|
movax posted:As for challenging x86...eh, I don’t think so, for RISC-V. At least not primarily; it’ll insert itself into the market in other segments first. Eating ARM's lunch in the super low end most likely.
|
# ? May 21, 2018 02:57 |
|
|
# ? Apr 23, 2024 22:22 |
|
ConanTheLibrarian posted:The Mill shall vanquish all lesser architectures. as soon as they get around to actually making one, I'm sure they'll dominate the world
|
# ? May 21, 2018 03:03 |
|
BobHoward posted:as soon as they get around to actually making one, I'm sure they'll dominate the world Speaking of Mill Computing posted:What do you get by joining us? LOL they still can't attract VC so they can't afford to pay anyone
|
# ? May 21, 2018 03:06 |
|
Every time someone tries VLIW architectures for anything other than DSPs, it ends up being too much of a pain in the rear end to be worth it.
|
# ? May 21, 2018 04:12 |
|
ConanTheLibrarian posted:The Mill shall vanquish all lesser architectures. Boy, are you guys going to have egg on your faces... a little birdie just told me that the Mill guys are in advanced negotiations to license the Intel Sufficiently Smart Compiler™ from the Itanium team.
|
# ? May 21, 2018 06:19 |
|
BobHoward posted:LOL they still can't attract VC so they can't afford to pay anyone
|
# ? May 21, 2018 10:47 |
|
lDDQD posted:Every time someone tries VLIW architectures for anything other than DSPs, it ends up being too much of a pain in the rear end to be worth it. Aren’t the Tensilica Xtensas VLIW? Or is more that they are super customizable to add on instructions and people often add VLIW type stuff. Somedays I wish I worked on projects with big enough budgets where playing with custom SIP is a requirement. Partially out of morbid curiosity as to what the toolchain and collateral is like for those cores and parts.
|
# ? May 21, 2018 12:37 |
|
movax posted:Aren’t the Tensilica Xtensas VLIW? If those are the AMD TrueAudio thingies they could go up to 2x VLIW for DSP work. e: LIW Longish Instruction Word
|
# ? May 21, 2018 13:25 |
|
looks like 8 core Coffee Lake really is coming, and maybe even compatible with Z370: https://videocardz.com/newz/intels-8-core-kaby-coffee-lake-chip-spotted
|
# ? May 21, 2018 18:13 |
|
If the base clock is 2.6Ghz color me unimpressed.
|
# ? May 21, 2018 18:30 |
|
Not sure why one would worry about base clock when there's all-core turbo boost? It's seems to be pretty useless number these days, though I wouldn't be surprised to learn that Intel uses it to calculate TDP (which is equally useless). ... and because no posting in the Intel megathread is complete without another 10nm delay: https://twitter.com/david_schor/status/998608222217891841 10 core eames fucked around with this message at 19:22 on May 21, 2018 |
# ? May 21, 2018 19:19 |
|
eames posted:Not sure why one would worry about base clock when there's all-core turbo boost? It's seems to be pretty useless number these days, though I wouldn't be surprised to learn that Intel uses it to calculate TDP (which is equally useless). Because it's not guaranteed to run at that speed for any length of time, depending on your workload, cooling, VRMs, etc.
|
# ? May 21, 2018 19:24 |
|
It's reasonable to put less weight in that number if it never clocks down that far under real workloads in normal conditions. Nobody buys an 8700/8700K expecting 3.2/3.7 GHz out of the box.
|
# ? May 21, 2018 19:34 |
|
PerrineClostermann posted:Because it's not guaranteed to run at that speed for any length of time, depending on your workload, cooling, VRMs, etc. That may be a concern for mobile CPUs but I don't think there's a single retail Z370/H390 board out there that can't sustain all-core turbo on a i7-8700. IIRC there was a OEM machine on computerbase that couldn't but that was because the BIOS had a hardcoded 65W power limit and the machine didn't have a single case fan. The reality is that retail Z370 boards launched with MCE enabled (= single core turbo on all cores).
|
# ? May 21, 2018 19:49 |
|
eames posted:That may be a concern for mobile CPUs but I don't think there's a single retail Z370/H390 board out there that can't sustain all-core turbo on a i7-8700. IIRC there was a OEM machine on computerbase that couldn't but that was because the BIOS had a hardcoded 65W power limit and the machine didn't have a single case fan. The reality is that retail Z370 boards launched with MCE enabled (= single core turbo on all cores). In der8auer's testing the lowest end Z370 boards could only sustain ~120W without airflow over the VRM's. My i7-8700K is a pretty bad overclocker and easily draws 150W under some heavy workloads, and that's with manual voltage settings. With stock voltages it would draw more, and I'm pretty sure I recall it drawing 170W in P95 at 4.7GHz with stock voltage. You can get away with MCE enabled by default anyway because even people who buy i7's don't run Blender CPU renders all day long, but low end Z370 boards really do have VRM's that are dimensioned for stock TDP on an i7, not all-core max turbo. Especially if you plan on running heavy workloads for extended periods of time - the caps will burn out eventually. TheFluff fucked around with this message at 23:02 on May 21, 2018 |
# ? May 21, 2018 22:58 |
|
Methylethylaldehyde posted:Having a branch predictor at all was a huge rear end improvement in the early days. Now we're getting the last few % out of whats left over after 15 or 20 years of improvements. movax posted:I had a really good relationship with my SW team at my last job, which was at the level where we would sit down and Id make a hardware design decision that was nowhere near optimal from my point of view understanding that as a result the SW guys would remain oversubscribed at 2x instead of 3x. Easy for me to live with knowing its helping my buddies out (while still getting the job done) and also knowing that I have some goodwill to burn (which I did) when I needed their help to fix my fuckups because hardware spins take time and money! Also, 4th Spectre variant identified. Good times.
|
# ? May 21, 2018 23:23 |
|
JawnV6 posted:Again, I'm finding this really handwavey and not really true to the x86 history as it existed. A "branch predictor" like we've been discussing on a 486 is wholly unnecessary. 100% static prediction. By the time of a Pentium MMX you need something dynamic or the bubbles will kill perf. It's ahistorical to say it was a huge improvement because there's no predictor-less model that gets some huge upgrade with only a predictor added. There's a 5-stage in-order core with a static predictor, then a 6-stage superscalar with a dynamic BTB. I was speaking more in the realm of 'microprocessors, generalized', where we went from 'holy poo poo photolithography is amazing' in the late 60s, with huge process, compiler and architecture improvements every year like clockwork, but now we're running into issues that are Not Fun(tm) to solve. Like EUV lithography and how basically nothing likes being bombarded with ionizing photons, and how any problem outside of embarrassingly parallelizable ones run into issues where even with infinite cores, it's fundamentally limited by the single threaded performance.
|
# ? May 22, 2018 00:09 |
Goddamndels law
|
|
# ? May 22, 2018 03:19 |
|
TheFluff posted:My i7-8700K is a pretty bad overclocker and easily draws 150W under some heavy workloads, and that's with manual voltage settings. With stock voltages it would draw more, and I'm pretty sure I recall it drawing 170W in P95 at 4.7GHz with stock voltage. Mine has the worst TIM under heatspreader known to man: 82C on a Cryorig H7 running P95 at -0.08 Vcore, 30C ambient. I always LOL everytime I check out old CPUs on Ebay: "Ohhh, an i5-3470 refurb at $90? i5-4590 and 12GB DDR3 refurb at $170? Too expensive, I just buy the super outdated Q9650 at $60 to use on a super beaten 10 year old mobo."
|
# ? May 22, 2018 04:08 |
|
Screw buying a $60 chip to use in LGA775, you can get an LGA1156 motherboard and X3440 both for that price. Or, if your LGA775 motherboard can OC, get an X3350 for $20-25 and you can easily run it as fast as the Q9650. Eletriarnation fucked around with this message at 17:30 on May 22, 2018 |
# ? May 22, 2018 17:26 |
|
movax posted:Aren’t the Tensilica Xtensas VLIW? Or is more that they are super customizable to add on instructions and people often add VLIW type stuff. Until just now, I wasn't even aware Cadence had designed their own processor. After a cursory glance, it seems that it's a RISC CPU, with (optionally) a DSP glued to it - the DSP is obviously VLIW. What I was alluding to are the famously-terrible Itanium arch, as well as older GPU architectures - they'd all attempted VLIW (with various degrees of success), but eventually gave up.
|
# ? May 22, 2018 17:45 |
|
lDDQD posted:Until just now, I wasn't even aware Cadence had designed their own processor. After a cursory glance, it seems that it's a RISC CPU, with (optionally) a DSP glued to it - the DSP is obviously VLIW. To be nitpicky, Cadence didn’t design it, they bought it. Tensilica had been around quite a while pre-Cadence. There’s all kinds of oddball deeply embedded CPU cores out there that you never really hear much about. ARC, for example (the Synopsys equivalent to Tensilica, and also an acquisition of a smaller company rather than an in house design). ARC is everywhere, actually. You might have an ARC core in the computer you’re using to read this! One of their bigger design wins was to be the CPU core in the Management Engine in several generations of Intel CPUs.
|
# ? May 23, 2018 03:11 |
|
BobHoward posted:To be nitpicky, Cadence didn’t design it, they bought it. Tensilica had been around quite a while pre-Cadence. ARC can likely trace some of its lineage back to the SuperFX in some old SNES cartridges. Argonaut RISC Machine, from Argonaut Games. They did the SuperFX and a few other chips for Nintendo. Back in the wild 1990s where spinning ASICs for specific video games was totally fine and normal. Silicon Valley fuckin’ sucks now.
|
# ? May 23, 2018 07:41 |
|
poo poo was crazy. I have an old Snap-On automotive scanner that came out in the early ‘90’s and had replaceable cartridges that you updated every few years as cars and systems became more complex. The big selling point was that Snap-On guaranteed them to be upgradable for a long time; everything up to that point became obsolete very quickly. The early cartridges were just memory with a few support chips, but by the late 90’s they had a microprocessor inside to deal with the OBDII protocol, and the last few years they were available(2004-ish), they had an ASIC, microprocessor and memory shoved into a little cartridge the size of an Intellivision cartridge. A $1500 Intellivision cartridge.
|
# ? May 23, 2018 14:15 |
|
https://twitter.com/InstLatX64/status/969560033922035713 jesus christ intel, how is anyone supposed to target this mess
|
# ? May 24, 2018 20:11 |
|
Well the entire Knights group you probably either ignored completely or were targeting specifically
|
# ? May 24, 2018 20:17 |
|
repiv posted:https://twitter.com/InstLatX64/status/969560033922035713
|
# ? May 24, 2018 20:41 |
|
Has Intel over-segmented its' technology into a confusing nightmare? AVX is a powerful tool, no? Does AMD or ARM offer simpler or easier ways to target performance advantages? Also, maybe I don't understand what this means.
|
# ? May 25, 2018 10:30 |
|
repiv posted:https://twitter.com/InstLatX64/status/969560033922035713 CUDA / OpenCL
|
# ? May 25, 2018 11:34 |
|
repiv posted:https://twitter.com/InstLatX64/status/969560033922035713
|
# ? May 25, 2018 11:45 |
|
Coffee Lake seems to have received a pretty big boost in Premiere thanks to a patch that utilizes the iGPU for rendering/encoding. It went from being the slowest of the bunch (vs 7980XE/TR 1950X/R7 2700X) to performing similar to the two HEDT platforms when rendering a single video. https://www.gamersnexus.net/guides/3310-adobe-premiere-benchmarks-rendering-8700k-gpu-vs-ryzen I wonder about quality difference and (QuickSync tends to be worse) HVEC/h265 compatibility but it looks good.
|
# ? May 25, 2018 12:06 |
|
If you're in a position to be targeting these specific branches of the AVX instructions, then odds are you using some specialized software that you own on top of large, consistent hardware platforms. Figure out what instructions your specific hardware supports, compile to target that, run it until the next hardware refresh. For standard commercial software, they're just going to target those inner three circles.
|
# ? May 25, 2018 16:25 |
|
Literally no one bought Mill.
|
# ? May 25, 2018 16:45 |
|
BangersInMyKnickers posted:If you're in a position to be targeting these specific branches of the AVX instructions, then odds are you using some specialized software that you own on top of large, consistent hardware platforms. Figure out what instructions your specific hardware supports, compile to target that, run it until the next hardware refresh. For standard commercial software, they're just going to target those inner three circles. Assuming any compilers other than icc even support all these avx variants... If you're targeting one of the more esoteric ones you might have to drop to assembly.
|
# ? May 25, 2018 16:49 |
|
That's going to be a problem for the people running high-performance computer clusters to sort out and luckily those people are turbo-nerds anyway.
|
# ? May 25, 2018 16:56 |
|
VostokProgram posted:Assuming any compilers other than icc even support all these avx variants... If you're targeting one of the more esoteric ones you might have to drop to assembly.
|
# ? May 25, 2018 23:41 |
|
In HPC, you tend to do that a lot anyway.
|
# ? May 26, 2018 00:45 |
|
The AVX feature difference between Cascade Lake and Cannon Lake. Eh.
|
# ? May 26, 2018 13:52 |
|
i7 8086k will $486 and turbo boost to 5Ghz
|
# ? May 30, 2018 15:38 |
|
|
# ? Apr 23, 2024 22:22 |
|
Jesus christ Intel.
|
# ? May 30, 2018 15:45 |