Intel Confirms Sapphire Rapids Uses Golden Cove
tomshardware.com
tomshardware.com
The should start fresh. The "Intel One"
Apple can do this because there’s literally only one processor, the M1.
Apple didn't want to ship anything with a Pentium inside.
Of course the fact that the codenames are used more often than the marketing names is itself a marketing failure.
You really need the architecture name to compare processors, and AMD currently has more usable architecture names; even if their processor number are still a disaster, zen 1 vs zen 3 is clear.
Name it something like "Intel Smart Core i3" with the introduction of little.BIG architecture. If those live up to the battery life hype, then the naming would even be justified.
The closest they get to “marketing” naming is the UPC-code-like numbers their processors get; and these, I believe, are optimized for the needs of enterprise buyers, who actually understand the ‘meaning’ of the parts of the number, since it’s pretty easy to see when you’re staring at it as one row among many of a spreadsheet of all the part-buy options they’re selling.
In this sense, they’re just retaining the very classical electronics parts-numbering tradition — see the naming of pretty much everything in this list (https://spectrum.ieee.org/tech-history/silicon-revolution/25...). They’re all just random codes, because parts buying on an industrial scale was always about vendors sending you these long sheets of different possible configurations each chip can be made in, and the buyer ticking off the ones they want as if they’re placing a dim-sum order.
I believe Intel still gets the majority of its business from enterprise buyers (OEMs, but these days also IaaSes doing their own integration), so this part-code is still the leading number for their products. Even though it’s mystifying to the average consumer.
Of course, they do try to hand their downstream integrators a bone by coming up with random marketing terms that apply to some of these processors but not others (i3/5/7/9, etc.) so that they can have something to badge a computer with. I do get the sense that this isn’t something Intel themselves think is very important, though; in about the same way that Nvidia thought “Max Q” was kind of silly (and first tried fixing that by enforcing a specific meaning to it, before just explicitly renouncing it, letting vendors use it however they like as pure marketing buzzspeak.)
The marketing names aren't much better. They sound like a frequent flier program (bronze, silver, gold, platinum) with seemingly meaningless numbers added. You'd think with so many generations and so many configurations, they could come up with a scheme that made sense. I'd name things like:
$GENERATION-$SOCKETS-$CORES-$FREQ
So for a 7th generation Xeon supporting 2 sockets worth of UPI, with 32 cores and a 2.9Ghz clock speed, you'd call it:
Xeon 7-2-32-29
Please Intel? Please?
Sounds too much like a bingo announcer to me...
Something I hope Pat Gelsinger will change it back to his Intel's days.
So I suppose some kind of opaque numbering, like "I7400" is shortest. I wish they adopted some position-coded scheme which would be easier to decipher, though. Something like architecture-revision-cores-sockets-cache-RAM-frequently-TDP kind, one letter each. Some Intel KCG2MDTM which would be hard to pronounce, but easy to look up, and easy to compare.
Not supporting ECC memory except in "enterprise" versions is pure segmentation. Good thing AMD broke this unwholesome tradition.
I think it started in the car industry.
I know it's not very insightful, don't hate me. Surely you can appreciate "iPhone 12" vs "iPhone DoDecaMax (Sapphire Rapids)+" or something like that. Just call it Intel 10nm 2021. How hard is that? Or Intel 10nm Q2. Just, anything.
From my understanding this is due to copyright/trademark.
Sapphire Rapids is not a product name, it is a code name for a family of processors, so the more appropriate comparison is "Sicilian with Firestorm and Icestorm cores" rather than "iPhone 12."
Yes, "xyz v2.1" is boring, but you know that it's the version after 2.0. That's more meaningful than "oh, 'purple tiger' is the new version, are you still using 'rainforest dragon'?"
For example, here's intels product listing for their Xeon Scalable Platinum product line (https://www.intel.com/content/www/us/en/products/details/pro...)
They use the code names prior to launch because they don't have SKUs out yet, and also because they are actually talking about different things, that sometimes don't have clean ordering.
AMD Genoa may have 96 cores and Ampere Altra Max has 128 (smaller) cores.
That sounds like a huge increase, if indeed they can keep all of them busy.
Various hardware costs get more expensive over time, so the fix is to make things bigger to amortize those costs. I think that even big players don't like running big instances (Google, Facebook, etc.) because when you lose a machine you lose a large % of the capacity for a service. Better to run multiple services so each of those lose a small % each.
You can easily and efficiently use 80 cores for inference
This should give more than enough memory bandwidth to double the number of cores.
I'm personally excited by the memory bandwidth and PCIe Gen5 more than the number of cores. I'm hoping we can build a single socket 800Gb/s Netflix OCA :)
[1] https://www.tomshardware.com/reviews/glossary-hbm-hbm2-high-...
So one stack of HBM2E can hold 24 GB, accessed at 410 GB/s. And we can have multiple stacks per package - i see lots of diagrams with four stacks. So a plausible scenario is 80 cores with 96 GB and 1640 GB/s of HBM2E.
Per core, that's 1.2 GB and 20.5 GB/s. That's ~320 million cache line fills per second. If the clock speed is 4.6 GHz, that's a cache line fill every 15 cycles, or a 64-bit load every 1.8 cycles. That seems pretty good!
OTOH if you run interpreted code, like Python or Ruby, your entire interpreter and bytecode are in L3 cache, but the amount of CPU they spend is large anyway.
These are actually perfect for building large C++ and Rust code bases.