Subj : Re: Intel: Once mighty, now falling?
To   : Accession
From : tenser
Date : Fri May 23 2025 02:38:14

On 20 May 2025 at 06:26p, Accession pondered and said...
 
 Ac> On Wed, 21 May 2025 00:51:26 +1200, you wrote:
 Ac> 
 Ac>  > What issues are those?
 Ac> 
 Ac> From what I've read, CPU burnout due to voltage spikes, and core parking
 Ac> apparantly not working correctly. Also performance issues (ie. not as
 Ac> good as originally advertised) due to them using the Zen2 codebase in
 Ac> order to rush it out the door because they were more familiar with it.

Yeah,the intertubes are pretty much all wrong, there.  Current
generation AMD microarchitectures are Zen 4 and Zen 5; the "code
base" in question might be AGESA, but mention of Zen 2 in there
is pretty sparse; Zen 2 is really ancient.  I can tell you that,
in particular, Turin is pretty zippy.

 Ac>  > I've been working with AMD chips for a few years now; this is on
 Ac>  > the server side, where we design and build our own boards, as we
 Ac>  > are a computer vendor.  We don't use AMD's proprietary UEFI-based
 Ac>  > firmware (that is, we use neither AGESA nor OpenSIL).  We wrote
 Ac>  > our own, directly in the operating system (we run Unix).
 Ac> 
 Ac> What do you recommend if one were to build a new homelab, replacing an
 Ac> Intel Xeon E3-1230 CPU? Mind you, I wouldn't be looking to spend a ton,
 Ac> but would at least want better than what I currently have (more than 4x2
 Ac> cores/threads).

Homelabs aren't really something I'm super up on, but that sounds
like Sandy Bridge; basically anything is going to be ok.

On the AMD side, you can probably get a Milan-based server, or even
something based on Genoa, pretty reasonably.  Personally, I'd go for
that.  You can probably get 16 cores/32 threads for under $2k, but
I'm speculating.  (Our machines are rack-scale, and go for about a
million dollars a pop; but you get 32 compute sleds with 128 HW threads
and 1TiB of RAM each, plus about 48 TiB of disk and 100 Gbps to a
custom switch).

 Ac>  > My sense is that Intel's failures are a) they were really late
 Ac>  > getting onto a 7nm process, basically missing the boat on that
 Ac>  > one, and they haven't been able to keep up with AMD on the power
 Ac>  > consumption/heat side.  Frankly, AMD is just producing a better
 Ac>  > chip.
 Ac> 
 Ac> Do we really need desktop processors that have 24 cores? It seems there
 Ac> have been issues ever since going over 8. I'm currently running a Core
 Ac> i9-9900k with 8 cores and 16 threads. Never felt the need to overclock,
 Ac> and I can throw anything at it that I would normally do, and yet it
 Ac> continues to run super smooth. *shrug*

Honestly, I have no idea.  I think that Intel on the desktop is
a losing proposition in the long term.  As for whether you _need_
it?  I suppose it depends on what you're using it for.  For high
end CAD or graphics processing, or a developer workstation where
you're doing lots of big builds, sure: more is better.  Since
Denard scaling stopped a few years ago, the only way we're going
to get better performance is to increase parallelism, and since
we want to maximize bandwidth to DRAM and number of memory channels
is in some sense a function of core counts, then yeah.

But for day-to-day consumer use, is a desktop machine even that
useful?  Probably not.  Most end users are probably better off with
a laptop and an external monitor+keyboard/mouse.

 Ac>  > Also, Intel keeps favoring the x86 business over other, possibly more
 Ac>  > lucrative spaces.  Canceling Tofino seemed like a huge mistake, for
 Ac>  > example; there was literally nothing else on the market that did what
 Ac>  > Tofino2 could do.  It was short-sighted madness.
 Ac> 
 Ac> I take it this is more in the commercial market? I only deal with my
 Ac> personal computers, I don't work in the field or anything, so am lacking
 Ac> a lot of knowledge as to what goes on in the industry.

Yeah.  Basically, they had an ASIC that was almost unique in the
industry in terms of functionality for building high-end switches
and routers, but they canceled it.  We had built our switch
product around it (32-ports of 200Gbps capacity with a high-speed
PCIe link to one of our compute sleds to control the whole thing),
and finding a replacement was an interesting exercise.

I remember when DEC was failing; they were selling off successful
business units to try and preserve their high-margin server
business, selling their big VAXen and Alpha boxes running VMS.
They gave away almost everything: Alpha, the networking division,
etc.  I see Intel making similar mistakes to try and preserve the
x86 business.

.... I know a good tagline when I steal one!

--- Mystic BBS v1.12 A48 (Linux/64)
 * Origin: Agency BBS | Dunedin, New Zealand | agency.bbs.nz (46:3/203)

.