[KLUG Hardware] AnandTech and Computex Coverage -- Hammer Time, Chipset Central ...

[Bryan J. Smith]

Sorry I'm a bit late on this.  I'll just review some basics, jump on
some goodies and leave the rest to the links below ... (which also 
include details on non-processor/non-chipset stuff).

Overview:
- It's Hammer Time (AMD K8 / Opteron)
- AMD K8 Hammer HyperTransport "Nodes"
- Intel "Prescott" and Processors
- Intel Pentium 4 Chipsets
- ATI goes OEM, IGP and boards
- Links ...


- It's Hammer Time (AMD K8 / Opteron)

In case you missed my previous "discussions," the 64-bit (x86-64) AMD K8
"Hammer" will be released in 3 flavors ...

 Official Name  Core    Pin  Mem+IO**/CPU  CPU  Release
 -------------  ------  ---- ------------  ---  -------
 Athlon "64"?   Claw    754  3.2/ 6.4GBps   1   2002Q4
 Opteron "DP"?  Claw    754  3.2/ 6.4GBps  1-2  2003H1
 Opteron "MP"?  Sledge  940  6.4/19.2GBps  2-8  2003H1

[ **NOTE:  I/O includes CPU-to-CPU interconnect ]

Again, if you're late to the party of the AMD K8, there is no such thing
as a "front side bus" (FSB) because there is no such thing as a
"northbridge" chip in the chipset.  The memory (PC2700/PC3200,
166/200MHz DDR) and I/O are completely on the chip now.  This also means
that each chip has its own memory bus.  For I/O as well as
multiprocessor/memory interconnect, ClawHammer has (1) 6.4GBps
(32+32-bit, 400MHz DDR?) HyperTransport and SledgeHammer has three (3)
independent 6.4GBps HyperTransport links.

The difference between the uniprocessor Athlon "64" and the
dual-processor Opteron "DP" is marketing and package "lock out" pins for
dual-processor as they both use the same Socket-754 and ClawHammer core.

- AMD K8 Hammer HyperTransport "Nodes"

Since the traditional chipset is now out, there are just HyperTransport
"Nodes" that are "glued" to the processors.  These "nodes" provide AGPx8
(32-bit, 266MHz AGP 3.0), PCI (32/64-bit, 33/66MHz) or PCI-X (64-bit,
66-133MHz) and legacy PC (ATA, ACR, etc...).  The Audio Communication
Riser (ACR) is the forward compatible replacement for the Audio Modem
Riser (AMR), and designed as a "more complete" competitor to the
Communication Networking Riser (CNR) -- since it can do both AMR and CNR
in one slot (that looks a like an up-side down PCI slot).

The greatest thing about HyperTransport is that its _already_proven_! 
Okay, maybe not at the 6.4GBps speeds that Hammer will use, but in
existing products upto 1GBps.  So just like the proven Alpha EV6 bus of
before the Athlon used it, AMD is NOT going in "blind" here with
Hammer.  "Gluing" more HyperTransport logic on a mainboard is radically
simplified from the more "custom" interconnect designs of mainboard's
prior.  No wonder it's taking over even Intel mainboards and chipsets! 
Since K8/Hammer "chipsets" can work with _any_ combination of
processors, 1-8, since they are just HyperTransport "Tunnels" or
"Hubs."  Logic of interest ...

  AMD's 8000 series
  8151 (AGPx8 Tunnel), 8131 (PCI-X Tunnel), 8111 (PCI/LPC Hub)

Usually only servers and high-end workstations will use the 8131, and
servers might not use the 8151.

  ALi's M1687 (AGPx8 Tunnel) and M1563 (PCI/LPC Hub)

The M1687+1563 will be one of the first Hammer logicsets on the market,
much like ALi's first DDR SDRAM chipset.  Hopefully it will be more
reliable than that "first."

  nVidia CK8 chipset (AGPx8 Tunnel + PCI/LPC Hub)

CK8 has a built-in NV17 (GeForce4 MX4x0 series), may or may not use main
Hammer memory.  The CK8 will be mirrored on the 32-bit Athlon platform
with the "Crush18/nForce2" product with PC3200 (200MHz DDR, 400MHz
effective) SDRAM support.

  SiS 755 (AGPx8 Tunnel) and 963 (PCI/LPC Hub)

All new SiS products have both USB 2.0 _and_ IEEE1394 FireWire. 
Finally, FireWire is _standard_ on the PC!  SerialATA wasn't mentioned,
so it may not be in the 963, but possibly a follow-on?

  ViA K8HTA (AGPx8 Tunnel) and ? southbridge (PCI/LPC Hub)

Interesting enough, it appears ViA does not yet have it's own
HyperTransport chip for typical southbridge support.  But that's the
power of HyperTransport, dropping in any other vendor's chip will do,
assuming they use the same width and range of speeds.  800MHz (8+8-bit,
200MHz DDR?) seems to be the typical HyperTransport connection coming
off the AGPx8 Tunnel HyperTransport nodes from AMD, ALi and SiS. 
Gigabyte, who only uses AMD and ViA, has been reported to be using an
ALi PCI/LPC Hub with the ViA K8HTA.  Others have reported the K8HTA
_does_ have a "legacy" VLink channel so connecting a legacy ViA
southbridge is an option.

- Intel "Prescott" and Processors ...

"Prescott" will be Intel's final IA-32 generation Pentium.  It will
appear in 2003Q3, although Intel is saying it could be released by as
early as 2003Q2.  It features 0.09um sizes, 1MB L2 cache, 3GHz+
operation (possibly 4GHz by end of 2003), 667MHz (and 800MHz planned)
FSBs along with a few tweaks including HyperThreading.  Intel should be
able to match the performance of any x86-64 processor from Intel.

IA-64 Itanium2 will still be a server processor, although a desktop
equivalent is planned for release in early 2004.  If AMD has succeeded
in capturing some of the market with its x86-64 Hammer products, Intel
plans on releasing its licensed AMD x86-64 clone "Yamhill" if
necessary.  E.g., if "Prescott" does not offer enough performance versus
Hammer or 64GBs is not enough for servers and/or IA-64 is too expensive
compared to the Hammer, "Yamhill" will most likely make an appearance.

- Intel Pentium 4 Chipsets ...

  Intel "Granite Bay" (dual-DDR) and ICH5 (SerialATA, USB 2.0, NIC)

As Intel phases out RDRAM, DDR SDRAM will take its place.  The "Granite
Bay" chipset houses the first dual-DDR SDRAM memory controller (outside
of the AMD SledgeHammer) on the PC.  It will be paired with the new ICH5
southbridge that offers SerialATA, upto 8 USB ports (some are
2.0/480Mbps capable) and an integrated NIC with everything for 802.3
(Ethernet) as well as 802.11 (WLAN) except PHY chip (which would be
provided on an CNR card).  No FireWire though

Yep, almost 5 years ago Intel _promised_ to put FireWire on the PIIX4
(i430TX and 440LX/BX southbridge) and would end up _never_ doing so. 
Despite all the performance issues with USB 2.0 -- namely the fact that
even ATA drives run "real world" at 1/4th-1/5th their native speed over
USB 2.0 (compared to about 1/2-2/3rd over FireWire, let alone all the
other features USB 2.0 doesn't have that FireWire does) -- Intel still
refuses to put FireWire in the chipset.  Why I still have no idea --
Apple doesn't ask much for a license and Intel's chipsets are the
highest priced of the lot anyway (so you _know_ they can afford it!).

  SiS 658 (PC1200 RDRAM) and 648 (PC2700 DDR) chipsets offer AGPx8

SiS, an official Pentium 4 chipset licensee, will be introducing the
first non-Intel RDRAM chipset with the 658.  The 658 will have a
dual-channel, 600MHz DRDRAM (2 x 16-bit x 1.2GHz effective) and support
FSBs of 533MHz (and 600MHz?).  The 648 has a single-channel, 200MHz DDR
SRAM (400MHz effective) memory controller with FSB of 533MHz (and
600MHz?).

  AMD HyperTransport rules Intel chipsets and mainboards

Both the SiS 963 and ALi M1563 HyperTransport nodes for PCI/LPC will
make their appearance Intel mainboards.  SiS and nVidia already use
HyperTransport on their current Intel chipsets, and ALi will join the
fray (in the M1681+1563 combo).  HyperTransport greatly simplifies
engineering effort and redesign of support logic, and gives a
significant speed boost over current ViA-Link (VLink), Intel ICH and
other, custom approaches (800MBps is typical versus 128-512MBps of
others) for a lot less noise/EMF.

- ATI goes OEM ... IGP and boards

ATI has traditionally been like the Adaptec or Promise only of the video
card market, retail-focused with only a few chips for mainboards and PC
OEM modification of existing, retail products.  But that is changing as
they are introducing a northbridge with an IGP (integrated graphics
processor) to compete with nVidia's nForce, as well as allowing OEMs to
sell their own ATI-based video cards.  Times are changing!  Maybe we'll
see accelerated GLX (OpenGL on X-Windows) drivers for Linux soon too!

- Links ...

Anand's Days 1, 1 (part 2), 2 and 3: 
http://www.anandtech.com/showdoc.html?i=1630 
http://www.anandtech.com/showdoc.html?i=1631 
http://www.anandtech.com/showdoc.html?i=1632 
http://www.anandtech.com/showdoc.html?i=1633 

Tom's Days 1, 2, 3 and 4:
http://www4.tomshardware.com/business/02q2/020603/index.html 
http://www4.tomshardware.com/business/02q2/020604/index.html 
http://www4.tomshardware.com/business/02q2/020605/index.html 
http://www4.tomshardware.com/business/02q2/020607/index.html 

-- Bryan

P.S.  I don't know what Tom's problem is, but he, and everyone on his
site, have gotten might anti-AMD as of late.  They do this to the point
of being very biased in reviews (e.g., stacking the deck against AMD in
his "selective" benchmarks that make heavy use of lossy SSE operations),
and is overly assumptive in his analysis/predictions (like the Athlon
"Ultra 3400+" BS).  He seems to have done that for 6-9 months with other
companies before -- e.g., 3dfx.

But with AMD, I find this odd.  Because _he_ was the first person to
document that Intel's SSE floating-point using dedicated SSE registers
and pipes was very "lossy" in comparison to AMD's 3DNow "Enhanced"
(which is basically SSE, right down to the opcode/parameters) which uses
microcode to leverage its existing 3-issue FPU which is very accurate. 
He did this in comparing quality tests in MPEG-4 encoding, Intel SSE v.
AMD SSE v. Intel FPU v. AMD FPU and made it known that the Intel SSE
instructions are NOT for precision computing -- or even high fidelity
graphics, like when rendering production 3D/video.

-- 
Many Windows users would rather buy and hunt down dozens of utils
for their Microsoft software than try Freedom Software that not
only includes those same functions inheritly, but lacks spyware.
-----------------------------------------------------------------
Bryan J. Smith, SmithConcepts, Inc.     mailto:b.j.smith@ieee.org
Engineers and IT Professionals       http://www.SmithConcepts.com