|
Kazinsal posted:That's going to cause some interesting stall situations. Imagine you have 16 threads (as in pre-emptive multitasking execution threads, not simultaneous multithreading logical processors) that all want to access the AI accelerator and now need to schedule what order they each get to run their code in. Now you need to worry about saving/flushing/uploading/restoring state between each invocation, across each core, including situations where now you need to save and flush your thread's AI accelerator state to memory because there are 15 other threads ahead of you in the accelerator's work queue and you can't just idle because there's other non-accelerator-touching code to be executed. Mainframe processors are extremely fucky and do weird things with pipelines, instructions, task scheduling, etc. with z/OS, so intuitive reasoning from 'normal' systems might not be that helpful. Mainframes are designed for I/O throughput so a thread (or lots of threads) waiting for output from an on-die 'peripheral' isn't an unusual or sub-optimal state. Conceptually, I do think it's sufficiently similar to CUDA, APUs or QuickSync. Honestly, a 6 TF inferencer is not that impressive and seems more like a spec-sheet check-off so the few hundred middle-managers that buy these things can get approval for their next gen mainframe replacement to their CIO as "AI enabled."
|
#
?
Sep 4, 2021 18:23
|
|
|
|
| # ? Apr 25, 2024 08:49 |
|
|
priznat posted:Lol I was wondering what the ARM royalties are like for Apple, so they have fixed ones they have to re-up after x years or is it more fluid? They have got to be one of the biggest licensees of it. Apple was one of the original investors in ARM and it wouldn’t surprise me if they got a perpetual platform license out of it.
|
|
#
?
Sep 4, 2021 19:38
|
|
|
eschaton posted:Apple was one of the original investors in ARM and it wouldn’t surprise me if they got a perpetual platform license out of it. Yeah, I thought I remembered reading that they have one of the few perpetual / arch licenses. Otherwise I feel they would have started heading away from ARM already / not invested as much into developing their own ARM cores.
|
|
#
?
Sep 4, 2021 20:19
|
|
|
Just ordered an eight core Power9 combo from Raptor Computing yesterday, and 32GB of ECC RAM for it. It's currently on backorder, but I'll give it a month before I start getting antsy. Anyone have prior experience that's useful in getting started, or thoughts on the platform?
|
|
#
?
Sep 25, 2021 13:48
|
|
|
Looks like Apple isn't the only producer of super-wide ARM processors. https://twitter.com/aschilling/status/1466436451361603596 https://twitter.com/aschilling/status/1466436457699201024 This pic in particular is interesting. The design is the opposite of Epyc's, a big compute die surrounded by many IO dies. That must be much better for inter-core latency. 
|
|
#
?
Dec 3, 2021 22:53
|
|
|
Yeah but is Amazon actually ever going to release it to Market outside of having to go through AWS?
|
|
#
?
Dec 3, 2021 23:06
|
|
|
Hah, never.
|
|
#
?
Dec 3, 2021 23:48
|
|
|
I’m sure you’ll be able to get one in an outpost.
|
|
#
?
Dec 3, 2021 23:49
|
|
|
I have the guts of a 90's Alphastation 200 4/233 2 cpus with 256 ram. It was in fairly bad physical shape when I got it, and I had no way of powering it up due to its rather odd configuration. So I took the decision to dismantle it and save the pieces which I still have. It was the administration server for the Computer Science faculty of the University of Melbourne in the mid-90's and clearly they just put it out to pasture and upgraded to better hardware/more modern OS since it was running Digital Unix 3.2c. The FDDI interface seems to have given them no end of trouble, always going down and the ethernet interface never had enough bandwidth and the network setup was complicated since they had all their students work via NFS linked. Fascinating historical document. I also have the obligatory SGI O2 but the CD doesn't work and I do need to replace the CMOS battery, that's going kaput but IRIX CDE is a nice environment, pity that the hardware is so underpowered.
|
|
#
?
Dec 13, 2021 02:53
|
|
|
Not sure where else I would put this, but this is insanely cool: https://youtu.be/ZAu3sXXCZuo
|
|
#
?
Dec 31, 2021 02:35
|
|
|
I have a question. Does anyone know of any electronic computers that stored and processed number in ways not integer or float? I don't mean software systems, like the complicated ways CAS's store and calculate number, or ASCII digits, I mean like pre- or non-IEEE 754.
|
|
#
?
Jan 1, 2022 23:49
|
|
|
Mofabio posted:I have a question. Does anyone know of any electronic computers that stored and processed number in ways not integer or float? I don't mean software systems, like the complicated ways CAS's store and calculate number, or ASCII digits, I mean like pre- or non-IEEE 754. Does this count? http://www.6502.org/users/dieter/bcd2/bcd2_6.htm Or like a 1s-complement system?
|
|
#
?
Jan 1, 2022 23:53
|
|
|
Phobeste posted:Does this count? http://www.6502.org/users/dieter/bcd2/bcd2_6.htm Or like a 1s-complement system? Sure! So the adder is doing native BCD, staying in base-10 rather than going base-10 input -> base-2 for calculation -> base-10 output? I guess I was imagining alternative or exotic number systems. One alternative but common number system example is coin-exchange, where each coin is in fixed fraction with another coin, and they can be combined to make any number (roman numerals are another example of these, where VII is 7 and not 511). These systems aren't generally practical for multiplication, and they aren't one-to-one with integer (roman numerals specifically have ordering and subtraction rules to make it so VII is the valid representation, and not IVI or IIV, but coins are orderless), so my guess is there hasn't been an IC like this, specifically. But I bring it up as an example of what I was thinking: non-integer and non-float number systems. Mofabio fucked around with this message at 00:26 on Jan 2, 2022 |
|
#
?
Jan 2, 2022 00:23
|
|
|
I don't think you're going to find any commercial computers which implement exotic number systems in hardware. Most are nightmarish to implement and nearly all have zero practical value, so why?Phobeste posted:Does this count? http://www.6502.org/users/dieter/bcd2/bcd2_6.htm Or like a 1s-complement system? By the way, IBM is still big on decimal math right up to this day. They threw their weight around enough to get decimal FP formats into IEEE 754-2008, and POWER9 supports decimal in several places. A lot of the demand for decimal math is related to IBM's mainstay for big computers, finance. There are tons of trivial and important decimal fractions which cannot be expressed in a finite number of binary digits. (0.1 base 10 = 0.0001100110011001100... binary). Accountants like it better when the computer's results are an exact match to what they'd see with old school manual computations. Most companies have regarded software implementation of decimal fractions as good enough to support such applications, but not IBM.
|
|
#
?
Jan 2, 2022 07:26
|
|
|
Mofabio posted:Sure! So the adder is doing native BCD, staying in base-10 rather than going base-10 input -> base-2 for calculation -> base-10 output? Maybe look into British business computing prior to the decimal changeover? e.g. https://en.wikipedia.org/wiki/%C2%A3sd#Computing
|
|
#
?
Jan 2, 2022 07:30
|
|
|
doing it in hardware is how you get the massive amount of bandwidth needed for the transaction processing loads mainframes handle a guy who I chat with on Twitter sometimes has been doing IBM’s free online mainframe courses and has shared some of the insane stuff they have now, like not just parity or error correction for RAM but straight up RAID-style hot-swapping
|
|
#
?
Jan 2, 2022 07:36
|
|
|
eschaton posted:doing it in hardware is how you get the massive amount of bandwidth needed for the transaction processing loads mainframes handle z systems have some wild capabilities. Parallel Sysplex, firmware-level hypervisor (to enable LPARs), hardware zip compression.
|
|
#
?
Jan 3, 2022 04:25
|
|
|
carry on then posted:z systems have some wild capabilities. Parallel Sysplex, firmware-level hypervisor (to enable LPARs), hardware zip compression. LPARs well predate the Z series. I believe they were introduced all the way back with System/370, but they were definitely available on ES/9000 systems under MVS. That's where I started my career. Also available at that time: hot-swappable CPUs. A decade later when I started working for Google I recognized their datacenter design (arbitrarily large collections of "ordinary" machines, any of which are allowed to fail at any time) as the inverse of IBM big iron (singular machines capable of terrific throughput, which are not allowed to fail under any circumstance). Edit: if you never worked with true mainframes, the scale of what they were capable of is kinda hard to realize. In 1998, when your "average" PC had just graduated to 8MB of system RAM, I was using IBM Hiperspaces at work. This was a technique that let you read 2GB chunks of data from mass storage, and shove all those records through your batch processing flow, completely in-memory, before writing them back out to storage after processing. mdxi fucked around with this message at 20:24 on Jan 3, 2022 |
|
#
?
Jan 3, 2022 20:19
|
|
|
mdxi posted:LPARs well predate the Z series. I believe they were introduced all the way back with System/370, but they were definitely available on ES/9000 systems under MVS. That's where I started my career. Also available at that time: hot-swappable CPUs. Well yeah they came earlier than z, but z is just a name. It's all the same lineage. Modern z systems are still pretty drat huge, and since it's all a vertical platform the Systems division gets to experiment with some interesting architectural decisions. The next generation is playing around with cache design: https://www.anandtech.com/show/16924/did-ibm-just-preview-the-future-of-caches. No idea if it's going to work but it's pretty out there.
|
|
#
?
Jan 3, 2022 20:43
|
|
|
Mofabio posted:I have a question. Does anyone know of any electronic computers that stored and processed number in ways not integer or float? I don't mean software systems, like the complicated ways CAS's store and calculate number, or ASCII digits, I mean like pre- or non-IEEE 754. Setun was a balanced ternary digital computer. There are electronic analog computers where the computed value is proportional to the voltage in a circuit.
|
|
#
?
Jan 4, 2022 15:59
|
|
|
I know nobody’s kicked this thread in a while, but I fell down a bit of a YouTube wormhole learning a bit about PA-RISC. Anybody have thoughts or impressions? Of all the machines I used in college computer labs or friend’s eBay-harvested collections, they’re one I never managed to run into.
Hasturtium fucked around with this message at 06:55 on Feb 12, 2022 |
|
#
?
Feb 11, 2022 03:45
|
|
|
Hasturtium posted:I know nobody’s kicked this thread in a while, but I fell down a bit of a YouTube wormhole learning a bit about PA-RISC. Anybody have thoughts or impressions? Of all the machines I ran into in college computer labs or friend’s eBay-harvested collections, they’re one I never managed to run into. I have one of these on top of my rack, came with several boxes of accessories, manuals, etc. 
|
|
#
?
Feb 11, 2022 03:50
|
|
|
Hasturtium posted:I know nobody’s kicked this thread in a while, but I fell down a bit of a YouTube wormhole learning a bit about PA-RISC. Anybody have thoughts or impressions? Of all the machines I ran into in college computer labs or friend’s eBay-harvested collections, they’re one I never managed to run into. The stack and the heap grow in the opposite direction to normal, that's the main 'weird' thing I remember about it. They also did the same thing Itanium later did (which HP co-developed) of having massive by the standards of the time caches to make up for so-so performance. You want really weird, try https://en.wikipedia.org/wiki/HP_FOCUS - stack architecture! As for hardware hypervisors and indeed LPARs, an AIX server will also do those just fine, it's not just a mainframe thing. feedmegin fucked around with this message at 13:00 on Feb 11, 2022 |
|
#
?
Feb 11, 2022 12:58
|
|
|
I loaned a PA-RISC system to a friend in college and he never returned it... I'm still slightly salty about that.
|
|
#
?
Feb 11, 2022 17:50
|
|
|
This thread reminded me I had been keeping an eye out for the BeagleV Risc-V SBC, but looks like it got cancelled  Any other Risc-V Pi-alikes coming out?
|
|
#
?
Feb 11, 2022 18:14
|
|
|
priznat posted:This thread reminded me I had been keeping an eye out for the BeagleV Risc-V SBC, but looks like it got cancelled I think it got spun off: https://www.techradar.com/news/a-risc-v-raspberry-pi-rival-is-about-to-hit-the-market
|
|
#
?
Feb 11, 2022 18:15
|
|
|
CommieGIR posted:I think it got spun off: Oh very cool. I'll keep an eye out for that one now!
|
|
#
?
Feb 11, 2022 18:17
|
|
|
I still want to see a Power Microwatt SoC, but god only knows when that’ll happen. Also vaguely bummed that Sifive is stopping production on the Hifive Unmatched. The CPU performance was underwhelming, but I still wanted to play around with a brand new ecosystem. And still waiting, since September, for my Raptor Computing Blackbird kit.
|
|
#
?
Feb 11, 2022 18:20
|
|
|
Will say I'm a little disappointed Nvidia won't get their hands on ARM's IP. That would've been exciting, monopoly aside.
|
|
#
?
Feb 11, 2022 18:22
|
|
|
I mean, they already can/do license pretty much any (non-competitor) IP arm has? $60 billion buys a lot of ip licenses even at a 2% retail royalty rate. They wouldn’t be able to use any of the parter-proprietary ip even if they did acquire them.
|
|
#
?
Feb 11, 2022 19:11
|
|
|
It’s interesting as a lot of SoCs are moving away from Arm to Risc-V, I imagine the sale falling through may slow it down a little bit as Arm may be more willing to get competitive with licensing fees? I don’t think the trend will stop though. Nvidia/arm would have been a solid combo both for enterprise and for consumer stuff. Now I wonder if nvidia joins the risc -v train even!
|
|
#
?
Feb 11, 2022 19:15
|
|
|
priznat posted:It’s interesting as a lot of SoCs are moving away from Arm to Risc-V, I imagine the sale falling through may slow it down a little bit as Arm may be more willing to get competitive with licensing fees? Is it? Honestly most SoCs I see for embedded stuff remain ARM and ARM still has a significant advantage in power consumption. I could be wrong.
|
|
#
?
Feb 11, 2022 19:41
|
|
|
CommieGIR posted:Is it? Honestly most SoCs I see for embedded stuff remain ARM and ARM still has a significant advantage in power consumption. I could be wrong. Well, it's fairly anecdotal I suppose but I know of a couple fairly major players going that way. And from talking to JTAG probe vendors they're seeing a lot of risc-v adoption happening. It won't be total but it's definitely going to take a bite out of Arm's business.
|
|
#
?
Feb 11, 2022 19:56
|
|
|
I have made a terrible, awful, wonderful mistake.
|
|
#
?
Feb 12, 2022 06:57
|
|
|
priznat posted:It’s interesting as a lot of SoCs are moving away from Arm to Risc-V, I imagine the sale falling through may slow it down a little bit as Arm may be more willing to get competitive with licensing fees? I think some SoC vendors and other members of the ecosystem are talking about it specifically to get better deals out of ARM. I don't think there's any kind of seismic shift actually happening. Same as how 32-bit ARM servers were constantly touted as the future a few years ago because that got datacentres better rates out of Intel. Then everyone actually making them, e.g. Cavium, went bust because noone was actually buying it. An NVidia acquisition would possibly have changed this, because NVidia+ARM is direct competition that makes its own silicon in a way that ARM now specifically is not. You don't want your competitor owning your architecture; ARM has got where it has by being conspicuously neutral. feedmegin fucked around with this message at 15:23 on Feb 12, 2022 |
|
#
?
Feb 12, 2022 15:20
|
|
|
feedmegin posted:I think some SoC vendors and other members of the ecosystem are talking about it specifically to get better deals out of ARM. I don't think there's any kind of seismic shift actually happening. Same as how 32-bit ARM servers were constantly touted as the future a few years ago because that got datacentres better rates out of Intel. Then everyone actually making them, e.g. Cavium, went bust because noone was actually buying it. I'm still pissed AMD and Intel dropped their Hybrid x86-64/ARM CPU plans. That could've been gamechanging.
|
|
#
?
Feb 12, 2022 17:35
|
|
|
priznat posted:It’s interesting as a lot of SoCs are moving away from Arm to Risc-V, I imagine the sale falling through may slow it down a little bit as Arm may be more willing to get competitive with licensing fees? The Falcon (embedded GPU control processor / security stuff) is RISC-V I’m pretty sure. For Tegra and friends though, they have the perpetual ARM architecture license still, right?
|
|
#
?
Feb 12, 2022 19:44
|
|
|
movax posted:The Falcon (embedded GPU control processor / security stuff) is RISC-V I’m pretty sure. For Tegra and friends though, they have the perpetual ARM architecture license still, right? I don’t know how the nvidia licensing works, but I would imagine they have a pretty solid relationship, so probably not moving away anytime soon. Especially with all the mellanox stuff being deeply intertwined with arm. For SoCs anything that isn’t requiring datapath processing will be great to move to risc-v. That’s most of what I’m familiar with.. Anything moving from a MIPS32 would be great on a risc-v!
|
|
#
?
Feb 12, 2022 19:50
|
|
|
Kazinsal posted:I have made a terrible, awful, wonderful mistake. Congratulations! You can run Ultrix and BSD and VMS! It’ll run VMS pretty nicely, too.
|
|
#
?
Feb 13, 2022 09:49
|
|
|
|
| # ? Apr 25, 2024 08:49 |
|
|
priznat posted:For SoCs anything that isn’t requiring datapath processing will be great to move to risc-v. That’s most of what I’m familiar with.. Anything moving from a MIPS32 would be great on a risc-v! Surely anything moving from a MIPS32 is on an ARM already, by now. Certainly that's been the case multiple places I've worked.
|
|
#
?
Feb 13, 2022 16:44
|
|