IBM's zEnterprise 196 CPU: Cache is king - 'The fastest CPU in the world.' | Hacker News Reader

IBM's zEnterprise 196 CPU: Cache is king - 'The fastest CPU in the world.'

theregister.co.uk

34 pointsableal19 comments

Not really all that impressive, actually underwhelming and anti-climactic.

The interconnects are where you see that Intel and AMD are catching up with them ... 60GB/sec interconnects? HyperTransport 3 can do 25.6GB/sec per CPU, and you can put up to 8 of them in a 3U or 4U box, and still use less power than just one of the IBM CPUs.

5.2Ghz, when they had all sorts of fancy water-cooling and other tricks available in what is essentially a cost-no-object CPU package? Yawn...

IMO what's interesting is the 94MB of L4 cache shared among 80 usable cores at 5.2Ghz. Latency is really just a question of the speed of light and there are workloads where this could be 20+ times as fast as that 8 CPU cluster in a 3U box. The value is "GPU with cache" not a fast x86.

I read that spec differently: 96MB of L4 per board (6 sockets/24 cores per board), and up to 16 boards in one machine. So the amount of cache, is vast, but not exactly shared equally between all sockets.

Call me back when an AMD box reaches 5 9's...

Seriously: x86 servers are good for some things. But these boxes can cut through a lot of workloads at speed and reliability levels a PC cannot dream of.

The first mainframe I used, in the mid-80's, had more than 200 3278s connected to it. I see a lot of current servers that can't handle 200 concurrent requests.

One of my oldest client's quad-Xeon (900Mhz CPUs), 6GB RAM system running Oracle has been up for 932 days at this point.

That is 5 9's (actually better, it is 100%) over 2.5 years. I am sure that similarly configured Opteron based systems are out there with similar uptimes.

Sure, mainframes are reliable, that wasn't my point however - AMD and Intel are catching up in terms of interconnects and other technology, at 1/1000th the price point.

The processors and interconnects may be faster, but that's not the whole story. The X-15 was fast, but you wouldn't use it to deliver FedEx packages. Or shipping containers.

x86s are fast and, if your problems allow that, you can gang a bunch of them to crunch your workload faster and cheaper than any mainframe could, but that does not make these machines even a little bit less impressive than they appear.

That's just absurd. Are we really that dense to compare interconnected x86s to shared memory parallel mainframes? Not even Myrinet or SCI (and the various other contenders of the throne) could be half as effective as point-to-point processor interconnects, and that's an overstatement.

To make such a comparison without understanding the workloads these mainframes are "crunching" is not very wise.

edit: Wow. Must've hit a nerve, I guess. Downmod me to oblivion, sir!

No way I would downmod you - that's exactly what I was saying. You may put certain workloads on racks of Dell servers, but there are workloads that are best served by these kinds of machines (SGI has some shared-memory monsters too, with x86 processors) and I say these IBM machines are impressive works of art.

But I am biased. I even love the 3278 font with those sleek 6's and 9's.

Well, actually it DOES make them less impressive... because once you look, actually look objectively, at how x86 continues to close the gap, you realize that the day is coming when mainframes will go away or be otherwise not distinguishable.

Look: when the 486 came out circa 1989, its memory bandwidth to RAM was about 70MB/second. That was the speed of a single disk channel for an IBM mainframe, a huge difference in performance specs.

Interconnects? NUMA? Heck the 486 wasn't even good for 2-way SMP - while the mainframe could be hooked up into a clustered sysplex (1994, or the earlier virtual coupling facility) and even mirrored over long distance links (GDPS).

Look at the way these differences have been minimized in recent times ... the mainframe chip borrows its GX IO chip from POWER, fer cryin' out loud...

Your logic is flawed. The impressiveness of current and near-future x86 processors (and they are, considering they are still compatible with 8088's, quite impressive) has no relation to how impressive the x196 and its support chips are.

To use my previous analogy, the X-15 or the SR-71 are impressive machines, but that doesn't make a 8,000 TEU freighter, or an aircraft carrier, less impressive.

The differences in basic technology are being reduced, but all it does is to make the current crop of x86s look less crippled compared to POWER and zCPUs. And, mind you, POWER makes a top-of-the-line x86 look like a toy.

I cannot find any specs regarding the interconnect in z196. Can you please point to where you got that 60GB/s number from?

Also, would anyone happen to know any details on it? With 6 interfaces, I suspect cross-connected CPs; is that accurate?

"RAID-protected DDR3 main memory", most desktop class machines don't even support ECC.

Uh... even 3 years ago, some x86 servers were supporting DIMM mirroring, which is "RAID1" .

See http://www.redbooks.ibm.com/abstracts/tips0671.html and note publication date of 2007.

Spoiler: It's not really the fastest, just the highest-clocked (5.2 GHz).

They couldn't remember the fastest CPU in the world, this is just a tribute.

and the peculiar thing was, the CPU that was released, it didn't perform anything like this CPU

The article mentions that it is the highest-clocked, but I didn't see anything saying that it's not the fastest, too. It has not only the raw clock speed but several cores and a truckload of cache.

You'd expect to see some benchmark numbers if so.

Impressive. I'll continue to dream about Big Blue.

Reply on news.ycombinator.com