements (e.g. Agena -> Deneb -> Thuban for Phenom I/II). We saw some pretty nice performance improvements there, and we're going to have to rely on these architectural improvements on Bulldozer to see enough competition to keep Intel on their toes.
As for price/performance comparisons, here are a few high-end builds and their prices (CPU prices from Newegg except otherwise stated).
1 I take entry-level boards for each chipset (the cheapest product on Newegg) without consideration for board features, such as overclockability; these are commodity features determined by third-party board manufacturers and I'll avoid complicating the price/performance comparison with them.
2 I use the K-parts rather than the non-K ones so as to compare unlocked processors; you can get a better price at the cost of overclockability by going with non-K variants.
3 I'll ignore the 990FX here as I'm not interested in comparing quad-SLI builds, so chipsets with 32X PCIe lanes are out. Its direct Intel match would be the X79.
4 Budget options for AMD are available in the form of 700-series chipsets; from Intel they come in the form of lower-end chipsets (H61). I opted not to post any builds with these as those lack SATA 6Gbps; anyone buying high-end shouldn't be skimping on such minor price differences.
No mid/low-end comparisons are done here as I believe I've posted my analysis of AMD's positions in those segments in earlier posts. I might do so once the quad-core FX processors are benchmarked.
* Prices are tagged by CPU brand
CF/SLI builds (ranked by price):
X6 1055T ($150) + Asrock 870 ($90) = $240
X6 1100T ($190) + Asrock 870 ($90) = $280
FX-8120 ($220) + Asrock 970 ($110) = $330
i5-2500K ($220) + MSI P67 ($115) = $335
FX-8150 ($250 MSRP) + Asrock 970 ($110) = $360
i7-2600K ($315) + MSI P67 ($115) = $430
Single-x16 builds (ranked by price):
X6 1055T ($150) + Biostar 870 ($60) = $210
X6 1100T ($190) + Biostar 870 ($60) = $250
i5-2500K ($220) + Biostar H67 ($70) = $290
FX-8120 ($220) + MSI 970 ($85) = $305
FX-8150 ($250 MSRP) + MSI 970 ($85) = $335
i7-2600K ($315) + Biostar H67 ($70) = $385
[edit] Some changes made in CF/SLI selections. If you read this post before the last edit date-time, please refresh again.
Knowing how the 8150 performs against the 1100T, 2500K and 2600K, I am hard-pressed to call it a "value platform" from those prices.
For non-SLI/CF builds, the i7 offers equivalent real-world multithreaded performance and much better light-threaded performance, as well as lower power consumption/heat, for 10% more.
For gaming builds (SLI/CF), the X6 offers lower multithreaded performance but roughly equivalent gaming performance for much lower cost. The savings can be put into a better graphics setup which is more likely to see better gaming performance.
per:
--- Quote from: kitamesume on October 16, 2011, 11:37:22 AM ---BUT, if they ever find a way to make bulldozer more worth it, then consider me getting one as well, that is, if ivy bridge isnt more worth it than bulldozer/piledriver.
--- End quote ---
From the server CPU side of things I have seen ivy bridge is a lot better than sandy bridge.
A 2x6-core blade with 48Gb RAM (+ssd) uses 120W peak power, and is faster than the old Xeon L. The gain in power efficiency is close to 30% (and this is with early silicon, and without low-power DDR3).
For us this means that we would save about $1M/year on power alone, and power usage costs goes from 50% to ~35% of the total cost over three years (we budget the servers for three years), and space savings since we need fewer nodes would save a bit too.
Compare this to buldoozer, where we would need 3x the power (around 6M), and power density concerns (we only have 20kW/rack or so) would cause us to be forces to use about 3x the space as well. And the CPU:s and chipsets are not even cheaper.
Really? What are they thinking with this CPU release? :o
And in our tests intels hyperthreading actually is _more_ efficient at boosting performance than AMD:s "true core":s. That is the most suprising thing to me.
kureshii:
--- Quote from: per on October 16, 2011, 03:11:38 PM ---And in our tests intels hyperthreading actually is _more_ efficient at boosting performance than AMD:s "true core":s. That is the most suprising thing to me.
--- End quote ---
How was that test conducted? Do you have a link? I really am quite curious, because AMD’s been hyping their Bulldozer Module as a better-than-hyperthreading technology—on paper, of course. But like we’ve seen with the Pentium 4, theories don’t always work out as expected.
Google turns up no comparison of the sort, but looking at this image gives me an idea for how to go about doing this:
Note that the point here is not to see how a quad-core, quad-module Bulldozer benchmarks against an i5, but to see how much of a performance gain hyperthreading gains for Sandy Bridge in various workloads, vs adding a second integer core in each Bulldozer module.
What use is this? It could help isolate Bulldozer’s bottlenecks in various workloads, particularly the single-threaded/lightly-threaded workloads. Those who have been convincingly bought over by AMD’s octo-core marketing might suddenly remember that the only truly separated resources in each module are the integer cores; both cores in each module still share the same fetch/decode units, as well as L2. The high latency of each cache, though necessary for AMD to achieve those high clock speeds, could adversely affect some memory-heavy benchmarks. The branch predictor is another possible bottleneck; Anand hypothesised this from the AIDA64 Queens benchmark, and it could potentially bottleneck other branch-heavy workloads as well.
My current thought is that the decode units might perhaps not be able to keep up both cores fully fed for light-compute workloads (where threads do not spend a lot of time in the integer execution units). Anandtech’s review has a comparison of decode capabilities of Thuban, FX and SB. This does not directly translate to raw instruction decode speed, but results from this test could help examine the results/implications of this change in decode resources, which could explain some of the discrepancies between Thuban and FX performance.
Add to this AMD’s claim that Windows 7’s scheduler is not optimised for a “non-uniform core architecture”, and you should be able to see how this experiment could lead to interesting results.
[edit] Some interesting results: http://www.xtremesystems.org/forums/showthread.php?275873-AMD-FX-quot-Bulldozer-quot-Review-%284%29-!exclusive!-Excuse-for-1-Threaded-Perf
It’s late and I’m tired, so comparisons with i5-vs-i7 will have to wait.
per:
--- Quote from: kureshii on October 16, 2011, 03:47:26 PM ---
--- Quote from: per on October 16, 2011, 03:11:38 PM ---And in our tests intels hyperthreading actually is _more_ efficient at boosting performance than AMD:s "true core":s. That is the most suprising thing to me.
--- End quote ---
How was that test conducted? Do you have a link? I really am quite curious, because AMD’s been hyping their Bulldozer Module as a better-than-hyperthreading technology—on paper, of course. But like we’ve seen with the Pentium 4, theories don’t always work out as expected.
--- End quote ---
We simply ran our server software on the nodes, first with only one instance and then with the normal 100/node instances, and noted that the intel hardware with 2x6 cores can do 20 times more, while the 2x"8"core AMD could only do 11 times more.
Granted, not all that scientific, especially since the nodes also does a lot of other thins, network access etc, which means that the single-instance test is not really only using one cpu (it's fairly close, though).
In practice our software normally spends significant time waiting on RAM, and does perhaps 15% float and 85% integer computations when it is not waiting around.
Lupin:
--- Quote from: kureshii on October 16, 2011, 03:47:26 PM ---[edit] Some interesting results: http://www.xtremesystems.org/forums/showthread.php?275873-AMD-FX-quot-Bulldozer-quot-Review-%284%29-!exclusive!-Excuse-for-1-Threaded-Perf
--- End quote ---
Nice find! I didn't know the new bios can disable individual cores.
Navigation
[0] Message Index
[#] Next page
[*] Previous page
Go to full version