POWER7 Spec

By: Linus Torvalds (torvalds.delete@this.linux-foundation.org), February 9, 2010 7:57 am
Room: Moderated Discussions
Ian Ameline (ian.ameline@gmail.omitthis.com) on 2/9/10 wrote:
>
>It is interesting, however, that with an L1$ that is twice
>as fast, double the register file size, double the L3 cache
>size, and double the FP execution resources, that they don't
>do even better than this.

I'm actually pretty impressed that they matched Nehalem
core-for-core at the same frequency. I didn't think they
would do that - their previous cores definitely didn't.

What I'm interested in is how much of that is the 4-way
SMT. IOW, can they actually beat Nehalem in single-thread
performance too? From the numbers, I think it's going to
be close.

Looking at the single-thread numbers, not only is POWER7
a four-way SMT, but historically IBM has gotten more out of
SMT than Intel has. So the fact that they are at parity
clock-for-clock in throughput per core doesn't translate to
parity at a single thread level.

And that's where things like "what's the ALU latency?" and
"how aggressive is the OoO really?" matter. The 4-way
SMT might be filling the unused cycles that Intel is so
good at filling with Nehalem even with a single thread.

If I had to take a wild guess, I'd guess that the POWER7
4-way SMT improves throughput by no more than ~40%. Iirc,
the 2-way one on POWER6 was a roughly 30% specint_rate
improvement, but I don't think POWER7 is going to get all
that much more from its four threads (if even that). Why?
Both because of diminishing returns, and simply because
it's OoO and lower latencies, thus likely to keep its
integer resources relatively busier even with a single
thread.

Nehalem, in contrast, doesn't get as much improvement from
its 2-way SMT on something like specint_rate, because it is
already so good at filling the integer resources anyway.
And turbo (depending on the chip) will make that even less
noticeable.

So at a guess, the fact that POWER7 matches Nehalem core-
for-core in throughput means that it likely has a single-
thread IPC that is about 25% lower. But at least in the
8-core version it then more than makes up for it with a
higher frequency (compared to the rumored 2.26GHz for
Nehalem-EX)

At 3.33Ghz, the four-core Nehalem will likely beat
POWER7, though, so Intel probably retains the SpecInt
crown. Anybody want to guess at actual numbers?

(Of course, Nehalem has that Intel "libquantum" advantage.
That thing should just be removed from specint)

Linus
< Previous Post in ThreadNext Post in Thread >
TopicPosted ByDate
POWER7 SpecParadox2010/02/08 11:05 AM
  POWER7 SpecThu Nguyen2010/02/08 12:58 PM
    POWER7 SpecIan Ameline2010/02/08 09:22 PM
      POWER7 SpecThu Nguyen2010/02/08 11:54 PM
        POWER7 SpecIan Ameline2010/02/09 06:46 AM
          POWER7 SpecLinus Torvalds2010/02/09 07:57 AM
            industry-standard single-threaded performance benchmarks absentMichael S2010/02/09 08:09 AM
              industry-standard single-threaded performance benchmarks absentParadox2010/02/09 08:33 AM
                industry-standard single-threaded performance benchmarks absentMichael S2010/02/09 09:30 AM
                  industry-standard single-threaded performance benchmarks absentParadox2010/02/09 10:52 AM
                    industry-standard single-threaded performance benchmarks absentJesper Frimann2010/02/09 11:33 AM
              industry-standard single-threaded performance benchmarks absentLinus Torvalds2010/02/09 08:48 AM
                industry-standard single-threaded performance benchmarks absentMichael S2010/02/09 09:26 AM
                  industry-standard single-threaded performance benchmarks absentLinus Torvalds2010/02/09 09:58 AM
                    industry-standard single-threaded performance benchmarks absentJesper Frimann2010/02/09 12:17 PM
                      industry-standard single-threaded performance benchmarks absentLinus Torvalds2010/02/09 12:54 PM
                        industry-standard single-threaded performance benchmarks absentJesper Frimann2010/02/09 02:10 PM
                    industry-standard single-threaded performance benchmarks absentParadox2010/02/09 01:22 PM
                industry-standard single-threaded performance benchmarks absentanon2010/02/09 10:21 AM
            POWER7 SpecJesper Frimann2010/02/09 12:30 PM
            POWER7 SpecDoug Siebert2010/02/09 05:38 PM
              POWER7 SpecLinus Torvalds2010/02/09 07:28 PM
                POWER7 SpecLinus Torvalds2010/02/09 07:28 PM
                  POWER7 SpecIan Ameline2010/02/09 08:02 PM
                    POWER7 SpecDoug Siebert2010/02/09 10:18 PM
                POWER7 Specsomeone2010/02/09 08:20 PM
                  POWER7 SpecLinus Torvalds2010/02/10 09:17 AM
                    POWER7 SpecLinus Torvalds2010/02/10 09:46 AM
                      POWER7 SpecEmil2010/02/10 11:06 AM
                    POWER7 SpecIan Ameline2010/02/10 10:13 AM
                    POWER7 Specsomeone2010/02/10 11:01 AM
                      POWER7 SpecLinus Torvalds2010/02/10 11:32 AM
                        POWER7 SpecMichael S2010/02/10 12:30 PM
                          POWER7 SpecLinus Torvalds2010/02/10 01:25 PM
                            POWER7 Specmpx2010/02/10 02:58 PM
                        POWER7 Specnemlis2010/02/11 12:24 AM
                          POWER7 Specnone2010/02/11 12:52 AM
                            POWER7 Specnemlis2010/02/11 01:52 PM
          POWER7 Specmpx2010/02/09 08:18 AM
          POWER7 SpecMichael S2010/02/09 09:08 AM
        POWER7 SpecJesper Frimann2010/02/16 02:29 AM
          POWER7 SpecJamie Lucier2010/02/16 06:35 AM
            POWER7 Specanon2010/02/16 07:16 AM
              POWER7 SpecJamie Lucier2010/02/16 10:46 AM
                POWER7 Specanon2010/02/16 03:42 PM
                  POWER7 SpecJamie Lucier2010/02/18 06:07 AM
              POWER7 SpecThu Nguyen2010/02/16 01:23 PM
            POWER7 SpecParadox2010/02/16 09:57 AM
              POWER7 SpecJamie Lucier2010/02/16 11:22 AM
                POWER7 SpecParadox2010/02/16 02:00 PM
            POWER7 SpecJesper Frimann2010/02/16 02:14 PM
              POWER7 SpecJamie Lucier2010/02/17 10:41 AM
  POWER7 SpecJesper Frimann2010/02/09 04:43 AM
Reply to this Topic
Name:
Email:
Topic:
Body: No Text
How do you spell tangerine? 🍊