Thursday, July 30
Daily News Stuff 29 July 2020
Turtles, Termites, And Traffic Jams Edition
Turtles, Termites, And Traffic Jams Edition
Tech News
- According to Wikipedia, all DDR5 memory modules are registered.
I took a quick look at the docs released by Micron, and they only discuss registered modules, but don't explicitly say that unbuffered modules don't / won't exist.
If they are all registered - and we already know that DDR5 mandates on-die ECC - then there's little difference between desktop RAM and server RAM. Just full-channel ECC in case of bus errors, and LR (load reduced) modules for maximum die stacking.
I'm not certain if DDR5 being registered will work exactly as it does with DDR4, but if it does then the next generation of desktop CPUs will support at least twice as much memory, without needing to wait years for denser DRAM.
Which means that the next-plus-one generation of even entry-level servers will likely support up to 32 cores and 256GB of RAM.
- Zen 3 is due this year. (AnandTech)
Speak the name three times and it will appear.
This means both server and desktop parts. Also next-generation Navi cards for both gaming and datacenter use. Oh, and Xbox Series X and Playstation 5. It will be a busy few months for AMD.
- A passively-cooled 10Gb home server/router from Supermicro for only $1500. (AnandTech)
It's no Cobalt Qube.
- Two of Intel's next-generation Ice Lake Xeons beat a single current-generation Epyc. (Tom's Hardware)
By 7%. If you use AVX-512, which Epyc doesn't have. Will wait for more benchmarks, because Ice Lake should be a significant upgrade, if and when it ever arrives.
- Someone needs to explain to physicists that humans don't live in trees. (Vice)
- A lot has been written about the technological singularity. My view is that (a) it's impossible (at least the way it is depicted in science fiction) and (b) as far as it is possible, it is happening right now which is part of the reason everything is so fucking weird.
I'm not the only one to make that observation. (Less Wrong)
Also of interest is this article about the potential speed of the singularity. (Sideways View)
The two things to expect when it really kicks into gear, is rapid shifting in the labour market leading to high unemployment, and at the same time, global GDP growth in the mid double digits.
The key point of the notiong of the Singularity, though, is that at some point the trends go literally vertical, and that prediction of what things will be like on the other side of that point on the graph are impossible. That is exactly what I predict will not happen, though if I'm wrong no-one will ever know.
Disclaimer: What's that? There's an alternate-universe version of Bakarina where she's already into the main plot before she bonks her head? Well, see you tomorrow, then!
Update: Although she's 15 years old rather than 8 this time, she's still the same lovable idiot.
Update: Although she's 15 years old rather than 8 this time, she's still the same lovable idiot.
Posted by: Pixy Misa at
12:34 AM
| Comments (12)
| Add Comment
| Trackbacks (Suck)
Post contains 490 words, total size 5 kb.
1
"If they are all registered - and we already know that DDR5 mandates
on-die ECC - then there's little difference between desktop RAM and
server RAM."
If anyone can find a way to maintain the distinction, you can trust Intel will do so.
If anyone can find a way to maintain the distinction, you can trust Intel will do so.
Posted by: Rick C at Thursday, July 30 2020 01:56 AM (Iwkd4)
2
Also, I see Wikipedia is panhandling again. Remember when they only did it once a year? I wonder what their overhead percentage is.
Posted by: Rick C at Thursday, July 30 2020 01:57 AM (Iwkd4)
3
The wiki says DDR5 DIMMs will support onboard VRMs, but every tech article I've seen says the VRMs *will* be on the DIMM, not that it *may* be. I guess we'll find out eventually.
Also, I had read that the on-die ECC, IIRC, is not the same as the ECC on current DIMMs--it's internal to the chips, again, IIRC. But then again, https://www.rambus.com/blogs/get-ready-for-ddr5-dimm-chipsets/ from Rambus says "With DDR5, each DIMM will have two channels. Each of these channels will be 40-bits wide: 32 data bits with eight ECC bits. ".
Also, I had read that the on-die ECC, IIRC, is not the same as the ECC on current DIMMs--it's internal to the chips, again, IIRC. But then again, https://www.rambus.com/blogs/get-ready-for-ddr5-dimm-chipsets/ from Rambus says "With DDR5, each DIMM will have two channels. Each of these channels will be 40-bits wide: 32 data bits with eight ECC bits. ".
Posted by: Rick C at Thursday, July 30 2020 02:04 AM (Iwkd4)
4
This doesn't sound promising for "all ECC all the time": "Third, there are two independent channels per DIMM. While earlier SDRAM
generations had one CA bus controlling 64 or 72 (non-ECC/ECC) data
lines, each DDR5 DIMM has two CA buses controlling 32 or 40
(non-ECC/ECC) data lines each, for a total of 64 or 80 data lines."
Posted by: Rick C at Thursday, July 30 2020 02:07 AM (Iwkd4)
5
That Vice article is stupid, but you could've predicted that from it being on Vice. First off, it makes the assumption that deforestation will happen everywhere, but there's decades of reversal of that trend in the US and, I assume, in Europe. To claim net deforestation happens in the US you have to move the goalposts by talking about "virgin forests" and discounting the massive amount of planting going on.
Posted by: Rick C at Thursday, July 30 2020 02:12 AM (Iwkd4)
6
https://psmag.com/environment/the-planet-now-has-more-trees-than-it-did-35-years-ago
Posted by: Jay at Thursday, July 30 2020 02:39 AM (0jVI9)
7
Jay--like I said, the "rebuttal" is that the new growth is somehow deficient because it's not old-growth (and, of course, notice that South America is down, and the rain forests are more important than any other kind).
Also, I didn't mention this in my prior comment, but in the US, tree acreage has basically been stable since about 1910 (and is actually a bit up since then, something like 740M acres now vs 730M then). In other words about the time (very roughly speaking) we realized we were cutting down too many forests we took steps to start replanting.)
Also, I didn't mention this in my prior comment, but in the US, tree acreage has basically been stable since about 1910 (and is actually a bit up since then, something like 740M acres now vs 730M then). In other words about the time (very roughly speaking) we realized we were cutting down too many forests we took steps to start replanting.)
Posted by: Rick C at Thursday, July 30 2020 07:28 AM (Iwkd4)
8
Rick C - DDR5 won't be all ECC all the time, but it will at least be some ECC all the time. On-die ECC is mandatory, but module-level ECC is optional.
So for big servers with terabytes of RAM, you'll definitely want module-level ECC, and load-reduced modules, but for the average small server - anything up to 32 cores - desktop RAM will be a viable option.
So for big servers with terabytes of RAM, you'll definitely want module-level ECC, and load-reduced modules, but for the average small server - anything up to 32 cores - desktop RAM will be a viable option.
Posted by: Pixy Misa at Thursday, July 30 2020 09:37 AM (PiXy!)
9
The "new-growth" vs "old-growth" agrument is silly. Old growth is furniture wood: very valuable, and scarce (and it always has been, because of wildfires which are perfectly natural). The stuff that goes into paper and other cheap sources is always pulp wood, which can't be old growth because that stuff makes terrible paper. There was a lot of old timber wood that was lost in the US from the early 1800s on, but that stuff is still harvestable on a 60 year timescale. Pulp wood is more like 20 year, and it's always been farmed because of what it is. I mean, the English have been coppicing and pollarding for a few decades, at least, like that's new technology.
The South American jungle is a whole dif'rent story: that stuff grows back at fantastic rates. And the deadfall and other matter on the jungle/forest floor mostly makes up for any purported oxygen production. (hint: algae is pretty much the end-all, beat-all for oxygen production, just because of its lifecycle and proliferation)
Posted by: normal at Thursday, July 30 2020 11:06 AM (obo9H)
10
"The "new-growth" vs "old-growth" agrument is silly. "
Of course it is, and it's telling that the watermelons have to fall back to saying that when it's shown there's more trees now.
Of course it is, and it's telling that the watermelons have to fall back to saying that when it's shown there's more trees now.
Posted by: Rick C at Thursday, July 30 2020 11:16 AM (Iwkd4)
11
Pixy: yeah. I'm not sure how to articulate it but my understanding is that the per-die ECC isn't really exposed to the computer as a whole the way it is in regular ECC memory; it's there for the internal use of the chip. But it's possible I misunderstood an article.
Posted by: Rick C at Thursday, July 30 2020 11:18 AM (Iwkd4)
12
I believe that's correct. It's great, but not a substitute for full module-level ECC in serious server applications.
Posted by: Pixy Misa at Thursday, July 30 2020 11:44 AM (PiXy!)
58kb generated in CPU 0.0162, elapsed 0.3319 seconds.
58 queries taking 0.3204 seconds, 359 records returned.
Powered by Minx 1.1.6c-pink.
58 queries taking 0.3204 seconds, 359 records returned.
Powered by Minx 1.1.6c-pink.
