AMD: R8xx Speculation

[FrameBuffer]

Ummm I think that the 5450 is the new 2400/3450/4350. It's not supposed to be impressive, it's supposed to sell for $30. I've picked up a few of its predecessors to use as ultra basic Bluray accelerator cards in some HTPCs. Actually I've had one overheat while doing it too because the ultra mega cheap heatsink sucked too much.

Another area where the 5400 falls flat.. price, as you said it replaces the 2400/3400/4400 line but at nearly twice the cost.. (nevermind the crippled Avivo!) While the power consumption looks great on paper, the product is all but trash (IMO) for the end user, better (older/cheaper) products give no reason to "upgrade". It seems to me ATI was just too damn cheap spec wise, I'm sure OEMs will LOVE it's diminutive size/power and DX11 check mark to their product lines though.

 

[willardjuice]

I think bitstreaming support is a pretty big plus for htpc (it is for me at least). If the interlacing issue can be resolved for video (a big if), it would be a perfect htpc card.

 

[Dave Baumann]

It seems to me ATI was just too damn cheap spec wise...

Really? Adding DX11, DirectCompute, higher performance OpenCL, Bitstream Audio, Eyefinity / more panel support aren't necessarily "cheap".

 

[GZ007]

Really? Adding DX11, DirectCompute, higher performance OpenCL, Bitstream Audio, Eyefinity / more panel support aren't necessarily "cheap".

DX11, DirectCompute , higher performance OpenCL with a single SIMD is quite questionable.
Hell, even on 5670 those things could be sometimes questionable.
I think that if u double the high-end performance with each generation , u should at least double the rest of the lineup or the whole thing doesnt have to much meaning in the long run.

 

[Silent_Buddha]

Meh, I personally don't see the need to up performance in the lowest budget sector. It is what it is. A cheap card that should offer more performance than integrated with more features.

Anyone that does any significant gaming is going to picking up cards closer to 80-100 USD on the low end.

Yes I realize that for enthusiasts performance in benchmarks and games means everything, but I don't see many enthusiasts seriously considering anything under 80-100 USD for gaming even on a HTPC.

More units would have meant more power (BOOOO). In this segment with it's intended purpose, I'm far more concerned with power consumption (directly related to heat output), video output capabilities, and audio output capabilities (over HDMI).

The deinterlacing thing doesn't particularly bother me. The 3450 is a far less capable chip, yet has played all my video content quite well. It's only the audio capabilities I've been wishing were better.

DirectCompute and OpenCL may turn out to be interesting depending on whether they get leveraged into HTPC friendly applications. Video transcoding assistance might be interesting for example. Especially with the GPU's ability to work in parallel (on the same task) with the CPU in OpenCL.

That said, it would be an added bonus for the "checkbox people" if it performed better in games, but really that's completely irrelevant to me.

Regards,
SB

 

[mczak]

DX11, DirectCompute , higher performance OpenCL with a single SIMD is quite questionable.
Hell, even on 5670 those things could be sometimes questionable.
I think that if u double the high-end performance with each generation , u should at least double the rest of the lineup or the whole thing doesnt have to much meaning in the long run.

Oh but it has two simds - they are just only 8 wide .
That said, AMD really went for "all features in no die size" this time around. rv710 was reportedly 73 mm², Cedar 59mm². I don't know how big a 16-wide simd would be in terms of die area vs. a 8-wide one, but I'm quite certain that with 2 16-wide simds the chip still would have been well below 70 mm². Might also have solved that video processing problem when it's just a bit too slow (unless that code really benefits from small batch size). Wouldn't have made it a killer chip but definitely faster than g210 and rv710. But hey nvidia's g210 certainly isn't any faster even now (and has pretty much the same die size), so from that point of view I guess the chip has "enough" performance. And I guess they aren't really expecting people to upgrade from IGP to these lowest-end chips any longer, hence there's no need to outclass those completely.

 

[FrameBuffer]

Really? Adding DX11, DirectCompute, higher performance OpenCL, Bitstream Audio, Eyefinity / more panel support aren't necessarily "cheap".

and of those how many are actually USABLE for such a low end part.. bitstream audio and eyefinity. Of which, Eyefinity is almost nil use for HTPC, leaving bitstream support.. which would be negated seeing as the 5400 series has castrated Avivo post processing (and atm bugged de-interlacing),.. enough so that anyone interested in the audio feature would pass over the 5400 series if they wanted decent HD representation.

 

[Silent_Buddha]

and of those how many are actually USABLE for such a low end part.. bitstream audio and eyefinity. Of which, Eyefinity is almost nil use for HTPC, leaving bitstream support.. which would be negated seeing as the 5400 series has castrated Avivo post processing (and atm bugged de-interlacing),.. enough so that anyone interested in the audio feature would pass over the 5400 series if they wanted decent HD representation.

Castrated? It is the same or improved in all ways (UVD) compared to the previous model. I'm not sure how you can call that castrated.

My 3450 has even worse theoretical processing. Yet I'd hardly call it castrated in any way.

Even Anandtech that went into it in great depth only called it a very minor problem that won't affect or even be noticed by the majority of users.

And how do you know DX compute and OpenCL won't be useful on this card. We don't even have any applications to try yet.

Regards,
SB

 

[Dave Baumann]

DX11, DirectCompute , higher performance OpenCL with a single SIMD is quite questionable.
Hell, even on 5670 those things could be sometimes questionable.

Relative to RV710, no though are not questionable. Relative to the system they are being either built with, or upgraded into? Again, transocding benefits over the CPU can be had relatively simply.

I think that if u double the high-end performance with each generation , u should at least double the rest of the lineup or the whole thing doesnt have to much meaning in the long run.

The parts are much more constrained by their TDP budgets because of where they are used. Doubling up from RV770 to Cypress corresponds to a greater TDP, a luxury that is not afforded in notebooks or many OEM systems. When it comes to the entry level a generational leap in features often means that the transistor budget is spent there for similar TDP ranges; its easier to get performance leaps intra-generational because the featureset changes are often less radical (i.e. RV610-->RV620-->RV710).

 

[Dave Baumann]

and of those how many are actually USABLE for such a low end part.. bitstream audio and eyefinity. Of which, Eyefinity is almost nil use for HTPC, leaving bitstream support.. which would be negated seeing as the 5400 series has castrated Avivo post processing (and atm bugged de-interlacing),.. enough so that anyone interested in the audio feature would pass over the 5400 series if they wanted decent HD representation.

It seems to be your impression that HTPC is the sole purpose of such a product?

Features like Eyefinity will be used very much in across the range of products that the chip is designed for. Likewise, dependant on the systems they are being put in, the other features can and will bring benefits - these are not designed to be put in Phenom X4 or Core i7 systems.

As for AVIVO, as pointed out, there's nothing worse than with prvious generations and the level of functionality offered scales with bandwidth,

 

[Blazkowicz]

it might get included in netbooks, similar to radeon 3410 or ION2.

AMD's underclocked 65nm sempron + Cedar would make a fine, high performance one. (or a fine "CULV" as it seems to be the new term for faster netbooks). though a bit on the hot running side.
wondering if we'll get to see DDR3 on a 32bit bus, as an alternative to 64bit DDR2

 

[Silent_Buddha]

Dave,

Are there any plans to release a 256 meg version of the 5450? I don't need a lot of memory on a budget HTPC card and would prefer less memory for slightly lower power useage... Not to mention being cheaper also.

Regards,
SB

 

[Blazkowicz]

maybe 256MB versions aren't done because they're a cost in having one more SKU in inventories and retail.

an alternative would be using video memory for non graphics use, such as disk cache. a linux hack allows this, especially on PS3 where you have little main memory and as much video memory.

It would have been nice if Microsoft had made such operation part of WDDM, then OS and vendor support would be there ; if you go windows and cheap for a HTPC, 1GB system ram + 512MB video ram would feel nice.

 

[hkultala]

Dave,

Are there any plans to release a 256 meg version of the 5450? I don't need a lot of memory on a budget HTPC card and would prefer less memory for slightly lower power useage... Not to mention being cheaper also.

Regards,
SB

hmm, 256 MB memory on 64-bit memory bus means two 32-bit 1 Gbit memory chips.
And 512 MB memory on 64-bit memory bus means two 32-bit 2-Gbit memory chips.

how much cheaper are gigabit chips compared to 2-gigabit chips, and do they consume less power?
(or can they actually consume more power if they are manufactured on older process?)

Price of gigabit chip is about $2 ?
Does 2-gigabit cost about $4, or more or less?

Assuming 2-gigabit chip costs $4, it would mean 256 MB version of card would be $4 cheaper to manufacture than 512MB version

 

[GZ007]

The parts are much more constrained by their TDP budgets because of where they are used. Doubling up from RV770 to Cypress corresponds to a greater TDP, a luxury that is not afforded in notebooks or many OEM systems. When it comes to the entry level a generational leap in features often means that the transistor budget is spent there for similar TDP ranges; its easier to get performance leaps intra-generational because the featureset changes are often less radical (i.e. RV610-->RV620-->RV710).

In AnandTech test the 5450 load temperatures and load power are way below 4550 so there was clearly a headroom in TDP. I wasnt saying about doubling the ROPs or TMUs just ALU performance.
This time the shader rate from high end to low end is 1600:80 while last generation it was 800:80. Thats 20 times difference against 10. Its degenaration not evolution i say.
Now nvidia wont need to make faster low-end to beat the increased performance from ati card too so the performance level is stalling again like the last few years the midrange cards from nvidia rebrabnding.(the 5700 and lower cards performance targets were clearly just against the weak nvidia cards and are dissapointment if u upgrade from 4k series )
3d cards were/are/will be always about performance. If someone can play at playable fps with dx9 and dx11 is unplayable than he will not bother with dx11. For perfect hpc there is the integrated graphic segment.

 

[Silent_Buddha]

In AnandTech test the 5450 load temperatures and load power are way below 4550 so there was clearly a headroom in TDP. I wasnt saying about doubling the ROPs or TMUs just ALU performance.
This time the shader rate from high end to low end is 1600:80 while last generation it was 800:80. Thats 20 times difference against 10. Its degenaration not evolution i say.
Now nvidia wont need to make faster low-end to beat the increased performance from ati card too so the performance level is stalling again like the last few years the midrange cards from nvidia rebrabnding.(the 5700 and lower cards performance targets were clearly just against the weak nvidia cards and are dissapointment if u upgrade from 4k series )
3d cards were/are/will be always about performance. If someone can play at playable fps with dx9 and dx11 is unplayable than he will not bother with dx11. For perfect hpc there is the integrated graphic segment.

What the heck are you babbling about?

The 57xx series is roughly the performance of the 48xx series and clearly faster than the 47xx series.

The 56xx series is clearly faster than the 46xx series.

So they have not only increased the speed in each segment of the line, they have added more features.

The 54xx appears bad because everyone is comparing it to the 45xx series as there was no 44xx series. I'd imagine the 55xx series should be faster than the 45xx series.

Regards,
SB

 

[rpg.314]

What the heck are you babbling about?

The 57xx series is roughly the performance of the 48xx series and clearly faster than the 47xx series.

The 56xx series is clearly faster than the 46xx series.

So they have not only increased the speed in each segment of the line, they have added more features.

The 54xx appears bad because everyone is comparing it to the 45xx series as there was no 44xx series. I'd imagine the 55xx series should be faster than the 45xx series.

Regards,
SB

Somewhat OT: Considering that LIano is supposed to have 240 alu's, this chip is probably the second last of it's breed.

Also since the next refresh of lineup is supposed to come in 2010H2, I wonder whether it will have replacements for Cedar (and redwood too?) at all, since all of it is coming awfully close to the launch of LIano.

Since Cedar has 400 alu's (not very far from LIano), I am guessing that <$100 market for discrete gpu's is toast in a couple of years. Where will all the poor G92 rebrands go then?

 

[hoom]

Maybe they should have called it 52XX series?

 

[Dave Baumann]

In AnandTech test the 5450 load temperatures and load power are way below 4550 so there was clearly a headroom in TDP. I wasnt saying about doubling the ROPs or TMUs just ALU performance.

A point comparisn of two different products do not tell you much about actual TDP differences across a product line - you have no clue of the leakage difference of the two chips tested.

 

[trinibwoy]

The 54xx appears bad because everyone is comparing it to the 45xx series as there was no 44xx series. I'd imagine the 55xx series should be faster than the 45xx series.

Regards,
SB

It looks like people are comparing specs, not brand names.