david@elroy.jpl.nasa.gov (David Robinson) (11/21/89)
Can anyone with any detailed knowledge compare the new MIPS R6000 with
the BIT 80Mhz SPARC (B5000)? BIT is claiming 65 MIPS while the R6000
is rated at 55 MIPS. I thought both the R6000 and B5000 were manufactured
by BIT and probably with the same ECL process.
I know most of the standard MIPS vs SPARC arguments (reg windows etc)
so I am interested in how the implementations differ and the types
of trade offs that were made. An article in SunTech Journal (aka
Sun's trade rag) talks about the trade offs in their cache design.
Any comments?
--
David Robinson elroy!david@csvax.caltech.edu ARPA
david@elroy.jpl.nasa.gov ARPA
{cit-vax,ames}!elroy!david UUCP
Disclaimer: No one listens to me anyway!sritacco@hpdml93.HP.COM (Steve Ritacco) (11/22/89)
As I understand it the R6000 is only running at 67 MHz the 80 MHz version will be out some time next year. Mips pays a lot of attention to cache design, and is quoting system performance when they say 55 mips, not raw CPU performance as the BIT SPARC numbers.
mslater@cup.portal.com (Michael Z Slater) (11/22/89)
>Can anyone with any detailed knowledge compare the new MIPS R6000 with >the BIT 80Mhz SPARC (B5000)? BIT is claiming 65 MIPS while the R6000 >is rated at 55 MIPS. I thought both the R6000 and B5000 were manufactured >by BIT and probably with the same ECL process. I wouldn't put much signficance on those performance numbers. The MIPS rating is MIPS Co.'s rating of their *system* performance, and thus includes the particular cache structure, memory system, etc. The SPARC numbers are just estimates based on looking at the CPU chip, and (I believe) neglect any cache effects. No one has yet shown a system based on the BIT SPARC chips; MIPS has shown their system, but concedes that its not ready for benchmarking. Note that I'm not saying that the SPARC chip might not be faster; just that with the information available, you can't tell. The implementation differences are major. The MIPS design includes some cache control and MMU logic on the CPU chip, while the SPARC design provides absolutely no support for MMU or cache. The MIPS design includes a bunch of ECL gate arrays to hook up the cache RAMs to the processor. Perhaps because they included more functions on the CPU chip, the MIPS design is initially rated at 66.7 MHz, while the SPARC design is rated at 80 MHz. Michael Slater, Microprocessor Report mslater@cup.portal.com 550 California Ave., Suite 320, Palo Alto, CA 94306 415/494-2677 fax: 415/494-3718
mash@mips.COM (John Mashey) (11/22/89)
In article <1989Nov21.015953.13817@elroy.jpl.nasa.gov> david@elroy.jpl.nasa.gov (David Robinson) writes: >Can anyone with any detailed knowledge compare the new MIPS R6000 with >the BIT 80Mhz SPARC (B5000)? BIT is claiming 65 MIPS while the R6000 >is rated at 55 MIPS. I thought both the R6000 and B5000 were manufactured >by BIT and probably with the same ECL process. Right now, this is probably premature in most possible ways. 1) It is no secret that both are manufactured by BIT with the same ECL process. 2) The BIT claims are for a chipset, whereas MIPS is claiming a system that yields 55-(MIPS)-mips. Sun has not yet published the detials of the SYSTEM(s) that they're building. Note that chipsets and systems are different [in my car model, one is measuring RPM at the engine, and the other speed on the road]. This doesn't say one is better or worse than the other, just that they're different. BIT isn't building systems, and their claims imply few committments about what Sun is doing with the chips. 3) As always, especially on these fast machines, relative performance ratings will vary around quite a bit, i.e., the faster things get, the worse an approximation a single-number-mips-rating is. 4) Remember that different vendors (legitimately) have different ideas of -mips ratings, and you believe they're equivalent at your peril; this has been true forever in the computer business, starting (most notably) with the difference between IBM-mainframe-mips and VAX-mips ratings. Vendors can call their -mips whatever they want, but they're essentially meaningless except when you have comparable & realistic benchmarks. REMEMBER A VENDOR CAN LABEL THEIR MACHINES WITH ANY MIPS-RATING THEY LIKE, AND MAYBE THEY'LL BE CONSISTENT WITHIN THEIR PRODUCT LINE (or not), BUT THE PROBLEM COMES WHEN YOU MAKE EVEN THE SLIGHTEST ASSUMPTION THAT YOU CAN COMPARE THEM. This is like claiming that one car has better gas mileage than another because its figures are given in km/gallon instead of miles/gallon. A useful thing to have is conversion factors, if such exist, between different vendor's -mips ratings. (I say, "if such exist", carefully: a) It might be that one number gives a +/-10^% estimate. b) It might be that one for integer and one for floating-point is needed. c) It might that a whole bunch more are needed, especially as one gets faster machines.) For example, one observes that on the SPEC benchmark set, although the relative performance bounced around, the (16-SPARC-mips) SPARCsystem 330 and (13-MIPS-mips) MIPS M/120 are quite similar in overall performance. (It really helps to see the charts superimposed, which unfortunately do not post very well on the net :-) Needless to say, cost/mips comparisons are mostly meaningless to compare vendors with, if you can't calibrate the kinds of -mips used. Right now, there is insufficient public data to decide which of the two ECL SYSTEMS is faster, or for which benchmarks, even for the simplest single-user cases, much less more complex environments. -- -john mashey DISCLAIMER: <generic disclaimer, I speak for me only, etc> UUCP: {ames,decwrl,prls,pyramid}!mips!mash OR mash@mips.com DDD: 408-991-0253 or 408-720-1700, x253 USPS: MIPS Computer Systems, 930 E. Arques, Sunnyvale, CA 94086
rathmann@Neon.Stanford.EDU (Peter K. Rathmann) (11/23/89)
Ok, R6000 vs BIT SPARC is too early to call, but other comparisons might be even more interesting. The R6000 based system was billed in the popular press as a mainframe, so does anyone have any numbers comparing it to more typical mainframes like the IBM 3090 or big Amdahl machines? Extrapolating from the xfroot numbers indicates that the R6000 should be pretty competitive, at least for floating point. Has anyone checked it for a business data processing mix of databases and COBOL? If the R6000 can compete with classic mainframes on their home ground, I might start believing this talk about killer micros. -Peter
mash@mips.COM (John Mashey) (11/29/89)
In article <1989Nov22.233508.20446@Neon.Stanford.EDU> rathmann@eclipse.Stanford.EDU (Peter K. Rathmann) writes: >Ok, R6000 vs BIT SPARC is too early to call, but other comparisons >might be even more interesting. The R6000 based system was billed in >the popular press as a mainframe, so does anyone have any numbers >comparing it to more typical mainframes like the IBM 3090 or big >Amdahl machines? Please note: it's called an "Enterprise Server" rather than a mainframe; the 6280 brochure doesn't ever say "mainframe", on purpose. The claim is that the CPU performance is mainframe-uniprocessor class, and that it's got pretty reasonable I/O. According to a posting of Mike Taylor's a while back, a few relevant numbers would be, for the fastest scalar mainframe for which I have numbers: Amdahl 5990 CPU RC6280 Cycle time 10ns 15ns (@67Mhz) LLNL 64, Harmonic 11.9 MFlops 8.8 MFLOPS LINPACK, 64, FORT 14.5 MFLOPS 10.3 MFLOPS LINPACK, 32, FORT 16.8 MFLOPS 13.9 MFLOPS Dhrystones (1.1) 90Kish? 109K, -O3 Grossly, as far as I can tell: An Amdahl 5990 CPU is faster than a 5890 (for sure!) and faster than the fastest current IBM 3090 CPU, assuming all scalar stuff (no vector units). @ 67.5Mhz, a 6280 CPU seems faster than a 5890 or 3090, but a little slower than a 5990, at least on floating-point. There is insufficient comparative data to really know much on the integer side. (Dhrystone is also good for S/370-bashing, not just VAX-bashing.) I suspect a 6280 might equal a 5990 on the integer stuff, but more data is really needed. One would have to see a lot more data to know a lot more, especially given the different memory systems [the Amdahl brings joy to SRAM vendors: up to 512MB of it...; the 6280's secondary cache is "only" 512KB. This certainly will help the Amdahl's performance in huge multi-user situations, although the 6280 has been designed to be pretty reasonable for that purpose also (512KB of physical, 2-way set-associative cache helps a lot).] Of course, you can put 4 CPUs in the Amdahl box, whereas we've announced 1. I think a 5990 CPU would be called about 60-65 VUPs (?; note that IBM-mips and VUPs are different, so this corresponds to nothing they claim.) A full 5990 box (according to a posting a while back) has a peak aggregate I/O rate of 500 MB/sec, or about 2 MB/sec per VAX-mip, assuming that each CPU is around 65 VUPs. The RC6280's biggest I/O configuration will allow 6 VME busses, for 33MB/sec peak aggregate rate, also around 2-3 MB/sec per VUP (well, it looks more like 3, but I suspect that a 5990 can get closer to its peak than we can, given the different memory architectures.) There is insufficient data in those numbers, given the completely different architectures, to know much, except that the biggest announced configurations sound like they have grossly have similar ratios of I/O bandwidth to compute power. The smallest 6280 has about .5 MB/sec per VUP. > >Extrapolating from the xfroot numbers indicates that the R6000 should >be pretty competitive, at least for floating point. Has anyone checked >it for a business data processing mix of databases and COBOL? It is specifically tuned to be good at such things; in particular, the cache design is partially aimed at that issue. > >If the R6000 can compete with classic mainframes on their home ground, >I might start believing this talk about killer micros. Again, it's NOT a mainframe: it doesn't have the direct, immense connectivity of a large mainframe. It probably does have the connectivity of a small mainframe, and (about) the CPU performance of the faster large uniprocessors. (The following is doing its best not to be a commercial, but it's hard :-) Just to fix a misconception I've now seen several times, such machines are NOT expected to put IBM or Amdahl out of business, and are NOT expected to be one-for-one replacements for big mainframes. No one in their right mind would position this product that way. If not that, then what are these things (big Killer-Micros, not just MIPS-based ones) for: a) Upgrades from smaller micro-based systems, in either commercial or technical markets. (Again, there are fairly large numbers of UNIX micros in the commercial markets, not just technical ones.) b) Offload selected existing applications from mainframes, mostly to avoid upgrades at awkward times, or to survive budget pressure. For example, big OLTP things running under MVS aren't about to move any time soon to any of the fast micros ["any time soon" means "in the forseeable future"]. On the other hand, suppose you have a mainframe complex that's running: OLTP with IMS or DB2 CAD number-crunching program development and it's out of steam. You might be tempted to move some of the CAD off, and maybe some program development (many computer centers have done this for years anyway, but some applications were "stuck" on the mainframes for speed or size needs). Thus, you would free up cycles for your OLTP work (which will probably never move). Also, maybe the the CAD department decides it could afford to have the comp center run a K.M. for them, for their dedicated use. "Avoiding upgrades at awkward times" : anybody who's spent any time in the mainframe business knows there are good times in product lifecycles to buy new things, and other times that are not so good. Sometimes, putting off a major upgrade for 6 months can save a lot of money. c) Suppose your company uses some relational database that runs on both mainframes and various UNIX systems, and you're implementing a new application based on it. In some cases, a Killer Micro offers another point in the cost/performance space that may fit what you need, where a bunch of small micro-based systems just don't fit. In general, if you look at the history of computing, new kinds of computers have either siphoned off growth from older kinds, or opened new markets, rather than engaging in direct 1-on-1 replacement battles. A reasonably objective view of the K.M. machines for the commercial market says: a) There are shops that will never, ever have them (not surprising, as there are shops that have only recently come to believe that companies like Amdahl, Tandem, or DEC are substantial enough to deal with. Some haven't gotten that far... for such places, companies like MIPS or even Sun simply don't exist...) Fotunately, many K.M.s will get sold thru other larger companies. b) Enough places simply want faster UNIX boxes, or if they can get mainframe CPU speeds from something whose cost/VUP is 10X better, and is fast enough to get the job done in time, will happily move selected applications off, or at least consider K.M. boxes as alternatives for growth. -- -john mashey DISCLAIMER: <generic disclaimer, I speak for me only, etc> UUCP: {ames,decwrl,prls,pyramid}!mips!mash OR mash@mips.com DDD: 408-991-0253 or 408-720-1700, x253 USPS: MIPS Computer Systems, 930 E. Arques, Sunnyvale, CA 94086