richard@gryphon.COM (Richard Sexton) (04/15/89)
In article <865@savax.UUCP> thompson@savax.UUCP (thompson mark) writes: >In article <1619@dretor.dciem.dnd.ca> king@dretor.dciem.dnd.ca (Stephen King) writes: > >Not foolish at all. Actually, the 120 MOPS is achieved on the 40MHz part >by simultaneously executing a floating point add, floating point multiply, >and an integer ALU operation in a single clock cycle. This doesn't even >include whatever operation the graphics block is performing. Consequently, >the 50 MHz part can attain 150 MOPS. Yes, the 80860 looks like the only decent piece of silicon to come out of intel since the 8051. However, I believe it would be impossible to sustain that FLOP rate as a math coprocessor in an Amiga. Consider the 68881/2 or the Weitek chip: when a floating point instruction is hit, BINGO, the hardware executes it. Bit since the 80860 is not a 680x0 coprocessor, some hardware/software interface will need to be worked out. This is a simplified version; it coulkd take a dozen instructions each for the 680x0 and 80860 to completely pass the operands and resukt back and forth. 68000: Hmm, I need to take the square root of 6.6 80860: dum de dum dd dum 68000: hey 80860, ar you busy ? 80860: nope 68000: here, 80860, heres a number 80860: ok, got it 68000: hey 80860, are you busy ? 80860: nope ? 68000: Ok, do a square root will ya ? 68000: 80860 are you done 80860: of course! Whereas a true 680x0 math coprocessor has none of this nonsense, it just executes inline floating point instructions. So it will take some calculations to determine if the 80860 is actually a superior math processor. Where it would really shine though, would be where you could unbundle a monolithic process, such as a matrix inversion or the like. Clearly this is something a pure math coprocessor could not do. -- ``Parents who have children, have children who have children'' richard@gryphon.COM decwrl!gryphon!richard gryphon!richard@elroy.jpl.NASA.GOV
bakken@arizona.edu (Dave Bakken) (04/15/89)
In article <14716@gryphon.COM>, richard@gryphon.COM (Richard Sexton) writes: > Yes, the 80860 looks like the only decent piece of silicon to come > out of intel since the 8051. Close. Their 80960 (aka p7), which is targetted for embedded applications, is very interesting. It has good performance, but has a 4 deep register cache (registers map into the stack frame) and, more significantly, the hardware can schedule tasks for you. (Interestingly, but perhaps not surprisingly, I have not seen one trade rag mention the HW scheduling. About 1.5 years ago I benchmarked Ada on the chip and worked on a debug monitor for it, and I think that feature is very significant). Also, I think they are offering versions with fault tolerant support (last I heard it was QMR, not TMR, unfortunately). Of course, this chip will live in niche applications that most of us never see. -- Dave Bakken bakken@arizona.edu uunet!arizona!bakken "The fact that I'm paranoid doesn't prove everyone's *not* out to get me"
karl@sugar.hackercorp.com (Karl Lehenbauer) (04/15/89)
In article <14716@gryphon.COM>, richard@gryphon.COM (Richard Sexton) writes: > 68000: Hmm, I need to take the square root of 6.6 > 80860: dum de dum dd dum I was thinking that the 860 could be a concurrent coprocessor; that is, it could run concurrently with the 68000, mostly executing within its own private memory. You could use it to get "reasonable" times on your photorealistic HAM raytracings (just kidding about the HAM part...hilk hilk) and your Mandlebrots -- kind of what the IBM "Wizard" board for the PiS/2 appears to do. -- -- uunet!sugar!karl | "Time is an illusion. Lunchtime doubly so." -- | -- Ford Prefect -- Usenet BBS (713) 438-5018
thompson@savax.UUCP (thompson mark) (04/21/89)
In article <14716@gryphon.COM> richard@gryphon.COM (Richard Sexton) writes: >Yes, the 80860 looks like the only decent piece of silicon to come >out of intel since the 8051. > >However, I believe it would be impossible to sustain that FLOP rate >as a math coprocessor in an Amiga. > >Consider the 68881/2 or the Weitek chip: when a floating point >instruction is hit, BINGO, the hardware executes it. > >But since the 80860 is not a 680x0 coprocessor..[stuff about a slow interface] I didn't realize the Weitek part was so compatible with 680x0's. Actually, I never envisioned the i860 (Intel N10) as replacing the 68881/2 but using it as a somewhat dedicated auxillary processor in some sort of multi-processor scheme with shared memory. The application I specifically had in mind was a graphics card using the i860 for display list traversal, 3D transformations, smooth shaded solid rendering, and hidden surface removal. The Amiga would simply DMA the display list into the i860's memory and off it would go cranking out Phong shaded unicycles in real time. (This implies that you put the display buffer on the i860 graphics card.) The thing to remember is that the i860 is processor all by itself (unlike the 68881/2), but I guess you knew that already. >Where it would really >shine though, would be where you could unbundle a monolithic >process, such as a matrix inversion or the like. Clearly this >is something a pure math coprocessor could not do. Yeah. -------------------------------------------------------------------------- | Mark Thompson | | decvax!savax!thompson Designing high performance graphics | | (603)885-9583 silicon today for a better tomorrow. | --------------------------------------------------------------------------
raz@kilowatt.uucp (Raz- Berry) (04/22/89)
From bilan@morder Fri Apr 21 17:08:47 1989 Return-Path: <bilan@morder> Received: from morder. by kilowatt. (4.0/SMI-4.0) id AA00571; Fri, 21 Apr 89 17:08:34 PDT Received: by morder. (4.0/SMI-4.0) id AA04597; Fri, 21 Apr 89 17:02:17 PDT Date: Fri, 21 Apr 89 17:02:17 PDT From: bilan@morder (Steve Bilan ext. 31031) Message-Id: <8904220002.AA04597@morder.> To: raz@kilowatt, rlam@neptune Subject: sysdiag bug with e-net pal change Cc: donald@dswalker, hsiegel@margo, iwerness@broadmoor, kho@ho, rwlee@onyx Status: R The test problem that you ran into (for ECO G579) does not really seem to be a bug. I ran sysdiag on two 3/E systems (with the new pals) overnight and only had one problem. When the ethernet cable was unplugged, sydiag generated the message: enet - no replies to broadcast echo packet on enet0 But after plugging the cable back in, the devtop test would work again. It is still running at this point; I will continue to run it over the In article <877@savax.UUCP> thompson@savax.UUCP (thompson mark) writes: )In article <14716@gryphon.COM> richard@gryphon.COM (Richard Sexton) writes: ))Yes, the 80860 looks like the only decent piece of silicon to come ))out of intel since the 8051. )) ))However, I believe it would be impossible to sustain that FLOP rate ))as a math coprocessor in an Amiga. )) ))Consider the 68881/2 or the Weitek chip: when a floating point ))instruction is hit, BINGO, the hardware executes it. )) ))But since the 80860 is not a 680x0 coprocessor..[stuff about a slow interface] ) )I didn't realize the Weitek part was so compatible with 680x0's. )Actually, I never envisioned the i860 (Intel N10) as replacing the 68881/2 I think we are a little confused here, There is no way that the Weitek can possibly emulate a 68881. Not without a major hardware kluge. I think that richard means that the Weitek chip SET (IU and FPU) together function simalirly to the 680X0 and coprocessor series. )but using it as a somewhat dedicated auxillary processor in some sort of )multi-processor scheme with shared memory. The application I specifically had )in mind was a graphics card using the i860 for display list traversal, )3D transformations, smooth shaded solid rendering, and hidden surface removal. God this sounds familiar. Didn't there used to be a company that did stuff like this? You Know what else this fictional board needs... a standard graphics language. How about PHIGS+? Call me an Amiga PHIG. )The Amiga would simply DMA the display list into the i860's memory and off )it would go cranking out Phong shaded unicycles in real time. (This implies )that you put the display buffer on the i860 graphics card.) The thing to )remember is that the i860 is processor all by itself (unlike the 68881/2), )but I guess you knew that already. Exactly! This would be the IDEAL application for this chip. Putting this baby on or at the mercy of the Amiga bus would strangle it's performance. Of course if you do the board right, you could include the Clown... This would be neet, but I fear that it would easily cost more than the entire machine (+Caligrapher).-- Steve -Raz- Berry Disclaimer: I didn't do nutin! UUCP: sun!kilowatt!raz ARPA: raz%kilowatt.EBay@sun.com "Fate, it protects little children, old women, and ships named Enterprize"
richard@gryphon.COM (Richard Sexton) (04/26/89)
In article <33739@kilowatt.uucp> raz@sun.UUCP (Steve -Raz- Berry) writes: > >I think we are a little confused here, There is no way that the Weitek >can possibly emulate a 68881. Not without a major hardware kluge. I think >that richard means that the Weitek chip SET (IU and FPU) together function >simalirly to the 680X0 and coprocessor series. No, Richard meant it was a co-processor, like the 68881/2. Thats what the ads implied. So I called Weitek. Buggers. The stupid thing is memory mapped. You write the operands into memory addresses, then give it an operation, then poll it for completion. Some co-processor. Some performence data: Linpack - single precision 8 MFLOPS double 6 MFLOPS Whenstone Single precision 2.0 MWhets (is that what he said ? Mwhets ?) Double 1.2 How do these figures compare with '881/2 and 860 ? Some interesting stuff about this part - it doesnt exist. It seems Weitek has a nice math chip, the 3116 or something like that. They sell a daughtorboard that has some glue logic and this chip for use with the 386. What they have for the 68000 is a board that uses the 3116 and some glue. Every time I pumped him for information, he pumped me wanting to know how many i needed, so thay could make their projections and deicide to make this chip or not. Hell, I told 'em 10,000. Couldnt hurt. Weitek is in Sunnyvale. -- ``Bring me the head of fettucini alfredo'' richard@gryphon.COM decwrl!gryphon!richard gryphon!richard@elroy.jpl.NASA.GOV
raz@kilowatt.uucp (Raz- Berry) (04/28/89)
In article <15147@gryphon.COM> richard@gryphon.COM (Richard Sexton) writes: )In article <33739@kilowatt.uucp> raz@sun.UUCP (Steve -Raz- Berry) writes: ))I think we are a little confused here, There is no way that the Weitek ))can possibly emulate a 68881. Not without a major hardware kluge. I think ))that richard means that the Weitek chip SET (IU and FPU) together function ))simalirly to the 680X0 and coprocessor series. )No, Richard meant it was a co-processor, like the 68881/2. Thats what )the ads implied. Your kidding. I don't claim to know all, but I thought I woulda heard that. Oh well. )So I called Weitek. )Buggers. )The stupid thing is memory mapped. You write the operands into )memory addresses, then give it an operation, then poll it for )completion. Some co-processor. Not a true co-processor if you ask me, ok so it fits the description... but it's a kluge. )Some performence data: )Linpack - single precision 8 MFLOPS ) double 6 MFLOPS )Whenstone Single precision 2.0 MWhets (is that what he said ? Mwhets ?) ) Double 1.2 Did it say how this was set up? What I mean is, is it a 680x0 peripheral or is it running it's own code memory (probably). If it needs it's own special memory, like cache, then you might as well build a seperate board and do it up as black box math server. If you are going to go to that much trouble, might as well go with the '860 and get REAL speed. I guess I define co-processor as a transparent hardware accelerator. If you can hook it up to the main processor, with minimum amount of hassel, and have it run out of processor memory space, then it's a co-processor. My definition. Weitek would probably disagree. )How do these figures compare with '881/2 and 860 ? I have no clue. I looked in the 881 manual and couldn't find the performance figures. )Some interesting stuff about this part - it doesnt exist. It seems )Weitek has a nice math chip, the 3116 or something like that. They )sell a daughtorboard that has some glue logic and this chip )for use with the 386. What they have for the 68000 is a board )that uses the 3116 and some glue. Is it compatable with the 881 instruction set? If not, why bother? You'ld have to write your own compiler for it. YECHH. Unless it does sin(x)/cos(y) in two cycles... )Every time I pumped him for information, he pumped me )wanting to know how many i needed, so thay could make their )projections and deicide to make this chip or not. Should have asked for a sample ;-) )Hell, I told 'em 10,000. Couldnt hurt. Weitek is in Sunnyvale. Gee, I can pick one up on my way to work! )richard@gryphon.COM decwrl!gryphon!richard gryphon!richard@elroy.jpl.NASA.GOV I don't like Weitek. When I was at Raster Tech. I talked to a few of the engineers that had worked on previous projects involving their 64 bit FPU/IU chip set. I heard nothing but horror stories about hardware bugs and instructions that didn't work as advertized. Abort and stall are good examples. -- Steve -Raz- Berry Disclaimer: I didn't do nutin! UUCP: sun!kilowatt!raz ARPA: raz%kilowatt.EBay@sun.com "Fate, it protects little children, old women, and ships named Enterprize"
dillon@POSTGRES.BERKELEY.EDU (Matt Dillon) (05/01/89)
:No, Richard meant it was a co-processor, like the 68881/2. Thats what :the ads implied. : :So I called Weitek. : :Buggers. : :The stupid thing is memory mapped. You write the operands into :memory addresses, then give it an operation, then poll it for :completion. Some co-processor. The neat thing about the Weitek chip is that it has a stack. Results are automatically pushed on the stack. So, something like: (a + b) + (c + d) is: fpush a, fpush b, fadd, fpush c, fpush d, fadd, fadd, fpop (to) Mem That is, temporaries need not be moved off the FP unit. The sequent (parallel processor UNIX machine (running dynix)) uses the Weitek chip. Our current configuration has 12 386 boards (when you are running UNIX you generally don't care about the processor), each with a Weitek fp unit. While not a super computer, the thing can, running 12 compiles in parallel, compile all 246 C files (about half of postgres, the rest being in lisp at the moment) in less than 5 minutes. -Matt
mcp@ziebmef.uucp (Marc Plumb) (05/04/89)
In article <15147@gryphon.COM> richard@gryphon.COM (Richard Sexton) writes: >The stupid thing is memory mapped. You write the operands into >memory addresses, then give it an operation, then poll it for >completion. Some co-processor. Well, it still counts as a coprocessor. The coprocessor instructions it ises aren't the F-line ones, they're loads and stores with magic values, but there are bytes you can stick in the instruction stream to make the thing compute. Isn't that the point? It is true that dedicated instructions can be decoded faster *if you really work at it*, but have you ever looked at the 68020 coporcessor protocol? Nice and general, but *slow*. The 6888[12] is also memory mapped, albeit in a separate address space (FC = 111), and the processor writes the instruction word to a certain place, then reads another to see what needs doing, and generally keeps polling the coprocessor until it tells the 020 it can go away. Generally disgusting. The only difference is the FC bits (that's why a 68010 helps a lot - it lets you control them explicitly) and the fact that the load/store sequence is done by microcode, not user code. And I'll bet long odds Weitek's protocol is more efficient. -- -Colin