Friday, August 28
Daily News Stuff 27 August 2020
Where Oh Where Have My GAA-FETs Gone Edition
Where Oh Where Have My GAA-FETs Gone Edition
Tech News
- How much memory did home computers realistically have in 1983? The C64 came out in 1982, so that's one datapoint. The C128 came out in 1985.
But the Atari 1040ST came out in 1986. The Amiga 1000 in 1985 had 768K with the 256K expansion module. By the time I bought mine at the start of 1987 that shipped as standard but I don't know when that started.
If you look up the specs it will say that it came with 256K on board plus 256K on the expansion module, but that's not quite true. The OS ROMs weren't ready at launch, or indeed 18 months later, and the system came with an extra 256K of RAM instead so you could load what should have been there from a "Kickstart" disk. Installing the ROMs - kits became available in 1987 - turned that 256K into usable fast RAM.
I'm not sure if the A1000 ever shipped without that extra 256K of RAM, though the A500 shipped with 512K RAM and working ROMs.
So 128K in 1983 and 256K in 1984 is not entirely unreasonable. By mid-85 256K was something you could patch into hardware because the software wasn't ready.
Here's a handy collection of old Bytes on Archive.org.
I found the hard data I needed in - of all places - the New York Times. In June 1983 - dead centre on my imaginary time window - 256k DRAMs were sampling but not yet in production, and 64k chips were $4.50 each.
So 128k would run $72, and 256k $144. Plus another 25-50% if we're using 10-bit bytes, depending on whether we're using x1 or x4 chips. By comparison the C64 launched at $599 but the price came down pretty fast.
Since 256k chips were sampling in June '83, I can reasonably posit that the Imagine 1000 - which seems like as good a name as any for a fictitious 10-bit system - was designed from the start to be upgraded from 64k to 256k chips.
Using 3 4416 chips per bank to give 10 bits plus parity, 24 chips could provide the original system with 128k in total, while the Model B could come with half that number of chips but double the memory. And the Model A could be upgraded to a maximum of 512k with a simple chip swap.
That works. Why did no-one make this thing? Not that I had remotely that much money as a kid...
Update: In the March '83 issue of Byte there are adds for 64k memory upgrades (with parity) for $50, and the C64 could be found for $299 after rebate.
- TSMC's 3nm node will stick with FIN-FETs, with GAA pushed back to 2nm. (AnandTech)
I don't remember if TSMC had previously said that 3nm would be GAA, or just spoken of 3nm and GAA in the same breath and the rest was speculation.
Oh, and 2nm is coming.
- A sneak peek at Tiger Lake does show a single-threaded / IPC advantage over Zen 2. (Tom's Hardware)
Single-threaded performance is on the order of 5% better than a Ryzen 4300U, and since this particular Tiger ran at a lower clock speed, IPC is around 10% better. On multi-threaded workloads the 4-core/4-thread Ryzen 4300U was 45% faster than this 2-core/4-thread Tiger Lake chip, so AMD will retain a very comfortable multi-threaded lead even while we wait for Zen 3.
Intel is also expected to catch up on the graphics side with these new chips, but the leaked benchmarks don't say anything about that.
- Salesforce recorded a record quarter and immediately laid off 1000 staff. (Tech Crunch)
Uh. Timing is everything, guys.
- The share price of Fucking Elastic, makers of Fucking Elasticsearch, is at an all-time high. (WCCFTech)
No, I am not at all annoyed with the deliberately-introduced deficiencies of Fucking Elasticsearch. No, I didn't spend yesterday working around a broken-as-designed upgrade that wrecked application features that used to run just fine. Why do you ask?
- Milan will be up to 20% faster than Rome, sort of. (WCCFTech)
Rumours have been pretty consistent that Zen 3 has 10-15% better IPC than Zen 2, with the other 5-10% coming from minor clock speed improvements similar to the slight nudges we saw with the 3800XT.
It looks like L3 cache remains the same, but L2 cache has doubled, which would contribute to those IPC gains and help reduce contention on the new unified L3 cache.
Meanwhile Genoa is due by 2022 with "more than 64 cores", Zen 4, DDR5, and PCIe 5.0.
- Arwes makes your application look like it belongs in the 21st century. (Arwes.dev)
The 21st century as seen in the 1980s, anyway, which - looking at you, 2020 - was more optimistic than accurate.
- ArangoDB 3.7 supports JSON Schemas and other good stuff. (ZDNet)
It's a multi-model database - relational, object, document, graph, whatever - though currently pushing the graph features and its use case as a machine learning back-end, because that is the new hotness this month.
I've followed its progress but not actually used it, so can't say how well it really holds up.
Disclaimer: You know you've really fucked up when experienced engineers are thinking of replacing your product with MongoDB.
Posted by: Pixy Misa at
01:43 AM
| No Comments
| Add Comment
| Trackbacks (Suck)
Post contains 887 words, total size 7 kb.
53kb generated in CPU 0.1011, elapsed 0.1573 seconds.
56 queries taking 0.1474 seconds, 345 records returned.
Powered by Minx 1.1.6c-pink.
56 queries taking 0.1474 seconds, 345 records returned.
Powered by Minx 1.1.6c-pink.
