#
✕ nick:delsol
#asahi · 2023-11-07 17:24
delsolahhh
#asahi · 2023-11-07 17:24
delsolin ESXi or Proxmox?
#asahi · 2023-11-07 17:24
delsolI'm basically backwards in this channel..... did most of my stuff on mac, later OSX.... and did a shitload of hackintoshing for a while....
#asahi · 2023-11-07 17:23
delsolso what was your era of experimenting with shit to see what you could accomplish?
#asahi · 2023-11-07 17:20
delsolactually worked on one with a buddy in college. We only ever got 4 CPU's operational with the adapters....
#asahi · 2023-11-07 17:19
delsoland six P2's, P3's or celerons.....
#asahi · 2023-11-07 17:19
delsolwith 6 socket 370 to socket 8 adapters.....
#asahi · 2023-11-07 17:19
delsolkidplayer666: no, a fun project was an ALR Revolution 6x6... with 6 pentium pro's (3 socket 8's per CPU daughterboard...)
#asahi · 2023-11-07 17:18
delsolyou can take a Geforce 4090 back to 1997, plug it into a PCI to PCIe bridge chip and put it on your overclocked dual socket celeron machine... and it still aint gonna do shit.
#asahi · 2023-11-07 17:17
delsolydalton: still need drivers.
#asahi · 2023-11-07 17:17
delsolnot really, that would take EPIC/VLIW style architecture.
#asahi · 2023-11-07 17:16
delsolseveral of my hackintosh's had G200e's
#asahi · 2023-11-07 17:16
delsolhey, don't fuck with the G200e, G200e is a great GPU for all sorts of shit.
#asahi · 2023-11-07 17:16
delsolbasically an emulated Matrox G200e
#asahi · 2023-11-07 17:15
delsolbecause you're emulating a GPU with CPU....
#asahi · 2023-11-07 17:15
delsolydalton: windows GPU drivers to run as a VM on M2-Ultra even.... not so great.
#asahi · 2023-11-07 17:15
delsolit already is, see also Nvidia Grace/Grace-Hopper
#asahi · 2023-11-07 17:14
delsol(and potentially license a bunch of the rest of the SoC stuff from mediatek)
#asahi · 2023-11-07 17:14
delsolkidplayer666: I don't think so. I think its the flipside.... NVIDIA is going to just build their own arm shit with the own GPU.
#asahi · 2023-11-07 17:13
delsollook at benchmarks of random tasks..... M2 at 3.4ghz was competing with intel P cores at 5.5ghz.
#asahi · 2023-11-07 17:12
delsolbut look at what it was doing.....
#asahi · 2023-11-07 17:12
delsolnon-GPU stuff appears to blow the pants off of windows on x86 laptops....
#asahi · 2023-11-07 17:10
delsolkidplayer666: you see people that have run the dev builds of windows for ARM on M1/M2?
#asahi · 2023-11-07 17:08
delsoltake a goood look at the clock speeds. ;)
#asahi · 2023-11-07 17:08
delsolBut look at the 7000 series....
#asahi · 2023-11-07 17:08
delsolsingle threaded performance increases of higher clocks made up the difference....
#asahi · 2023-11-07 17:08
delsolYes, on non-bandwidth limited workloads.
#asahi · 2023-11-07 17:07
delsol(memory bandwidth and cache is also a large consumer of power.... and for a laptop chip, it makes a reasonable difference in potential battery life)
#asahi · 2023-11-07 17:05
delsolSee also: no consumer 5000 series threadripper..... because at the pricepoint the Threadripper Pro was plenty beefy to handle most of the single-workload setups as fast if not faster than a Dual EPYC....
#asahi · 2023-11-07 17:04
delsolbecause the big xeons don't clock for crap, chug power, and even four times as many cores doesn't make a difference if 10-15% of your workload execution time is single threaded and your single threaded performance is 20% lower.....
#asahi · 2023-11-07 17:03
delsoland at the clockspeeds apple is now pushing on M3 cores... if the cores are as powerful as expected, plenty of "big xeon" workloads would likely be faster on an M3-Max than the xeon.
#asahi · 2023-11-07 17:02
delsolkidplayer666: Amdahl's law basically says that those workloads are few and far between.
#asahi · 2023-11-07 17:01
delsolsingle threaded blows chunks, but if you are doing a metric fuckload of separate VM's that don't need single threaded..... throughput is huge for the price.
#asahi · 2023-11-07 17:01
delsolSee also Ampere Altra benchmarks....
#asahi · 2023-11-07 17:01
delsollol, no
#asahi · 2023-11-07 17:00
delsolbut if linux was better supported on it, plenty of big linux server workloads would run fucking amazeballs on it, and justify the 10k/15k/20k pricetag such a machine would demand.
#asahi · 2023-11-07 17:00
delsolin all seriousness, most apple workloads wouldn't benefit much from an -Extreme 4x die setup.....
#asahi · 2023-11-07 16:59
delsolyou want an M3-ShortIntelStockToday.... 4 M3-Max dies bolted to an IOdie with another couple memory channels for those huge memory footprint workloads, along with a pile of PCIe lanes....
#asahi · 2023-11-07 16:58
delsolChaosPrincess: then you don't want an M3Pro or M3Max...
#asahi · 2023-11-07 16:58
delsoland how much die space do you want to spend on it....
#asahi · 2023-11-07 16:57
delsolChaosPrincess: other than how many workloads are going to be actually more performant on 8p4e vs 6p6e?
#asahi · 2023-11-07 16:56
delsolkidplayer666: thats the other gotcha.... the previous setup GPU wasn't able to use all the memory bandwidth either.
#asahi · 2023-11-07 16:55
delsolkidplayer666: yes and no.
#asahi · 2023-11-07 16:55
delsoland whats worse, is to get the 250GB/s from memory to cores on sapphire rapids... you need 32 or more cores.
#asahi · 2023-11-07 16:54
delsolSapphire Rapids with its 8 channel DDR5 controller... and socket TDP of 380w.... only gets real world 250GB/s from memory to cores.
#asahi · 2023-11-07 16:54
delsolM2Max may have 400GB/s of theoretical memory bandwidth... but real world from memory to cores.... is only about 225GB/s
#asahi · 2023-11-07 16:53
delsoland the people bitching about memory bandwidth are fools.
#asahi · 2023-11-07 16:53
delsolthats more than 50% of the M2Max dies you are cutting 40% off of...
#asahi · 2023-11-07 16:53
delsolif you need to sell 100k M2, 20k M2Pro and 15k M2Max.....
#asahi · 2023-11-07 16:52
delsolbefore EVERY single M1Pro/M2Pro that they shipped... was a crippled M1Max/M2Max that they cut off half of the GPU.

plus ›