sliken

7 ms·

61.
by sliken 1y ago
Apparently the frameworks desktop's 5g bit network isn't fast enough to scale well with LLM inference workloads, even for a modest GPU. Anyone know what kind of network is required to scale well for a single modest GPU?
62.
by sliken 1y ago
Indeed. Pretty much all silicon today comes on 12" or so wafers, broken into chip sized pieces, and each chip is tested and the ones that failed are thrown away. Cerebras uses the entire 12" and builds in redundancy so that with
63.
by sliken 1y ago
Or comcast's 1 gb service with a 1TB limit, which means you hit the limit in 140 minutes at full rate.
64.
by sliken 1y ago
I was looking at $200 cameras that were PoE, designed for outdoors, and had an IR emitter to protect my car that someone tried to steal from my driveway. Ended up with a pair of reolinks for $120, the spec was a bit less, but I got two came
65.
by sliken 1y ago
When I looked they had premium pricing, gen old hardware, and dependencies on either cloud with subscription or proprietary hardware. Same with their doorbell cam, proprietary hardware or cloud subscription. I bought 2 reolink cameras for t
66.
by sliken 1y ago
Sounds like my macbook 16" with the Intel i9. Just about anything, full screen video call, backups, patching, etc it sounds like a hair dryer. I'm not surprised there's various docks, stands, etc that include supplemental
67.
by sliken 1y ago
Ah, yes, vector general. I think that was a bit earlier, but much slower. I think most of the footage was rendered a frame a time then a camera took a photo.
68.
by sliken 1y ago
I did use the screen a bit, it was connected to the unibus interface (which provides memory access) to read vector data straight out of the PDP-11's memory. One oddity of the vector screen is that if you drew long diagonal lines from
69.
by sliken 1y ago
One problem with minatures, positioning the camera for each photo, then turning the result into special effects shot for a movie is that it can take days to record a scene, get the film developed, and view the result. Only when played bac
70.
by sliken 1y ago
There do exist serial hybrids where the engine produces electricity while maintaining the ideal RPM, and then is connected to a battery+electric motors. However in practice the vast majority of hybrids do not use this approach and have moto
71.
by sliken 1y ago
Not sure I'd call it safe to touch. Getting with 5cm for an hour gives you as much radiation as a 8 hour flight. I wouldn't want touch it, make jewelry from it, or any substantial near promity. Not to mention if it was "on
72.
by sliken 1y ago
Not really, the max acceleration/gravity you can get from earth is around 1G. Even if you tunneled to the center of the earth, you'd get ... zero g. However with an earth max black hole you could get closer, which would have a hi
73.
by sliken 1y ago
Sort of. Ceres can't turn mass into energy at a ratio that makes fusion and fusion look pretty lame. On the other hand taking 0.00016 of earths mass, turning it into a blackhole, and shooting it through the earth isn't likely to
74.
by sliken 1y ago
Obviously we've managed to orbit the moon. To find the blackhole that replaced the moon just enter lunar orbit and keep firing your thrusters to slow down.
75.
by sliken 1y ago
Well trick is optane was terrible as ram and pricey as storage. From what I can tell it was targetted: * where persistence was important, better than ram * for cases where more ram didn't help * for cases where adding multiple
76.
by sliken 1y ago
I've had 3 serious failures. 1. Santa Cruz aluminum frame, snapped a few inches above the bottom bracket, at a weld line for a reinforcement. I was sprinting and heard it, frame didn't fail/fold but felt weird, didn&#x2
77.
by sliken 2y ago
I'd agree with you, if the goal was to decrease the deficit. But so far it sounds like the main goal is decreasing taxes on the rich and corps.
78.
by sliken 2y ago
Beware, once you get used to watching blurays streaming video looks like crap.
79.
by sliken 2y ago
One of the reasons I ditched plex was that it didn't allow local playback without a network connection. Jellyfin+roku works great for me.
80.
by sliken 2y ago
As plex got more annoying to use I switched to Jellyfin on my Roku. It's been great. Reading the threads it does seem like there's some issues on the Apple TV, but on the Roku it's great.
81.
by sliken 2y ago
Plex got increasingly annoying to use, I switched to jellyfin, which I mostly use from my Roku, and it's great.
82.
by sliken 2y ago
Indeed, 23% more money for double the memory bandwidth.
83.
by sliken 2y ago
Ah, thanks, I didn't know that.
84.
by sliken 2y ago
Yes, you can spend $12k on a mac studio. However the Mac studio with the m4 max has 128GB ram, double the memory bandwidth, and costs $3,699. LESS than the DGX spark. Granted it doesn't have 200gbe. If it's the memory bandwid
85.
by sliken 2y ago
1000 TOPS sounds good, until you notice the 273GB/sec. This will allow larger models than consumer GPUs, but be rather slow running them.
86.
by sliken 2y ago
Sort of. Sure you can game on the framework desktop, but I'd expect them to do similar on inference workloads and come with the same max ram and bandwidth.
87.
by sliken 2y ago
Depends which one, the HP ZBook Ultra uses LPDDR5x-8533, a dead match for what nvidia claims (273GB/sec). Although the DGX spark now costs $4,000 instead of the "starting at $2,999" mentioned a few months ago. So the bandwid
88.
by sliken 2y ago
Sort of. They come in different speeds, and those prices are for the 400gbe version. The 200gbe, like in the GX10 and spark are more like $1250. Not to mention you have to need that the 200gbe (to cluster 2 of them) and I'd expect t
89.
by sliken 2y ago
Cheaper? The DGX spark is $3,000 with 128GB ram. A framework desktop with the AMD strix halo 395 with 128gb ram is $2,000 and has better memory bandwidth. No price I can see for the identically spec'd (and nearly identical looking)
90.
by sliken 2y ago
273GB/sec with good FP4 performance should be fine for developers playing with inference. This isn't the kind of thing that you'd use for inference workloads supporting millions of users. I'd like to see a inference ben

More ›