AMD: R8xx Speculation

[aaronspink]

I was quoting realworldtech ... the point being that 64B data transfers are not always needed. Though, honestly my data is a bit old, maybe in days of Win7 and quad cores bursting large memory chunk is the way?

Doesn't matter what you need. Cachelines these days are generally 64B so that is what you have. You can always used non-caches access with orders of magnitude higher latency and lower bandwidth!

Of course adding more "generic" channels is better, but atm AMD can't afford socket change. Or rather they don't want to change it.

then the side-port question is moot, cause if they can't change the socket then they can't add it when they integrate graphics.

Intel thinks that quad-and-more core CPU needs triple channel DDR3 memory, yet you claim AMD's quad-cores can't use dual-channel ...

The data is available if you want to look. There are factors in the AMD design preventing them from being able to utilize the 2x raw handwidth available with DDR3. The numbers from various memory benchmark results show this. Either they have a bottleneck in the internal interconnects, they can't support enough misses to satisfy the round trip latency, or they have an issue within their memory controller.

And benchmarks show that 780G + 128MB sideport is way faster than same chipset without it.
So, perhaps someone is distorting reality & benchmarks & Intel +-

I'm sure it is, the 780G is sitting at the arse end of the universe with a slow HT about as much bandwidth as the sideport. Now move the graphics so it isn't sitting at the arse end of the universe...

 

[chavvdarrr]

hopefully we'll see in few months what AMD thinks about Fusion , I'm just a bit pessimistic about supposed performance level.
If they started a new line, now, why should they "kill it" with IGP twice more powerful?

 

[mboeller]

hopefully we'll see in few months what AMD thinks about Fusion , I'm just a bit pessimistic about supposed performance level.
If they started a new line, now, why should they "kill it" with IGP twice more powerful?

Wait till tomorrow. AMD will present some aspects of the Fusion-APU today in the afternoon at ISSCC 2010

Hopefully the presentation will include some aspects of the GPU-side of the APU too:

The 32nm implementation of an AMD x86-64 core occupying 9.69mm2 and containing more than 35 million transistors (excluding L2 cache), operates at frequencies >3GHz. The core incorporates numerous design and power improvements to enable an operating range of 2.5 to 25W and a zero-power gated state that make the core well-suited to a broad range of mobile and desktop products.

[edit]: spelling error

 

[mczak]

Intel thinks that quad-and-more core CPU needs triple channel DDR3 memory, yet you claim AMD's quad-cores can't use dual-channel ...

The inability for the AMD chips to actually really use that bandwidth was already mentioned, but it's not exactly the case that a triple channel core i7 blows away a dual channel core i7 neither... Looks to me like the third memory channel is there mostly to be able to use more memory plus for future chips with more cores...

And benchmarks show that 780G + 128MB sideport is way faster than same chipset without it.

Way faster? Here's some direct comparison for some 785G chipsets (and 790G), one with sideport, one without. http://www.technic3d.com/article-91...nimal-update-drei-mainboards-im-vergleich.htm (that's the first I found, I'm pretty sure you can find others). A 5% difference doesn't exactly blow me away...

 

[Mindfury]

5990, 5890, 5860. There I said it. Probably April, since all 5970/5870/5850 are EOL end of March (EOL as in unavailable to order from then on). Hope my sources are correct or I expect to be quoted in sigs in the near future

http://www.rage3d.com/board/showthread.php?t=33958553

Is this reliable?

 

[Bouncing Zabaglione Bros.]

http://www.rage3d.com/board/showthread.php?t=33958553

Is this reliable?

Kombatant usually is - he's been a long time beta tester and community ATI supporter. That's not to say he can't be wrong, but his words deserve more weight than the average forum speculation.

 

[neliz]

I sure hope no one is following that fake stuff from wikipedia.

 

[ferro]

Is SilentPCReview atypically the first with an HD5570 review? Or have I missed something?

 

[Albuquerque]

Is SilentPCReview atypically the first with an HD5570 review? Or have I missed something?

From page 6 on, I get a "Not authorized to view this page"

That 5570 seems like a pretty decent setup for someone like my brother who needs a card with a bit of "oomph" but without the external power connector.

 

[chavvdarrr]

Way faster? Here's some direct comparison for some 785G chipsets (and 790G), one with sideport, one without. http://www.technic3d.com/article-91...nimal-update-drei-mainboards-im-vergleich.htm (that's the first I found, I'm pretty sure you can find others). A 5% difference doesn't exactly blow me away...

looks like I was wrong

 

[ferro]

From page 6 on, I get a "Not authorized to view this page"

That 5570 seems like a pretty decent setup for someone like my brother who needs a card with a bit of "oomph" but without the external power connector.

The article is completely retracted now.

 

[Albuquerque]

The article is completely retracted now.

FFFuuuuuuuuuuuuuu.....

Well, quick recap for those who didnt' see it: half height, 400SP's @ 650Mhz (~500 GFlops), 1Gb GDD3 at 900Mhz (~10.8Gbps), ~28W more power draw (at full load) than the same system using Intel IGP, and a small active cooler.

 

[mczak]

looks like I was wrong

Or maybe you've seen scores with different setups. I wouldn't rule out that it made more difference with DDR2 and/or with old (HT 1000Mhz) cpus, though those score quite a bit lower (well at least with slow HT) anyway. Reviews testing IGPs with/without sideport and different ram/cpus are quite rare, unfortunately.

 

[mczak]

That 5570 seems like a pretty decent setup for someone like my brother who needs a card with a bit of "oomph" but without the external power connector.

If you just don't want the external power connector, you could also opt for the 5670...
That said, I guess the active fan means "fail" for the HTPC crowd. Not that some manufacturers won't come up with passive designs .
Too bad the article is gone, though I guess there weren't really any 3d benchmarks in there anyway. In any case, I'd certainly expect the performance drop to be much bigger relative to HD5670 than it was for gt240 going from gddr5 to ddr3 - not only because clocks are lower too, but also because it's a faster card to begin with hence it'll be more limited by memory bandwidth.

 

[ferro]

The article is completely retracted now.

The only performance numbers, salvaged from the retracted review:

                                     3DMark05   3DMark06

ATI Radeon HD 5570 1GB               8761       6326

ATI HD 4200 (Integrated, Sideport)   4095       1789

ATI HD 3300 (Integrated, Sideport)   4884       2205

ATI Radeon HD 5450 512MB             6231       3435

GeForce G210M (Asus UL80Vt)          6847       3460

PowerColor HD 4650 512MB (DDR2)      7622       4088

ATI Radeon HD 4670 512MB             8962       6341

Edit: and the conclusion:

Positioned just above the HD 5450 on ATI's depth chart, the $80 HD 5570 does deliver a sizable gain in 3D performance, putting it about on par with the HD 4670. Both are similarly priced with the 4670 costing about $10 less. In terms of power consumption, the 4670 does better at idle, while the 5570's advantage is on load. There's also the faster HD 5670 to consider; it can be had for as low as $95. A price reduction would help make a decision between the two easier.

Like the 5450, our 5570 sample does bitstreaming but lacks support for Eyefinity. The versions that will include it will undoubtedly drive the cost up somewhat, putting it in the same price bracket as the 5670. AiB partners will also likely come up with quieter cooling solutions as well for this 30W GPU. The reference cooler on our sample was adequate, but a little undersized and unnecessarily noisy given the card's high energy efficiency.

 

[thatdude90210]

That said, I guess the active fan means "fail" for the HTPC crowd. Not that some manufacturers won't come up with passive designs .

I would imagine the htpc crowd might be better off with the 5450. Their (now gone) chart show both idle and full load power at under 10W.

 

[Kaotik]

I would imagine the htpc crowd might be better off with the 5450. Their (now gone) chart show both idle and full load power at under 10W.

HD5450 is having some problems with deinterlacing though, compared to for example HD5600+ (and prolly 5500+, too)

 

[mczak]

The only performance numbers, salvaged from the retracted review:

                                     3DMark05   3DMark06
ATI Radeon HD 5570 1GB               8761       6326
ATI Radeon HD 4670 512MB             8962       6341

That's actually fairly interesting. Despite the 5570 having 10% less memory bandwidth, half the texturing rate, pretty much about the same alu capability (thanks to lower clock) it still is pretty much as fast.
Makes me wonder why AMD went with the 8-wide simds for the 4670 then, should have been a fair bit faster with the same config as redwood. I think in terms of transistors/die area it wouldn't have made much of a difference neither - you have more alus but less control logic and tmus (only 5 instead 8 simds/quad tmus). There seems to be next to nil advantage for having smaller simds (hence smaller batch size and more tmus).

 

[EduardoS]

Maybe because of the size of the chip, the SIMD cores are long, if HD4670 had 16-wide SIMDs it would be more rectangular instead of square, and maybe this is the reason for HD5450 having 8-wide SIMDs and the reason why I beat Llamo GPU will have 6 8-wide SIMDs for a total of 240 SPs

 

[Albuquerque]

If you just don't want the external power connector, you could also opt for the 5670...

Ah yes, duly noted. I neglected to mention that my brother is using a slimline tower from Dell, and finding half-height 5670's hasn't been easy so far. Or else, I'm just not looking in the right places.