sliken

9 ms·

121.
by sliken 2y ago
A local NAS makes sense for some use cases. Privacy, local control, and network bandwidth adds up to having a NAS for some use cases. I see similar coming for AI. Tagging local photos, reading/summarizing private documents, helping y
122.
by sliken 2y ago
I've seen similar claims, with nothing in particular to support it. Hopefully Nvidia clarifies this soon. AMD Strix Halo has a 128GB config with 256GB/sec. Apple has a MBP m4 max with 128GB ram for $4,700 with a 546GB/sec in
123.
by sliken 2y ago
The Nvidia Project Digits looked rather interesting. Much like a Apple studio ultra. 2 dies, "high" bandwidth unified memory, 20 cores, tiny mac mini sized case, and 2 SFP+ for 2 connectX ports (unsure if that's IB or ether
124.
by sliken 2y ago
As several other threads of mentioned, Nvidia has statements in various places along the lines of "All Nvidia Digit systems will have 128GB". Looks like you might be able to select different storage or something or maybe the numb
125.
by sliken 2y ago
Much of the last 5 years of AI improvements have been FP32 -> FP16 -> FP8 -> FP4, not any particularly impressive hardware or software improvements. Similarly on the GPU side much of the improvements is more fake pixels and more fa
126.
by sliken 2y ago
Ah, sure the Oryon cores are decent, GPU is a bit weak, even compared to the normal cheap/low end APUs.
127.
by sliken 2y ago
I think it's 1/40th of the GB200, but I think that's the one with two blackwells, so 1/20th of a full blackwell.
128.
by sliken 2y ago
This Nvidia is quite a bit like the AMD Strix Halo, which has the same 128GB max memory, x86, unified memory, and a 256 bit wide bus. I've seen much hopeful speculation that the nvidia is 512 bits wide, but I'm dubious.
129.
by sliken 2y ago
10 cortex-x925 and 10 cortex-a725 I believe.
130.
by sliken 2y ago
I'm not sure it has a video out. There is a AI produced image of the digit and a monitor though.
131.
by sliken 2y ago
Sure, but less than half the memory bandwidth and inference is largely bandwidth bound.
132.
by sliken 2y ago
Agreed. I care more about LLM size than tokens/sec so the GB10 or Strix Halo with 128GB are my leading choices. Both look to be cheaper than a similar mac studio with 128GB (minimum $4,800 currently). Will have to wait on final conf
133.
by sliken 2y ago
Use a filesystem that snapshots AND do a complete backup.
134.
by sliken 2y ago
Grace + Hopper, Grace + blackwell, and discussed GB10 are much like the currently shipping AMD MI300A. I do hope that a AMD Strix Halo ships with 2 LPCAMM2 slots for a total width of 256 bits.
135.
by sliken 2y ago
Not sure exactly, but they mentioned linking to together with ConnectX, which could be ethernet or IB. No idea on the speed though.
136.
by sliken 2y ago
Indeed, generational improvements are at an all time low. Most of the "revolutionary" AI and/or GPU improvements are less precision (fp32 -> fp16 -> fp8 -> fp4) or adding ever more fake pixels, fake frames, and now
137.
by sliken 2y ago
rk3588 is pretty close, I believe it's usable today, just missing a few corner cases with HDMI or some such. I believe that last patches are either pending or already applied to an RC.
138.
by sliken 2y ago
Keep in mind the "full" grace is a completely different beast with Neoverse cores. This new GB10 uses different cores and might well have a different memory interface. I believe the "120 GB" config includes ECC overhea
139.
by sliken 2y ago
Well presumably there's some different configurations supported, otherwise they would say $2,999 instead of starting at $2,999.
140.
by sliken 2y ago
Sure, but it's very limited. It doesn't detect or fix errors in the dimm (outside the chips), motherboard traces, CPU socket, or CPU.
141.
by sliken 2y ago
It's been far from clear what config options are going to be available, and the $3,000 price is the "starting at" price. Not sure what the options will be, but people have collectively found statements that imply all configs
142.
by sliken 2y ago
The cortex X925 is pretty competitive. Seems similarly aggressive to the Apple M4. Both have 10 wide decode and disbatch, pretty exceptional single thread, and good overall performance. Seems close enough that it might well come down to i
143.
by sliken 2y ago
Ah, looks like the 5070 = 4070 is not anywhere close to fair. It requires generating multiple frames per game engine frame and running the older DLSS on the 4070. The real world differences with the same settings looks to be MUCH slower.
144.
by sliken 2y ago
Hmm, seems pretty close. Not sure how the memory channels related to the performance. But the ampere board above has 8 64 bit channels @ 3200 MHz, the AMD Turins have 24 32 bit channels @ 6400 Mhz. So the AMD memory system is 50% wider,
145.
by sliken 2y ago
Seems like pretty heavily stretched truth. Looks like the actual performance uplift is more like 30%. The 5070=4090 comes from generating multiple fake frames per actual frame and using different versions of DLSS on the cards. Multiple f
146.
by sliken 2y ago
Double the bandwidth, double the ram, double the pins, and double the power isn't cheap. I wouldn't be surprised if the profit on the 4090 was less than the 4080, especially since any R&D costs will be spread over significant
147.
by sliken 2y ago
I believe the cards that are intended for compute instead of GPU default to ECC being on and report memory performance with the overheads included.
148.
by sliken 2y ago
I originally thought so, since the previous Grace CPUs used the neoverse N2, which loses to Apple's M4 cores. However apparently 10 of the cores are the Cortex-X925 CPUs, which are a serious upgrade. Basically 10 performance cores and
149.
by sliken 2y ago
TheRegister mentions: Specs we’ve seen suggest the GB10 features a 20-core Grace CPU and a GPU that packs manages a 40th the performance of the twin Blackwell GPUs used in Nvidia’s GB200 AI server.
150.
by sliken 2y ago
The CPU is apparently the result of a "secret" project that wasn't on published roadmaps. It's called the GB110. So maybe they will offer differently binned CPU/GPUs with a different fraction of cores disabled an

More ›