Article: Tukwila Update
By: Default (structuredmayhem.delete@this.hotmail.com), February 6, 2009 9:57 am
Room: Moderated Discussions
Since your post was long, I'll try to keep my response simple.
128 cores with small caches one a chip is a pipe-dream. Yes, the logical core of the Itanium is small, but it's not that small.
Take for example, Madison, the first "real" Itanium worth talking about, because it had a respectable amount of cache to keep the processor fed. At 374mm^2 on a 0.13 micron process, it's a big one, and at best, you'll fit two of these boys on a 90nm process (Montecito), four on a 65nm process (Tukwila), and eight on a 45nm process, and sixteen on a 32nm process.
Even if you nix some more cache, the best you can get is around 16/32 cores using your cutting-edge 45nm/32nm processes (this is ignoring the overhead of cache coherency support, which gets more complex as you add more cores). In addition, cutting the cache has a LARGE detrimental effect on performance with EPIC, because the processor takes every branch until the test is evaluated (see my post above). You're just not going to get 128 cores on a chip with EPIC.
As for getting rid of vectors...vectorization is a necessity today! Vectorization cuts-down on the instruction decode overhead of having individual instructions, placing the burden on software. Sure, it sucks for many tasks, but the reality is that the costs of adding it are less than the costs of adding another ALU/FPU and another dispatch pipe to keep it fed.
And I wouldn't decrease the number of instructions in a bundle - the number of instructions within a bundle defines the upper-bound of IPC for Itanium. Changing this would also mean you'd need a new set of optimizing compilers.
128 cores with small caches one a chip is a pipe-dream. Yes, the logical core of the Itanium is small, but it's not that small.
Take for example, Madison, the first "real" Itanium worth talking about, because it had a respectable amount of cache to keep the processor fed. At 374mm^2 on a 0.13 micron process, it's a big one, and at best, you'll fit two of these boys on a 90nm process (Montecito), four on a 65nm process (Tukwila), and eight on a 45nm process, and sixteen on a 32nm process.
Even if you nix some more cache, the best you can get is around 16/32 cores using your cutting-edge 45nm/32nm processes (this is ignoring the overhead of cache coherency support, which gets more complex as you add more cores). In addition, cutting the cache has a LARGE detrimental effect on performance with EPIC, because the processor takes every branch until the test is evaluated (see my post above). You're just not going to get 128 cores on a chip with EPIC.
As for getting rid of vectors...vectorization is a necessity today! Vectorization cuts-down on the instruction decode overhead of having individual instructions, placing the burden on software. Sure, it sucks for many tasks, but the reality is that the costs of adding it are less than the costs of adding another ALU/FPU and another dispatch pipe to keep it fed.
And I wouldn't decrease the number of instructions in a bundle - the number of instructions within a bundle defines the upper-bound of IPC for Itanium. Changing this would also mean you'd need a new set of optimizing compilers.
Topic | Posted By | Date |
---|---|---|
Tukwila Update - article online | David Kanter | 2009/02/05 12:03 AM |
Tukwila Update - article online | Dan | 2009/02/05 03:17 AM |
Tukwila Update - article online | Joe Chang | 2009/02/05 09:16 AM |
Tukwila Update - article online | Temp | 2009/02/05 09:25 AM |
Tukwila Update - article online | Paul | 2009/02/05 12:29 PM |
Tukwila Update - article online | David Kanter | 2009/02/05 06:32 PM |
Tukwila Update - article online | Phil | 2009/02/06 01:24 AM |
Great. Finally hard numbers | Michael S | 2009/02/06 04:46 AM |
Tukwila Update - article online | lubemark | 2009/02/06 05:54 AM |
Tukwila Update - article online | Phil | 2009/02/06 07:29 AM |
Tukwila Update - article online | RagingDragon | 2009/02/07 03:39 PM |
Tukwila Update - article online | Michael S | 2009/02/07 04:09 PM |
Tukwila Update - article online | savantu | 2009/02/06 06:23 AM |
Tukwila Update - article online | Michael S | 2009/02/06 07:13 AM |
Tukwila Update - article online | someone | 2009/02/06 07:18 AM |
Tukwila Update - article online | Phil | 2009/02/06 07:47 AM |
Tukwila Update - article online | someone | 2009/02/06 08:17 AM |
Tukwila Update - article online | RagingDragon | 2009/02/07 03:51 PM |
Tukwila Update - article online | Linus Torvalds | 2009/02/06 08:37 AM |
Tukwila Update - article online | someone | 2009/02/06 09:19 AM |
Tukwila Update - article online | savantu | 2009/02/06 10:19 AM |
Tukwila Update - article online | Linus Torvalds | 2009/02/06 10:40 AM |
Tukwila Update - article online | savantu | 2009/02/06 11:00 AM |
Tukwila Update - article online | Phil | 2009/02/09 04:54 AM |
Tukwila Update - article online | Doug Siebert | 2009/02/09 10:40 PM |
Tukwila Update - article online | Jouni Osmala | 2009/02/10 01:03 AM |
Tukwila Update - article online | someone | 2009/02/10 06:15 AM |
Tukwila Update - article online | slacker | 2009/02/10 06:22 PM |
Tukwila Update - article online | Michael S | 2009/02/05 03:56 PM |
Tukwila Update - article online | David Kanter | 2009/02/05 04:55 PM |
Tukwila Update - article online | someone | 2009/02/05 05:47 PM |
Tukwila Update - article online | anon | 2009/02/05 10:16 PM |
Tukwila Update - article online | RagingDragon | 2009/02/05 10:27 PM |
Tukwila Update - article online | someone | 2009/02/06 07:32 AM |
Tukwila Update - article online | anon | 2009/02/06 09:25 AM |
Tukwila Update - article online | someone | 2009/02/06 09:40 AM |
POWER6 memory bandwidth | Michael S | 2009/02/05 03:30 AM |
POWER6 memory bandwidth | someone | 2009/02/05 07:00 AM |
POWER6 memory bandwidth | Michael S | 2009/02/05 07:36 AM |
POWER6 interconnect | confused | 2009/02/05 10:50 AM |
POWER6 interconnect | foobar | 2009/02/05 02:12 PM |
POWER6 memory bandwidth | Wes Felter | 2009/02/05 12:57 PM |
POWER6 memory bandwidth | Jesper Frimann | 2009/02/09 11:54 PM |
POWER6 memory bandwidth | Michael S | 2009/02/10 07:21 AM |
Why the platform focus? | Linus Torvalds | 2009/02/05 08:40 AM |
Why the platform focus? | savantu | 2009/02/05 08:50 AM |
Why the platform focus? | Vincent Diepeveen | 2009/02/05 09:29 AM |
Why the platform focus? | savantu | 2009/02/05 10:34 PM |
Why the platform focus? | Vincent Diepeveen | 2009/02/05 11:09 PM |
Why the platform focus? | Phil | 2009/02/06 01:10 AM |
Why the platform focus? | savantu | 2009/02/06 01:50 AM |
Why the platform focus? | Phil | 2009/02/06 07:09 AM |
Why the platform focus? | savantu | 2009/02/06 10:08 AM |
Why the platform focus? | someone | 2009/02/06 10:21 AM |
Why the platform focus? | mpx | 2009/02/06 02:04 PM |
Why the platform focus? | someone | 2009/02/06 02:16 PM |
Why the platform focus? | RagingDragon | 2009/02/07 04:16 PM |
Why the platform focus? | mas | 2009/02/25 08:28 AM |
itanium bigger than entire car industry | Vincent Diepeveen | 2009/02/06 07:12 AM |
itanium bigger than entire car industry | Devon Welles | 2009/02/06 07:51 AM |
itanium bigger than entire car industry | Vincent Diepeveen | 2009/02/06 10:41 AM |
itanium bigger than entire car industry | Dean Kent | 2009/02/06 07:56 PM |
Unit sales is meaningless when ASP grows faster | someone | 2009/02/07 09:38 AM |
Unit sales is meaningless when ASP grows faster | Dean Kent | 2009/02/07 03:10 PM |
Unit sales is meaningless when ASP grows faster | RagingDragon | 2009/02/07 04:34 PM |
itanium bigger than entire car industry | Vincent Diepeveen | 2009/02/08 05:35 AM |
itanium bigger than entire car industry | RagingDragon | 2009/02/07 04:40 PM |
Why the platform focus? | Vincent Diepeveen | 2009/02/06 07:47 AM |
Yes it doesm performance matters | bob | 2009/02/05 10:51 AM |
Yes it doesm performance matters | Venki | 2009/02/05 11:06 AM |
Why the platform focus? | Doug Siebert | 2009/02/06 01:07 AM |
Why the platform focus? | savantu | 2009/02/06 02:00 AM |
Why the platform focus? | someone | 2009/02/05 10:49 AM |
Why the platform focus? | Linus Torvalds | 2009/02/05 12:03 PM |
Why the platform focus? | Default | 2009/02/05 01:29 PM |
Why the platform focus? | anon | 2009/02/05 02:08 PM |
Why the platform focus? | someone | 2009/02/05 02:24 PM |
Why the platform focus? | Linus Torvalds | 2009/02/05 03:30 PM |
Why the platform focus? | Paradox | 2009/02/05 11:22 AM |
Why the platform focus? | slacker | 2009/02/05 01:41 PM |
Why the platform focus? | RagingDragon | 2009/02/05 10:57 PM |
Why the platform focus? | Michael S | 2009/02/06 06:11 AM |
Why the platform focus? | slacker | 2009/02/06 01:58 PM |
Why the platform focus? | Michael S | 2009/02/08 02:24 AM |
Why the platform focus? | someone | 2009/02/08 09:38 AM |
Why the platform focus? | David Kanter | 2009/02/08 04:27 PM |
Why the platform focus? | someone | 2009/02/08 07:26 PM |
Why the platform focus? | savantu | 2009/02/09 12:35 AM |
Why the platform focus? | someone | 2009/02/08 09:53 AM |
All x86 SpecInt scores are useless due to autopar (NT) | Michael S | 2009/02/05 03:15 PM |
Auto parallelization | David Kanter | 2009/02/05 06:17 PM |
All x86 SpecInt scores are useless due to autopar (NT) | Paradox | 2009/02/06 08:47 AM |
Why the platform focus? | David Kanter | 2009/02/05 04:49 PM |
Why the platform focus? | David Kanter | 2009/02/06 01:09 AM |
Why the platform focus? | Linus Torvalds | 2009/02/06 08:14 AM |
Why the platform focus? | savantu | 2009/02/06 10:37 AM |
Why the platform focus? | Linus Torvalds | 2009/02/06 12:49 PM |
Why the platform focus? | Linus Torvalds | 2009/02/06 01:09 PM |
Intel puts its money where its mouth is | someone | 2009/02/06 02:08 PM |
Intel puts its money where its mouth is | RagingDragon | 2009/02/07 06:01 PM |
Intel puts its money where its mouth is | someone | 2009/02/08 02:24 PM |
mission-critical | Michael S | 2009/02/08 05:06 PM |
mission-critical | mpx | 2009/02/09 02:30 AM |
mission-critical | rwessel | 2009/02/09 03:23 PM |
mission-critical | anon | 2009/02/09 03:55 AM |
mission-critical | EduardoS | 2009/02/09 05:17 PM |
mission-critical | Dean Kent | 2009/02/09 08:11 PM |
mission-critical | Michael S | 2009/02/10 05:20 AM |
mission-critical | Dean Kent | 2009/02/10 07:26 AM |
mission-critical | Michael S | 2009/02/10 08:01 AM |
mission-critical | Dean Kent | 2009/02/10 01:36 PM |
mission-critical | someone | 2009/02/10 09:05 AM |
mission-critical | Dean Kent | 2009/02/10 01:22 PM |
mission-critical | Zt | 2009/02/22 04:54 PM |
mission-critical | anon | 2009/02/10 10:41 PM |
mission-critical | EduardoS | 2009/02/10 01:46 PM |
mission-critical | Dean Kent | 2009/02/10 02:31 PM |
mission-critical | slacker | 2009/02/10 07:30 PM |
mission-critical | Vincent Diepeveen | 2009/02/18 07:20 AM |
Mission critical | mpx | 2009/02/09 01:00 AM |
Why the platform focus? | savantu | 2009/02/07 01:15 AM |
Sun and x86 server differentiation | David Kanter | 2009/02/07 01:34 AM |
Sun and x86 server differentiation | max | 2009/02/07 03:30 AM |
Sun and x86 server differentiation | someone | 2009/02/07 10:19 AM |
Sun and x86 server differentiation | Linus Torvalds | 2009/02/07 10:44 AM |
Sun and x86 server differentiation | RagingDragon | 2009/02/07 06:09 PM |
Sun and x86 server differentiation | Michael S | 2009/02/08 05:05 AM |
Sun and x86 server differentiation | RagingDragon | 2009/02/10 12:03 AM |
Sun and x86 server differentiation | Jesper Frimann | 2009/02/10 12:51 AM |
Sun and x86 server differentiation | Alex Jones | 2009/02/10 01:43 PM |
Why the platform focus? | bob | 2009/02/08 04:51 AM |
Why the platform focus? | someone | 2009/02/08 09:23 AM |
missing the big picture | AM | 2009/02/18 06:43 AM |
missing the big picture | Michael S | 2009/02/18 08:42 AM |
missing the big picture | AM | 2009/02/18 09:03 AM |
Why the platform focus? | mpx | 2009/02/06 12:47 PM |
Itanium - slowest and most obsolete server CPU family in the world, NOW. | mpx | 2009/02/06 04:48 PM |
Itanium - slowest and most obsolete server CPU family in the world, NOW. | Paul | 2009/02/07 02:56 PM |
z series? | Michael S | 2009/02/07 03:12 PM |
Itanium - slowest and most obsolete server CPU family in the world, NOW. | someone else | 2009/02/24 04:37 AM |
Itanium - slowest and most obsolete server CPU family in the world, NOW. | EduardoS | 2009/02/24 06:55 AM |
Itanium - slowest and most obsolete server CPU family in the world, NOW. | someone else | 2009/02/25 01:55 AM |
Itanium - slowest and most obsolete server CPU family in the world, NOW. | Michael S | 2009/02/25 02:27 AM |
Why the platform focus? | RagingDragon | 2009/02/07 06:18 PM |
Why the platform focus? | Paul | 2009/02/08 01:10 PM |
Why the platform focus? | Jukka Larja | 2009/02/08 11:04 PM |
Why the platform focus? | slacker | 2009/02/06 02:10 PM |
Why the platform focus? | Linus Torvalds | 2009/02/06 02:40 PM |
Why the platform focus? | savantu | 2009/02/06 02:51 PM |
Why the platform focus? | someone | 2009/02/06 02:58 PM |
Why the platform focus? | Linus Torvalds | 2009/02/07 09:26 AM |
Why the platform focus? | someone | 2009/02/07 10:10 AM |
Why the platform focus? | Linus Torvalds | 2009/02/07 10:40 AM |
Why the platform focus? | someone | 2009/02/07 12:24 PM |
Why the platform focus? | Doug Siebert | 2009/02/08 12:32 AM |
Why the platform focus? | max | 2009/02/08 04:57 AM |
Why the platform focus? | Michael S | 2009/02/08 05:20 AM |
Why the platform focus? | someone | 2009/02/08 09:15 AM |
Why the platform focus? | Doug Siebert | 2009/02/08 11:36 PM |
Why the platform focus? | hobold | 2009/02/09 05:49 AM |
Why the platform focus? | anon | 2009/02/24 01:57 AM |
Why the platform focus? | Linus Torvalds | 2009/02/24 09:45 AM |
Why the platform focus? | savantu | 2009/02/24 12:30 PM |
Why the platform focus? | slacker | 2009/02/24 01:51 PM |
Why the platform focus? | savantu | 2009/02/25 12:04 AM |
Why the platform focus? | Michael S | 2009/02/25 02:34 AM |
Why the platform focus? | anon | 2009/02/25 10:17 AM |
Why the platform focus? | max | 2009/02/25 11:15 AM |
Why the platform focus? | someone | 2009/02/24 05:43 PM |
Why the platform focus? | Doug Siebert | 2009/02/24 08:26 PM |
Why the platform focus? | Howard Chu | 2009/02/25 03:07 AM |
Why the platform focus? | someone | 2009/02/25 06:48 AM |
Why the platform focus? | someone | 2009/02/25 06:41 AM |
Why the platform focus? | Linus Torvalds | 2009/02/25 09:17 AM |
Why the platform focus? | someone | 2009/02/25 09:55 AM |
has anyone seen Tukwila silicon? | anon | 2009/02/25 10:38 AM |
Why the platform focus? | Linus Torvalds | 2009/02/25 11:05 AM |
Why the platform focus? | slacker | 2009/02/25 01:11 PM |
Why the platform focus? | a reader | 2009/02/26 09:11 PM |
Why the platform focus? | rcf | 2009/02/27 01:32 PM |
Why the platform focus? | max | 2009/02/27 02:11 PM |
Why the platform focus? | rcf | 2009/02/27 03:50 PM |
Why the platform focus? | Vincent Diepeveen | 2009/02/25 04:30 PM |
$40M sale to $16M company | bob | 2009/02/25 08:25 PM |
$40M sale to $16M company | Richard Cownie | 2009/02/26 12:21 PM |
Why the platform focus? | anonymous | 2009/02/24 11:52 AM |
Why the platform focus? | savantu | 2009/02/24 12:20 PM |
Why the platform focus? | anonymous | 2009/02/24 03:31 PM |
Why the platform focus? | savantu | 2009/02/25 12:05 AM |
Why the platform focus? | someone else | 2009/02/25 01:04 AM |
Why the platform focus? | Michael S | 2009/02/25 01:42 AM |
Put me down for $500 that Poulson doesn't arrive earlier than Q4/2011 (NT) | slacker | 2009/02/25 12:39 PM |
Why the platform focus? | someone | 2009/02/25 06:54 AM |
Why the platform focus? | anonymous | 2009/02/25 09:46 AM |
Why the platform focus? | someone | 2009/02/25 10:22 AM |
Why the platform focus? | anon | 2009/02/25 11:01 AM |
Why the platform focus? | anonymous | 2009/02/25 11:54 AM |
Why the platform focus? | mpx | 2009/02/24 02:11 PM |
Why the platform focus? | anon | 2009/02/24 08:57 PM |
Why the platform focus? | Doug Siebert | 2009/02/24 10:04 PM |
Why the platform focus? | anon | 2009/02/24 10:46 PM |
Why the platform focus? | Doug Siebert | 2009/02/25 05:13 PM |
Why the platform focus? | anon | 2009/02/25 08:53 PM |
Why the platform focus? | bob | 2009/02/25 09:00 PM |
Please try to keep up (NT) | anon | 2009/02/25 09:49 PM |
Why the platform focus? | Doug Siebert | 2009/02/26 12:09 AM |
Why the platform focus? | anon | 2009/02/26 01:12 AM |
Why the platform focus? | Michael S | 2009/02/26 02:16 AM |
Why the platform focus? | James | 2009/02/26 06:09 AM |
sufficiently intimate with the OS | Michael S | 2009/02/26 06:29 AM |
sufficiently intimate with the OS | anon | 2009/02/27 01:01 AM |
sufficiently intimate with the OS | Howard Chu | 2009/02/27 01:37 AM |
Why the platform focus? | Michael S | 2009/02/25 02:02 AM |
Why the platform focus? | anon | 2009/02/25 03:07 AM |
Why the platform focus? | anon | 2009/02/07 01:18 PM |
Why the platform focus? | Vincent Diepeveen | 2009/02/08 10:16 AM |
Why the platform focus? | anon | 2009/02/25 07:40 AM |
Intels financial status | Vincent Diepeveen | 2009/02/25 12:02 PM |
Why the platform focus? | someone | 2009/02/25 07:54 AM |
Why the platform focus? | Vincent Diepeveen | 2009/02/06 08:20 AM |
Why the platform focus? | Default | 2009/02/06 09:57 AM |
Why the platform focus? | Vincent Diepeveen | 2009/02/06 10:59 AM |
Why the platform focus? | RagingDragon | 2009/02/07 06:43 PM |
Tukwila Update - article online | Vincent Diepeveen | 2009/02/05 09:11 AM |