Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!cornell!rochester!udel!nelson From: nelson@udel.EDU (Mark Nelson) Newsgroups: comp.arch Subject: Re: CDC 205 et al Keywords: MegaFLOPS Message-ID: <14353@louie.udel.EDU> Date: 30 Apr 89 21:13:44 GMT References: <568309@vaxa.uwa.oz> Sender: usenet@udel.EDU Reply-To: nelson@udel.EDU (Mark Nelson) Organization: University of Delaware and Cray Research Lines: 87 In article <568309@vaxa.uwa.oz> g_ahrendt@vaxa.uwa.oz writes: >1. In regard to an article posted a while back stating that the CDC CYBER 205 is >slower than a Cray 1. This is incorrect. The CDC CYBER 205-424 has been clocked >around 800 MIPS, making it faster than the Cray 1, ETA10-P, ETA10-Q, NEC SX-1, ^^^^^^^^ >Cray X/MP-2, IBM Sierra 3090-400/VF, NEC SX-1E..... > It's important to make the distinction between MIPS and MEGAFLOPS. The 205 has a 20-nanosecond clock cycle, making it capable of 50 MIPS (one instruction issued per clock). For vector floating point, it has up to four parallel vector "pipes". And each pipeline is capable of performing a simultaneous add and multiply (within strong limitations). So we're up to 50 * 4 * 2 = 400 MegaFLOPS. The machine gains another factor of 2 by going to 32-bit floating point data instead of the normal (for supercomputers) 64-bit floating point. That gives a peak of 800 MegaFLOPS, but 50 MIPS. Looking at another uniprocessor, the NEC SX-2 has a 6-nanosecond clock (167 MIPS), 4 pipes, and two operations per pipe yielding 167 * 4 * 2 = 1,333 MegaFLOPS. For comparison purposes, the Cray-1 gives: 12.5-nsec clock = 80 MIPS with 2 operations/clock yielding 160 MegaFLOPS. Going on to multi-processor machines: Machine Clock MIPS Pipes FLOPS/ Processors 32-bit Total (nsec) pipe speed-up (MegaFLOPS) Cray X-MP/4 8.5 118 2 4 941 Cray-2 4.1 122* 2 4 1951 ETA-10G 6.5? 154 2 2 8 2 9846 Cray Y-MP/8 6. 167 2 8 2667 NEC SX-X 2.9 345 4 4?? 4 22069 Notes: MIPS I'm referring to MIPS in one processor. * The Cray-2 issues instructions at most every other clock period so this is the actual peak MIPS/processor. For vector instructions, results are produced every clock cycle. ? I'm not positive about this number. Can anybody tell me what the fastest clock cycle on any shipped ETA machine is? ?? Since the SX-X is still the apple in a corp (pun intended) of NEC engineers' eyes, I don't have very complete information about it. I have numbers for all the other columns for the SX-X, but without this fudge factor I can't get them to work out. I'm sorry if I left out your favorite machine--these were the only machines I knew the numbers for off the top of my head, other than the Cray-3, and since I work for Cray I didn't feel it would be appropriate to comment on unreleased Cray machines (numbers for the Cray-3 are easily available). Remember that theses are peak speeds for the machines; sort of computational speeds of light, and average performance for any mix of programs is likely to be much lower. >2. A second article displayed the following : >> ... > >This is also incorrect, whereas the following gives a more accurate picture, >based on machines actually available, and not scheduled for production. > >1. ETA US 10-G 10.3 Billion Instructions per Second I get 1230 MIPS >2. Thinking Machines US CM-2 10 Billion Instructions per Second >3. West German Govt. D Suprenum 1 5 Billion Instructions per Second >4. ETA US 10-E 3.4 Billion Instructions per Second 760 MIPS (10.5 nsec clock 8 processors) >5. NEC J SX-X44 3.2 Billion Instructions per Second 1380 MIPS (not yet available) >6. Cray US X/MP-416 2.8 Billion Instructions per Second 470 MIPS (basically no longer available new) >7. Thinking Machines US CM-1 2 Billion Instructions per Second >8. NEC J SX-2 1.3 Billion Instructions per Second 167 MIPS 9. Cray US Y-MP 1333 MIPS I've never seen any numbers on the Suprenum project, so I can't comment on it. I don't have any of the details on the Thinking Machines CM-1 and CM-2 with me (I'm home, they're at work) so all I can say is that since they are SIMD architectures the above BIPS numbers can't be right, although the numbers might be right for BOPS. By the way, Thinking Machines has great sales brouchures--at least one even has moving parts. Mark Nelson ...!rutgers!udel!nelson or nelson@udel.edu This function is occasionally useful as an argument to other functions that require functions as arguments. -- Guy Steele