When I started playing with computers (1967) mainframes were the thing (we called them computers). The University of Waterloo got an IBM SYSTEM/360-75, the biggest computer in Canada, in 1967 or 1968. It filled a lot of the "Big Red Room", a jewel case the size of gymnasium. Eventually, it had 1MiB of core memory and 2MiB of LCS (slow core memory). It ran most of the computing work for the whole campus, including student programs. The successor to this hardware is IBM's System Z. The latest announced processor chip has 512MiB of cache per socket. I have no idea how many sockets a large system would have. The /360 architecture only allowed for 24 bits of address -- 16M. The /370 eventually allowed for 31 bits of address. I don't know what Z's limits are. Anyway, the whole contents of the 360-75 core would occupy 1/512 of the cache of one socket of the Z. The -75 didn't even have a cache. The -85, introduced a couple of years later, did. When an Ottawa company (SDL?) got one, that became the largest computer in Canada. The fastest instructions on the 360-75 (for example, adding two registers) took .39 microseconds. On my main computer in those days, the IBM 1710 Model 2, the fastest instruction took 70ns, plus 10ns for each pair of digits processed. The -75 was a LOT faster. Now ordinary PCs are 2GHz or more and execute better than 1 instruction per cycle. Perhaps 1000 times faster than the 360-75. Bulk RAM is perhaps only 10 times faster. But cache has a large effect. Disk latencies might not be a lot better: 2,400 RPM vs 10,000 RPM but seek times, capacity, and bandwidth are. Not to mention SSDs. I cannot even estimate power requirements. And you have to add power for cooling. The price is perhaps 10,000 times lower. You can see why I'm not really impressed with x86 improvements since Haswell.
On Thu, Sep 02, 2021 at 12:04:51PM -0400, D. Hugh Redelmeier via talk wrote:
When I started playing with computers (1967) mainframes were the thing (we called them computers).
The University of Waterloo got an IBM SYSTEM/360-75, the biggest computer in Canada, in 1967 or 1968. It filled a lot of the "Big Red Room", a jewel case the size of gymnasium.
Eventually, it had 1MiB of core memory and 2MiB of LCS (slow core memory). It ran most of the computing work for the whole campus, including student programs.
The successor to this hardware is IBM's System Z. The latest announced processor chip has 512MiB of cache per socket. I have no idea how many sockets a large system would have.
Well they claim up to 190 user configurable cores (no idea how many total cores that means given some are reserved for the system I guess). At 5.2GHz. They don't make it easy to understand. Looks like up to 5 drawers, with up 4 sockets each, and 12 cores with 2 threads, so 48 cores per drawer but only 41 or 43 active in a drawer, and clearly not all available for the customer to use. And then 8TB effective ram per drawer (using 10TB actual ram in some kind of "RAID" they call "RAIM"). All fit in 4 racks, which is rather tiny compared to the old machines.
The /360 architecture only allowed for 24 bits of address -- 16M. The /370 eventually allowed for 31 bits of address. I don't know what Z's limits are. Anyway, the whole contents of the 360-75 core would occupy 1/512 of the cache of one socket of the Z.
I believe all z series have been 64 bit address space.
The -75 didn't even have a cache. The -85, introduced a couple of years later, did. When an Ottawa company (SDL?) got one, that became the largest computer in Canada.
The fastest instructions on the 360-75 (for example, adding two registers) took .39 microseconds.
On my main computer in those days, the IBM 1710 Model 2, the fastest instruction took 70ns, plus 10ns for each pair of digits processed. The -75 was a LOT faster.
Now ordinary PCs are 2GHz or more and execute better than 1 instruction per cycle. Perhaps 1000 times faster than the 360-75.
Bulk RAM is perhaps only 10 times faster. But cache has a large effect.
Disk latencies might not be a lot better: 2,400 RPM vs 10,000 RPM but seek times, capacity, and bandwidth are. Not to mention SSDs.
I cannot even estimate power requirements. And you have to add power for cooling.
The price is perhaps 10,000 times lower.
You can see why I'm not really impressed with x86 improvements since Haswell.
Oh intel has been slacking off lately. -- Len Sorensen
On 9/2/21 4:18 PM, Lennart Sorensen via talk wrote:
On Thu, Sep 02, 2021 at 12:04:51PM -0400, D. Hugh Redelmeier via talk wrote:
When I started playing with computers (1967) mainframes were the thing (we called them computers).
The University of Waterloo got an IBM SYSTEM/360-75, the biggest computer in Canada, in 1967 or 1968. It filled a lot of the "Big Red Room", a jewel case the size of gymnasium.
Eventually, it had 1MiB of core memory and 2MiB of LCS (slow core memory). It ran most of the computing work for the whole campus, including student programs.
The successor to this hardware is IBM's System Z. The latest announced processor chip has 512MiB of cache per socket. I have no idea how many sockets a large system would have.
Well they claim up to 190 user configurable cores (no idea how many total cores that means given some are reserved for the system I guess). At 5.2GHz. They don't make it easy to understand.
Looks like up to 5 drawers, with up 4 sockets each, and 12 cores with 2 threads, so 48 cores per drawer but only 41 or 43 active in a drawer, and clearly not all available for the customer to use. And then 8TB effective ram per drawer (using 10TB actual ram in some kind of "RAID" they call "RAIM"). All fit in 4 racks, which is rather tiny compared to the old machines.
The /360 architecture only allowed for 24 bits of address -- 16M. The /370 eventually allowed for 31 bits of address. I don't know what Z's limits are. Anyway, the whole contents of the 360-75 core would occupy 1/512 of the cache of one socket of the Z.
I believe all z series have been 64 bit address space.
The -75 didn't even have a cache. The -85, introduced a couple of years later, did. When an Ottawa company (SDL?) got one, that became the largest computer in Canada.
The fastest instructions on the 360-75 (for example, adding two registers) took .39 microseconds.
On my main computer in those days, the IBM 1710 Model 2, the fastest instruction took 70ns, plus 10ns for each pair of digits processed. The -75 was a LOT faster.
Now ordinary PCs are 2GHz or more and execute better than 1 instruction per cycle. Perhaps 1000 times faster than the 360-75.
Bulk RAM is perhaps only 10 times faster. But cache has a large effect.
Disk latencies might not be a lot better: 2,400 RPM vs 10,000 RPM but seek times, capacity, and bandwidth are. Not to mention SSDs.
I cannot even estimate power requirements. And you have to add power for cooling.
The price is perhaps 10,000 times lower.
You can see why I'm not really impressed with x86 improvements since Haswell.
Oh intel has been slacking off lately.
Sure and considering AMD seems to be going to 5nm. Alongside, doing 128cores/256 threads: https://www.tweaktown.com/news/80337/amd-epyc-bergamo-cpus-zen-4-128-cores-2... The article is a little old but this is what seems to be predicted for Zen4 in the enterprise. Nick
participants (3)
-
D. Hugh Redelmeier -
Lennart Sorensen -
Nicholas Krause