rrss

55 ms·

541.
by rrss 7y ago
Up over 130% in the last 12 months, but up ~20% in the last 13-14 since AMD took a nosedive in October 2018 with the rest of the semiconductor stocks.
542.
by rrss 7y ago
GP100 and GP102 are different sizes. They are certainly different silicon - the die area can't be changed in firmware or drivers. From techpowerup: GP100: 610 mm^2 GP102: 471 mm^2
543.
by rrss 7y ago
The server class gpus are generally different silicon than the consumer gpus. That's a technical difference.
544.
by rrss 7y ago
What's drastically different between Radeon VII and 2080 Ti? They have basically the same peak fp32 tflops.
545.
by rrss 7y ago
Gyroscopes drift. Every system using gyroscopes that needs precision has some external tracking system as another reference (e.g. GPS, star tracking). In this case, that external system is a camera on top of the tower.
546.
by rrss 7y ago
> what they have is machine vision as a novel input into a conventional control loop Control loops are everywhere, but CD-ROM drives do not use machine vision. That's why op said that's the novel part.
547.
by rrss 7y ago
How do you align an array of hundreds of mirrors to a single spot with four photodiodes?
548.
by rrss 7y ago
Genuine question: how would open sourcing help with that compatibility issue? Open source software is nice, but doesn't magically make all compatibility problems go away.
549.
by rrss 7y ago
> it does work at about 90% the speed of an NVIDIA RTX 2080 Ti, but is much cheaper, which can be a good trade-off. Is this 90% of fp32 performance, of 90% of fp16 performance?
550.
by rrss 7y ago
> But the positive news is that v1 & v2 only matter at all if you do in-process sandboxing of untrusted code. I don't think this is accurate. It seems to be a widespread misunderstanding that started because the original proof o
551.
by rrss 7y ago
Who actually supports opencl 2.0? ROCM, which AFAICT is AMD's new preferred platform, only supports opencl 1.2 also.
552.
by rrss 7y ago
FYI the comparison is to the A72, not A73. A72 is a mobile chip, so still not particularly relevant to tuna-fish's claim about high end servers. The raspberry pi 4 is not a high end server. > half the die area Much of this is free w
553.
by rrss 7y ago
I think you have an extra factor of two there. If you need 20 MHz without aliasing, you either need real sampling at 40 MS/s or IQ sampling at 20 MS/s, but IQ sampling at 40 MS/s is overkill. Nyquist works differently if your
554.
by rrss 7y ago
Given the impressive inference results shown on this page, I find it interesting that graphcore did not participate in mlperf inference v0.5.
555.
by rrss 7y ago
"IPU" was made up by graphcore to describe their processor, so you call choose to use the term for as many or as few processors as you want. IMO, any processor can be an "intelligence" processor.
556.
by rrss 7y ago
AMD processors were also vulnerable to spectre v2. I don't know the status of the mitigations or whether it was fixed in zen 2. EDIT: I found the list I made a few months back. no guarantees, but i think it is mostly accurate. Meltdown
557.
by rrss 7y ago
> going all in with lying in benchmarks against AMD I'm assuming this is referencing the "intentionally misleading benchmarks" piece in servethehome. It's worth reading the follow up, in which the author discovers the
558.
by rrss 7y ago
Transistors haven't really been getting smaller for the last few nodes, they've mostly just been getting denser. AFAIK nobody currently knows how make finfets much smaller without making them awful.
559.
by rrss 7y ago
Gotcha, thanks.
560.
MLPerf Inference v0.5 Results(mlperf.org)
1 points by rrss 7y ago|0 comments
561.
Graphics Chip Chronicles – Nvidia's RIVA 128(electronicdesign.com)
1 points by rrss 7y ago|0 comments
562.
by rrss 7y ago
spleeter is using neural networks.
563.
by rrss 7y ago
So, based on the response to Spectre and friends ("intel knowingly sacrificed security in the pursuit of performance and everyone should sue them") [0][1], is the correct response here "browser vendors knowingly sacrificed se
564.
by rrss 7y ago
7nm instead of 16nm probably helps a lot.
565.
by rrss 7y ago
The vast majority of power lines in France are overhead. RTE overhead lines: 100000 km RTE underground lines: 5400 km https://www.rte-france.com/fr/document/statistiques-reseaux-... If you are intending to disting
566.
by rrss 7y ago
It is pretty common for cell towers to have backup power. Not all of them do, but many.
567.
by rrss 7y ago
That $0.02 value is entirely bogus. It is based in the published 30 "teraops" for Xavier, which it uses as teraflops and assumes is comparable to the other flops values. ($699/30 "TFLOPS" = $0.02 $/"GFLOP&
568.
by rrss 7y ago
SRAM cell area is a very common benchmark for processes.
569.
by rrss 7y ago
That sentence explains perfectly well what it means by democratizing, and how is independent of the platform being tied to nvidia.
570.
by rrss 7y ago
I've heard rumors that 10nm was late and management at Intel put a lot of pressure on the process teams, which led to the "old guard" quitting and the younger generation left fumbling with unusable yields. But they're on

More ›