Fusion die-shot - 2009 Analyst Day

Just curious regarding Llano, how does the raw shader/texture performance compare to current consoles? 2x faster or so?
Considering we don't really know the numbers for llano that's a bit speculative isn't it...
Xbox 360 has 192 shader alus compared to the 400 of redwood. Texturing is closer with 16 tmus vs. 20. Redwood also ships with significantly higher clock, so if you assume llano is similar it's definitely more than twice the alu throughput. Of course that ignores all other differences (shader alus of redwood are significantly more capable than xenos, differences in rops (edram),...).
 
<a bit speculative mode>
Xenos is ~192SPs (48*Vec4) @500Mhz.
In raw max theoretical shader power (ie ignoring stuff related to the eDRAM & functional differences) that is somewhere between an x1900GT & XT.
A 5550 would be about 1.8X (320SP @550Mhz) while 5670 would be over 3x (400SP @775Mhz).

For bandwidth: Xbox360 has 22.4GB/s to RAM shared between GPU & CPU.
Dual-channel DDR3 1600 is 25.6 GB/s.

X1900GT has 42.2GB/s, XT 46.4GB/s.
3870 72GB/s (GDDR4 version)
5550 25.6GB/s (GDDR3)
5670 64GB/s

So bandwidth wise Llano is probably going to be in the same ball-park as Xbox360/5550.
From that I would expect its GPU performance to be about same as 5550 too.

Thats a slower core & much lower bandwidth than 3870 but still vastly more powerful than existing integrated GPUs eg 4290 on the new AMD 890G chipset with only 40SPs @700MHz & with 2.6GB/s sideport + ~20GB/s shared main memory.

Sources:
http://www.xbitlabs.com/images/news/2004-04/xbox2_scheme_bg.gif
http://en.wikipedia.org/wiki/Ati_graphics_card_comparison
http://en.wikipedia.org/wiki/List_of_device_bandwidths

Also has anyone been able to work out how that 48*Vec4 fits into the Xenos die shot posted by Jawed? I see 4 groups of 3 that seem to be about equivalent to the SP blocks in the RV770 die shot.

They must be 16 wide ie 4*Vec4 each though & seem small for that? I guess some of the area opposite should be included & somewhere along the way ATI did re-jig the relationship between ALU quads and TMUs/ROPs which might account for some of that structure looking similar but for fewer units?
 
Thats a slower core & much lower bandwidth than 3870 but still vastly more powerful than existing integrated GPUs eg 4290 on the new AMD 890G chipset with only 40SPs @700MHz & with 2.6GB/s sideport + ~20GB/s shared main memory.
20? Anh... 4GT/s 16 bit hypertransport means 8GB/s, 16GB/s if you are nice and aggregate read and write bandwidth...
 
Hmm, yeah good point. ~20 is DDR3 1333 & the HT is between the RAM & GPU.

Anyway, reinforces that Llano should be much better than existing integrated GPUs for both bandwidth & raw power.
 
Also has anyone been able to work out how that 48*Vec4 fits into the Xenos die shot posted by Jawed? I see 4 groups of 3 that seem to be about equivalent to the SP blocks in the RV770 die shot.
<snip>
They must be 16 wide ie 4*Vec4 each though & seem small for that? I guess some of the area opposite should be included & somewhere along the way ATI did re-jig the relationship between ALU quads and TMUs/ROPs which might account for some of that structure looking similar but for fewer units?
Another look on Xenos:

xenos.png


Green boxes should be the shader pipes, yellow -- texture units; the four blocks in cyan are mystery.
 

Oh, in terms of i/o this is an almost exact copy of pinetrail atom. Single-channel ddr3. Connection to "southbridge" which is essentially 4xPCIE (Gen1). 4 x1 PCIE (Gen1 or 2 - the pic is inconsistent there, it could be the same pseudo-gen2 pcie as intel is often using - that is including the minimal protocol changes of gen2 but at gen1 speed). Though probably the display connections won't be that crippled and I wonder how many displays this can drive simultaneously. What the heck is "embedded probe header"? I guess it will be interesting where this thing will end up performance, price, and/or power wise compared to Atom/CULV.
 
If fusion will be mainstream and so will support external graphic, to what will connect if cpu+fch support only 12x pcie?

And what happened to HT3? 0_0
pull out a recent picture! make me talk with him! i want a proof that he's still alive! 0_0
 
http://en.wikipedia.org/wiki/PCI

33.33 MHz clock with synchronous transfers
peak transfer rate of 133 MB/s (133 megabytes per second) for 32-bit bus width (33.33 MHz × 32 bits ÷ 8 bits/byte = 133 MB/s)

Thanks, although I do understand PCI but my outburst is rather centered about why?

None of the computers we use at work (or my home computer) have PCI.

I thought pretty much everything today were PCIE.
 
Thanks, although I do understand PCI but my outburst is rather centered about why?

None of the computers we use at work (or my home computer) have PCI.

I thought pretty much everything today were PCIE.

World is full of PCI hardware.

Many separate network,usb,firewire and disk controller chips are still pci (though pci may become bottleneck for network and disk controller)
Most sound chips/sound cards are pci.
Many "legacy io" (serial, parallel, or some "custom industrial bus") controller cards are pci.

Even if the motherboard does not have any pci connectors, it may have internal PCI interface to some chip in the motherboard.


Only some new disk and network controllers are pci-e, but there are still much more of those available for pci.
 
World is full of PCI hardware.

Many separate network,usb,firewire and disk controller chips are still pci (though pci may become bottleneck for network and disk controller)
Most sound chips/sound cards are pci.
Many "legacy io" (serial, parallel, or some "custom industrial bus") controller cards are pci.

Even if the motherboard does not have any pci connectors, it may have internal PCI interface to some chip in the motherboard.


Only some new disk and network controllers are pci-e, but there are still much more of those available for pci.

Aren't most of these controller chips sitting on SB these days?
 
Back
Top