Barcelona Architecture: AMD on the Counterattack
by Anand Lal Shimpi on March 1, 2007 12:05 AM EST- Posted in
- CPUs
SSE128
Many of the "major" changes to Barcelona were driven by one significant change: what AMD is calling SSE128. In the K8 architecture AMD can execute two SSE operations in parallel; however the SSE execution units are only 64-bits wide. For 128-bit SSE operations, the K8 had to handle them as two 64-bit operations. This also means that when a 128-bit SSE instruction is fetched, it is first decoded into two micro-ops (one for each 64-bit half of the instruction), thus taking up an extra decode port for a single instruction.
Barcelona widens the execution units that handle SSE operations from 64-bits to 128-bits, so now 128-bit SSE operations don't have to be broken up into two 64-bit operations. This also means that you get more usable decode bandwidth since 128-bit SSE instructions now map to a single micro-op instead of two. The FP scheduler can now handle these 128-bit SSE operations as well.
It's the increase to SSE execution width that drove a number of other changes within the core. Since you effectively have more decode bandwidth when executing 128-bit SSE instructions AMD discovered a new bottleneck: instruction fetch bandwidth. These 128-bit SSE instructions tend to be quite large, and in order to maximize the number decoded in parallel the Barcelona core can now fetch 32-bytes per cycle, up from 16-bytes in K8. The 32B instruction fetch not only benefits SSE code but also seems to benefit integer code as well. Bigger instructions in general will see a performance boost here.
Now that you can fetch and decode more instructions, you need to be able to get more data to the execution core and thus AMD widened the interface between the L1 data cache and Barcelona's SSE registers. Barcelona can now perform two 128-bit SSE loads per cycle from the L1-D cache compared to two 64-bit loads per cycle in K8. AMD then widened the interface between the L2 cache and the memory controller so that now 128-bits can be transferred per cycle, once again to balance out all of the aforementioned changes.
The culmination of the SSE128 improvements is very similar to some of the changes made in the Yonah to Merom transition. Prior to Conroe/Merom, Yonah could not keep up with AMD's K8 when it came to FP/SSE performance. Almost a year and a half ago we did an article where we compared AMD's K8 to Intel's Yonah running at the same clock speed. While Yonah was able to equal the K8's performance in general applications, professional 3D rendering and games, it could not compete when it came to video encoding.
There were a number of SSE performance improvements made to Yonah but it wasn't until Intel's Core 2 processors that Intel was really able to outperform AMD in our video encoding tests. Whether the improvements were due to the single cycle SSE throughput introduced in Core 2 or the wider front end or a combination of both remains to be seen. Although it's difficult to compare specs between two very different architectures, encoding performance is a sore spot for AMD today, and it's something that the SSE128 changes can only help.
| AMD Architecture Comparison | ||
| K8 | Barcelona | |
| SSE Execution Width | 64-bit | 128-bit |
| Instruction Fetch Bandwidth | 16 bytes/cycle | 32 bytes/cycle |
| Data Cache Bandwidth | 2 x 64-bit loads/cycle | 2 x 128-bit loads/cycle |
| L2/Northbridge Bandwidth | 64 bits/cycle | 128 bits/cycle |
| FP Scheduler Depth | 36 Dedicated x 64-bit ops | 36 Dedicated x 128-bit ops |
Many of the "major" changes to Barcelona were driven by one significant change: what AMD is calling SSE128. In the K8 architecture AMD can execute two SSE operations in parallel; however the SSE execution units are only 64-bits wide. For 128-bit SSE operations, the K8 had to handle them as two 64-bit operations. This also means that when a 128-bit SSE instruction is fetched, it is first decoded into two micro-ops (one for each 64-bit half of the instruction), thus taking up an extra decode port for a single instruction.
Barcelona widens the execution units that handle SSE operations from 64-bits to 128-bits, so now 128-bit SSE operations don't have to be broken up into two 64-bit operations. This also means that you get more usable decode bandwidth since 128-bit SSE instructions now map to a single micro-op instead of two. The FP scheduler can now handle these 128-bit SSE operations as well.
It's the increase to SSE execution width that drove a number of other changes within the core. Since you effectively have more decode bandwidth when executing 128-bit SSE instructions AMD discovered a new bottleneck: instruction fetch bandwidth. These 128-bit SSE instructions tend to be quite large, and in order to maximize the number decoded in parallel the Barcelona core can now fetch 32-bytes per cycle, up from 16-bytes in K8. The 32B instruction fetch not only benefits SSE code but also seems to benefit integer code as well. Bigger instructions in general will see a performance boost here.
Now that you can fetch and decode more instructions, you need to be able to get more data to the execution core and thus AMD widened the interface between the L1 data cache and Barcelona's SSE registers. Barcelona can now perform two 128-bit SSE loads per cycle from the L1-D cache compared to two 64-bit loads per cycle in K8. AMD then widened the interface between the L2 cache and the memory controller so that now 128-bits can be transferred per cycle, once again to balance out all of the aforementioned changes.
The culmination of the SSE128 improvements is very similar to some of the changes made in the Yonah to Merom transition. Prior to Conroe/Merom, Yonah could not keep up with AMD's K8 when it came to FP/SSE performance. Almost a year and a half ago we did an article where we compared AMD's K8 to Intel's Yonah running at the same clock speed. While Yonah was able to equal the K8's performance in general applications, professional 3D rendering and games, it could not compete when it came to video encoding.
There were a number of SSE performance improvements made to Yonah but it wasn't until Intel's Core 2 processors that Intel was really able to outperform AMD in our video encoding tests. Whether the improvements were due to the single cycle SSE throughput introduced in Core 2 or the wider front end or a combination of both remains to be seen. Although it's difficult to compare specs between two very different architectures, encoding performance is a sore spot for AMD today, and it's something that the SSE128 changes can only help.
Facebook
Twitter
RSS
83 Comments
View All Comments
chucky2 - Friday, March 2, 2007 - link
Can you post the link that originates at AMD's own website then that says specifically that AM2+ CPU's are guaranteed to work - understandably maybe not supporting every new feature - in current AM2 boards?Not a news post from DailyTech, The Inquirer, Toms, whatever...one that's on AMD's site itself.
And No, AMD could make AM2+ completely incompatible with current AM2 boards and they probably wouldn't see much drop if at all from the large OEM's. The large OEM's would just ensure that when the AM2+ CPU's came in, AM2+ motherboards would likewise come in.
Believe me, I want to see the link...because I'm desperately awaiting 690G or MCP68, whichever comes first (which is probably MCP68 at the pace AMD is moving on 690G).
Chuck
yacoub - Thursday, March 1, 2007 - link
You say 128kb L1 per core but the diagram image just beneath that text shows a 64bit L1 cache. Please confirm which it is.
Thanks.
Awesome article, btw. Seems like quite a significant group of changes to the CPU. Looking forward to seeing how it stacks up against the best Quad Core2 Intel can offer. =)
yacoub - Thursday, March 1, 2007 - link
also, please forgive my hasty typing - I wrote "128kb" and "64bit" - I meant "128KB" and "64KB"JarredWalton - Thursday, March 1, 2007 - link
L1 is 128K total - 64K data and 64K instruction.Beenthere - Thursday, March 1, 2007 - link
AMD doesn't do knee-jerk reactions like Intel because AMD has superior products. AMD continues to take market share from Intel in every segment and Barcelona will continue that trend. Barcelona looks to be every bit as superior to Intel's hacked/patched/glued together chips as Opteron was when introduced. Intel's chips depend on huge cache size for their performance and that crutch won't work after the intro of Barcelona.For those without a clue, AMD didn't start design of Barcelona last week or last year. It's been in the development pipeline for many years and thr performance will demonstrate exactly why AMD's long term platform stability is the right choice for most enterprise buyers. Intel is gonna feel the pain again.
Roy2001 - Thursday, March 1, 2007 - link
Facts please, no BS.zsdersw - Thursday, March 1, 2007 - link
Idiocy incarnate.Regs - Thursday, March 1, 2007 - link
AMD, like Intel, start numerious projects. Just not all of them get to this finish line. Actually a lot of them don't even reach the end of the planning phase before being scratched.As for Intel and their large caches...well I'd say it's amazing how half their die (if not more) is used for cache and still had enough space for all the core logic that's kicking the crap out of the K8 now.
Common sense!
erwos - Thursday, March 1, 2007 - link
Looks like some good improvements coming down the pipe. The cache size issue makes me nervous, though - 512kb per core is starting to look a little antiquated, and there's no information about the bandwidth to the L3 cache (which, presumably, is slower than L2).SmokeRngs - Thursday, March 1, 2007 - link
In the past, AMD did not need the large cache sizes that Intel did for their processors. This was very obvious in regards to the Netburst architecture. However, while Core2 is much better than Netburst there are still disadvantages for Intel.I'll explain a little background as far as I understand it. In the K7 and Netburst days, Intel had to have the cache to make up for their long pipeline. Branch mispredictions are going to happen and the penalty on the long pipeline of the Netburst processors hurt their IPC badly. The shorter pipeline on the K7 did not have the same performance penalty due to the shorter pipeline. With K8, the on die memory controller also negated the need for large L2 caches due to the reduced latency when accessing main memory. This has been one of the major performance aspects for the K8 architecture.
The Core2 architecture obviously does not have the on die memory controller so the need for larger caches is still present and Intel sees improvement due to the larger caches. Barcelona still has the on die memory controller and the previous efficiency is still there and still negates the need for large caches. This is just the difference between architectures. While having a larger cache on the K8 did improve performance some in some usage scenarios, it wasn't on the same scale as the improvements Intel received with a larger cache.
AMD can't compete with Intel in regards to cache size. However, other architecture differences make up for the lack of large amounts of cache. Barcelona having a smaller cache does not seem to be a big problem. If it was a big problem, AMD probably would have gone with a larger cache to get the extra performance. Bigger does not always mean better or at least enough better to warrant the extra.
Smaller cache will mean fewer transistors which should mean better yields, lower power consumption and cheaper to produce.