AMD rips covers off 64-core Threadripper desktop monster, plus laptop chips, leaving Intel gesturing vaguely at 2021 AMD this week touted a bunch of new laptop and desktop silicon that put main rival Intel to shame. The raft of Radeon and Ryzen components – launched at this year's Consumer Electronics Show (CES) in Las Vegas, USA – will run between 10 and 20 per cent faster than Intel equivalents, according to AMD’s own testing. Included in …
Nothing says "power" like large projects like glibc and the Linux kernel compiling in seconds.
128 threads. That's technically 128 instances of GCC running at the same time (make bzImage -j 128). You could probably even push it harder with 256 instances.
This is the CPU for the impatient Linux developer.
They make a great development workstation for testing virtual machines.
You'll run into problems with throughput and memory (only has a max. of 256GB RAM). It also isn't designed as a server platform (motherboard etc.). I've been using a Ryzen 7 as a VM testbed for a couple of years and it is a great solution, but I wouldn't trust it for a live, production environment.
"It is fabricated using TSMC's 7nm process, and sports 32KB of L1 instruction and 32KB of L1 data cache per core, 512KB of L2 cache per core, and 256MB of shared L3 cache..."
It's been a while since I've been au fait with CPUs, but I'm surprised at the small size if the L1 cache. I suppose as its per core, it adds up, as well as depending on the speed ratios with the L2/L3 cache.
For comparison, the Pentium Pro (1995-1998) had an L2 cache between 512KiB-1MiB over its life. The Pentium MMXs that were the consumer grade chips around at the same time had 16KiB-32KiB. Am I missing something?
L1 cache size is chosen with a particular latency in mind. For example, you might be able to manage a 2 cycle latency at 16KB but need 3 cycles if you go larger.
That's the whole reason you have multiple levels of cache. Otherwise you'd just have a 16MB "L1" with a couple dozen cycles of latency and call it a day.
Caches are complex, they usually have multiple "ways" to make them more efficient, there is snooping involved with higher levels of cache since there is no longer such a thing a single CPU system, and must handle exclusive access for certain primitives used for locking.
So yes it is partly path length, but there is a lot more than that and the bigger it is the more time these things require - thus more cycles of latency.
Your Pentium from the 1990s is, like, single core, right? So there's space on the die for cache. With 64 cores, you can't bung too much on without producing dies that smash your yield targets.
Look at it this way: there's a total of 4MB L1 cache, 32MB L2, and 256MB of L3 in the 3990X.
And despite leaps in processor technology, it's likely today's software still works comfortably within 32KB working sets anyway, what with the latency issue DougS mentions above in mind.
C.
Er ... Threadrippers aren't a single die; they're multiple chiplets (dies).
And whilst the working set of a single piece of software may well fit into 32Kbytes, if you're only running one piece of software I'd be very surprised - I'm currently running nearly 700 processes here and nearly 900 at home (on a first generation threadripper).
Small L1 cache sizes like this are annoying if you’re into DSP. For example if you’re doing a lot of FFTs, it’s convenient if the twiddle factors (which remain constant), input, intermediated and output (which may overwrite the input) all fit in L1. That way the fastest possible compute times are achieved. With only 32kB of L1 cache, the size of FFT that fits entirely in L1 is fairly small. This may be ok for some applications but it can be an annoying limitation for others.
The Cell processors (as found in Sony’s PS3) were radically different. They had no cache at all, but had 8 math cores each with 256kB of static on chip RAM instead of L1, and huge bandwidth between those RAM chunks and external RAM. This was excellent because it was a lot bigger, but were notoriously difficult to code for (unless you were already into that kind of thing). Though sadly abandoned by IBM who’d developed it, it took 10-ish years for Intel to finally design a CPU with more maths performance than the Cell. It was a truly awesome piece of silicon and drew only 80W.
Some POWER CPUs from IBM have 64kB L1.
Edit: Intel's new Ice Lake has changed its L1 config which might break quite a lot of optimisations
Yes it might!
A thing Intel did get fairly right I think was to do the MKL / IPP libraries, and encourage adoption of those. Apologies if you're already deeply familiar with those, ignore the rest of this post!
Use those, let them work out combinations of threads, etc. that best suit the CPU architecture that the program finds itself running on, and the developer only really needs to write single threaded code to get good general purpose performance. Optimising code for a specific CPU by hand can result in better performance, but one has to be pretty determined to buck the cost / benefit curve for that to be worthwhile.
A difficulty with the approach is that it only works optimally on newer CPUs if one has a sufficiently recent version of the MKL / IPP library that knew how to optimise itself on them. This means you have to keep issuing software updates. The OSes don't provide these libraries as part of their default set of packages, meaning their software update mechanism isn't doing this.
Anyway, provided that this is addressed, a software dev doing it the Intel way probably can probably ignore the changes in L1 config. Unless Intel have fallen behind the curve too.
...after vomiting up that load of tripe:
Here’s the canned excito-quote for this range, attributed to Radeon head Scott Herkelman: “From heart-pounding esports competitions to eye-popping AAA blockbusters, today’s games demand more performance, higher framerates and lower latency than ever before… We’re committed to providing all 1080p gamers with the raw horsepower and incredible features that enable the absolute best gaming experiences with all settings maxed out.”
I've just upgraded form a quad core, to a 8core/16 thread monster...and most of the time all 16 threads are ticking along at <5% usage.
Still, transcoding video using Handbrake does use all the cores at full speed, and lets you find out if your cooling system is up to the job.
my computer goes a bit like this >>>>>
I watched a presentation by Sophie Wilson on CPU design progress. It was from 2016, and she said that the latest and greatest phone CPUs couldn't dissipate enough heat to run flat-out, and half the cores had to be off (on average, I guess).
So, yeah. It's like the old F15 vs Concorde thing. the F15 will take it in a sprint, but Concorde has the legs to just keep going. And that analogy shows my age, but there you go...
27nm - that was the other thing that stuck in my mind from that lecture. 27nm was the optimum feature size for cost. Any further required crazy interferometry. And her description of x-ray lithography was just astonishing.
As n ardent AMD CPU subscriber for my personal computers and small business servers in 1990s and early 2000s, I am pleased that AMD can now offer chipset that are proven superior to anything from Intel.
However the concerns for me and many of my clients now seems to be that the company has geared all it's development towards the Microsoft Windows realm almost exclusively, thus eliminating significant adoption in the technology spheres where almost all the implementation, and therefore sales are Intel based - like Data Centers, Cloud Computing infrastructure, Supercomputing research for medical, aerospace, defense and climate change, web hosting and Mobile telecommunications data centers that all deploy Linux or BSD UNIX-like (*NIX) software.
Even in the area if graphics and 3-modeling, I am aware that "every" major animation studio now use Linux design workstations instead of Apple MacOS and even set up their rendering farms on Linux instead of Windows server - at least according to three of these large animations companies' official statements, and confirmed public knowledge that most all animated movies from Toy Story1 through 4, the Shrek series to Avatar and beyond use Linux, not Windows.
Even NASA, Airbus and Elon Musk's SpaceX company use non-Windows almost exclusively in their engineering, design and programming software models.
Why then is AMD so hung up on Microsoft Windows? - except maybe for Gaming base, which will never bring the company to substantial earning, especially in the industries mentioned above. Note: Microsoft Azure is 64% plus Free/Open Source Software (FOSS) based, including *NIX verbally enunciated by company tech executives during a SuSE Linux Expert Days conference in NYC in early and again in late November 2019.
Financial success has and always be in enterprise and government sales, not niche enclaves.
As n ardent AMD CPU subscriber for my personal computers and small business servers in 1990s and early 2000s, I am pleased that AMD can now offer chipset that are proven superior to anything from Intel.
However the concerns for me and many of my clients now seems to be that the company has geared all it's development towards the Microsoft Windows realm almost exclusively, thus eliminating significant adoption in the technology spheres where almost all the implementation, and therefore sales are Intel based - like Data Centers, Cloud Computing infrastructure, Supercomputing research for medical, aerospace, defense and climate change, web hosting and Mobile telecommunications data centers that all deploy Linux or BSD UNIX-like (*NIX) software.
Even in the area if graphics and 3-modeling, I am aware that "every" major animation studio now use Linux design workstations instead of Apple MacOS and even set up their rendering farms on Linux instead of Windows server - at least according to three of these large animations companies' official statements, and confirmed public knowledge that most all animated movies from Toy Story1 through 4, the Shrek series to Avatar and beyond use Linux, not Windows.
Even NASA, Airbus and Elon Musk's SpaceX company use non-Windows almost exclusively in their engineering, design and programming software models.
Why then is AMD so hung up on Microsoft Windows? - except maybe for Gaming base, which will never bring the company to substantial earning, especially in the industries mentioned above. Note: Microsoft Azure is 64% plus Free/Open Source Software (FOSS) based, including *NIX verbally enunciated by company tech executives during a SuSE Linux Expert Days conference in NYC in early and again in late November 2019.
Financial success has and always will be in enterprise and government sales, not niche enclaves.
Short answer; because there's a lot more desktops and laptops out there, running Windows in people's homes, than there is large animation studios.
Sure, AMD probably make more margin (which is a polite way of saying "overcharge") on their server chips, but the low end market outpaces that in volume, by spades.
Financial success comes from the end user market, not government and enterprise.
This commenter is basically very ignorant about the technology industry in regard financial earnings.
The US Defense department alone just awarded a 10 plus billion dollar contract for Cloud Computing, which at this time is almost 100 % Intel server based. The vast majority of earnings for IBM, Oracle and other technology behemoths is in Intel data center servers . Wall Street and all USA/International Banking and Stock Exchange earnings reports for all major companies spell this out.
In what country is this person living? Even if they resided in North Korea, or Bangladesh, Myanmar or other less sophisticated, developed nation they would know this.
TheRegister needs to attract more informed readers.
Thanks for that particularly passive-aggressive insult, I suppose it's more gentile than "omg ur a idiot".
Anyway, a quick glance at AMD's most recent earnings show that they made $1.28B in revenue from desktop and mobile CPUs, and only $525M from enterprise and embedded (which includes the CPUs for the PS4 and XboxOne, I can't find any figures that just break out the server/enterprise revenue).
AMD are concentrating on where they make the most of the their money (ie consumer), rather than enterprise, an area where Intel completely dominate.
So perhaps I should clarify my previous comment, AMD's financial success comes from the end user market.
Latest earnings from Intel show a 50-50 split between Desktop/Laptop revenue and Datacenter revenue of all kinds. So actually its very balanced and AMD doing better in the consumer market will better them for the enterprise and large corp/governments. No one has trusted AMD for a decade and that takes time to recapture even with very good products because Intel have such a huge market share and are making billions in all markets.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
