AMD: R8xx Speculation

[DavidGraham]

About HD5830 ..

Guys .. about HD5830 , the card has identical ROPs/frequency specifications to HD4890 , with 50 MHz less Core frequency and 100MHz more RAM frequency .

So essentially it is the same as HD4890 except for 16 more Texture Units and 320 more shader processors .

As such it should perform faster than HD4890 ,or at least have similar performance .. yet it is STILL slower (8% according to Anand).. how does that happen ? is there a logical explanation for this ?

 

[Mindfury]

Guys .. about HD5830 , the card has identical ROPs/frequency specifications to HD4890 , with 50 MHz less Core frequency and 100MHz more RAM frequency .

So essentially it is the same as HD4890 except for 16 more Texture Units and 320 more shader processors .

As such it should perform faster than HD4890 ,or at least have similar performance .. yet it is STILL slower (8% according to Anand).. how does that happen ? is there a logical explanation for this ?

Rop bottleneck....5830 is suck at AA performance..

 

[Alexko]

It's not enough to explain why the 4890 is sometimes 10% faster when its ROPs are only clocked 6.25% higher than the 5830's, especially since it has much less shader or texturing power, and less memory bandwidth.

 

[mczak]

Meh what a disappointment. Performance much closer to 5770 rather than 5850, getting beaten for some odd reason by HD4890, and price closer to 5850 rather than 5770. Not to mention the power draw (60% higher than 5770, more than 5850, similar to 4890).
Now I understand AMD wants to sell the garbage bin, but as a customer I see zero reason why you'd want to buy that card over a 5770 or 5850 (depending on your budget) it just doesn't make sense at that price (unless you want the cheapest card with dual precision or something).
Though I'm curios about that pic here:
http://ht4u.net/reviews/2010/amd_ati_radeon_hd_5830/index2.php
This says 2 rop partitions (at 2 quad-rops) are disabled. Is this really true? It would mean there's only half the L2 cache. I'd have expected every second quad-rop (from left to right) be disabled instead, leaving all the L2 cache intact.

 

[DavidGraham]

It would mean there's only half the L2 cache. I'd have expected every second quad-rop (from left to right) be disabled instead, leaving all the L2 cache intact.

Could that be the reason for lower than HD4890 performance?

 

[mczak]

Could that be the reason for lower than HD4890 performance?

No, Dave confirmed in another thread that it has all the L2 cache (and hence that diagram must be bogus).
So being slower than HD4890 seems to be due to clock mostly, though I really don't understand why it is sometimes slower than even what the clock difference suggests. And even if it's "only" slower by the amount of clock difference, that is imho still quite strange, as surely even if you'd think it's limited quite a bit by the low rop count you'd still expect the much higher alu/tex throughput to make some difference...
Maybe it's the same reason why HD5870 seems to do quite a bit worse in practice compared to its theoretical specs vs. HD5570 (that is, it's nowhere near twice as fast). Though for that one speculation was insufficient internal bandwidth, which compared to HD4890 doesn't make sense as that should have the same bandwidth (at the same clock). Unless the dual rasterizers somehow make things more inefficient.

 

[Squilliam]

Could it be that the HD 5830 for whatever reason is running as an effectively 800SP card? Perhaps this is due to how the layout works and the loss of the 16rops from one side effectively makes the card an 800SP card for reasons unknown, bandwidth? Drivers? Architecture?

Edit: Drivers it was. More up to date Catalyst 10.2 drivers puts the 5830 where it ought to be. Much better looking card now we have a true picture of its performance.

 

[Alexko]

Could it be that the HD 5830 for whatever reason is running as an effectively 800SP card? Perhaps this is due to how the layout works and the loss of the 16rops from one side effectively makes the card an 800SP card for reasons unknown, bandwidth? Drivers? Architecture?

Edit: Drivers it was. More up to date Catalyst 10.2 drivers puts the 5830 where it ought to be. Much better looking card now we have a true picture of its performance.

How so? Here's what Anandtech says about the divers: "The drivers AMD shipped with the Radeon HD 5830 are version 8.703 RC2, dated February 11th".

Yet their review still shows the 5830 as clearly slower than the 4890.

 

[Jawed]

Unless the dual rasterizers somehow make things more inefficient.

L1s in HD5830 will be "starved" in comparison with L1s in HD4890 as total L2->L1 bandwidth is spread more thinly in HD5830 (rate is no better, per clock, than HD4890, but there are 4 more SIMDs).

Also, of course, Cypress ALUs are doing interpolation.

Jawed

 

[FrameBuffer]

Well, the performance is definitely disappointing given the suggested price, IMO ATi has failed to live up to the high standards it setup with the HD4000 series. It's about on par with the much older 4870/4890, costs substantially more, though in the areas it could do well in due to newer features (bitstreaming, DC, OCL, Eyefinity) it just doesn't "fit". In the HTPC crowd such a card (big, hot, loud) it would never find a home, it's size is severely prohibitive where half height cards are a must. In Eyefinity the 5830 seems like it would barely meet the minimums to power 3 x 1080p displays all while having to discard most any of the image enhancing effects (lower AA, AF etc). With DC and OCL, the castration of the amount of shaders will again most likely be a contributing factor to poor performance. In Fluid 3D it's about 10-15% faster then the nearly $100 cheaper 5770. In Julia 4D it's about 8-14% faster then the 5770 and only in Waves3D does it come close the the 5850 (about 8%).

At $200 it'd be a worthy successor but at 20% greater cost, if performance is the main factor the obvious choice is either 4870/4890 or GTX 275 (if you F@H then it's not really a choice.. NV simply wipes the walls with ATI from top to bottom) or pony up the cash for the 5850. If per to price ratio is a concern then the 5770 is a much better choice at nearly $80 less MSRP. IMO the 5830's price does not justify it's poor performance in relation to it's price.. some might say "well it gives 80% performance of a 5850 at 80% the price (actual vs msrp)".. to that I simply would counter that the entire 5800 line is priced too high, mainly due to the lack of competition. Just look at the actual cost of the 5850.. was in Microcenter this weekend (where the 275 was $198) the the XFX 3850 was tagged at $379.99 (yes thats a three) but "on sale" for $315 iirc.

Lastly I see we are back to the age of paper launches.. sad. The more time passes the more ATi sounds and looks like NV of old.

 

[mczak]

L1s in HD5830 will be "starved" in comparison with L1s in HD4890 as total L2->L1 bandwidth is spread more thinly in HD5830 (rate is no better, per clock, than HD4890, but there are 4 more SIMDs).

Yes, but the absolute L2 bandwidth is still the same (per clock), so you'd still only expect the HD5830 to be the same performance if it's 100% limited by L2 bandwidth (which seems unlikely as that would mean there's simply no point in more than 10 simds...).

Also, of course, Cypress ALUs are doing interpolation.

Clearly, the HD5830 should more than compensate for that by its much higher alu throughput. Unless you're hinting at the additional delays or something like that - but Juniper doesn't show any disadvantage there really so this doesn't make much sense.

 

[Dave Baumann]

Lastly I see we are back to the age of paper launches.. sad. The more time passes the more ATi sounds and looks like NV of old.

I have partners that want to ship ASAP (and product is flowing into the channel right now), I've just come off the back of Chinese new year and I have two major tradeshows / confernces in the next couple of weeks, one of which a large portion of the European journalists will be busy with. I don't see how this constitutes your comment, especially off the back of 10 products that have been launched in the the last 6 months.

 

[Kaotik]

Lastly I see we are back to the age of paper launches.. sad. The more time passes the more ATi sounds and looks like NV of old.

These cards are already in stock in some shops, despite what any review might say
http://www.newegg.com/Product/Produ...0&bop=And&ActiveSearchResult=True&Order=PRICE

 

[eastmen]

Could it be that the HD 5830 for whatever reason is running as an effectively 800SP card? Perhaps this is due to how the layout works and the loss of the 16rops from one side effectively makes the card an 800SP card for reasons unknown, bandwidth? Drivers? Architecture?

Edit: Drivers it was. More up to date Catalyst 10.2 drivers puts the 5830 where it ought to be. Much better looking card now we have a true picture of its performance.

http://forum.beyond3d.com/showpost.php?p=1399161&postcount=545

see my post here

But the run down cats 10.3

dirt 2

1920x1200 highest details 4x post processing high. dx 11

5850 39/44
5830 32/36.14
5770 27/32.67

ITs 5/3 faster than a 5770. Yet its 7/7 slower than a 5850

at 2560x1600 highest details 8xaa non dx 11 mind you

5850 42/49.29
5830 30/37.49
5770 26/33.16

4/4 faster than the 5770 yet 12/11 slower than the 5830.

Dragon age

dragon age

5850 41/57.09
5830 29/42.26
5770 25/36.24

I mean look at the performance hit here. The 5850 is 12/14.83 faster. The 5770 is just 4/6 slower. Once again the price diffrences between these cards make no sense.



Currently 1 gig 5770s go as low as $150 on newegg. This card at $240 is priced much closer to the 5850 than the 5770. ITs basicly $60 less than the 5830 yet performs closer to the $90 cheaper 5770.

I'm also sure that once the 5830 comes out the 5770s will drop a little more.

 

[FrameBuffer]

These cards are already in stock in some shops, despite what any review might say
http://www.newegg.com/Product/Produ...0&bop=And&ActiveSearchResult=True&Order=PRICE

weird wasn't there when I posted earlier ..

 

[Babel-17]

weird wasn't there when I posted earlier ..

Yeah, they need to update as it currently can't be found by searching the 5xxx series.

http://www.newegg.com/Product/Produ...10380048 106793261&name=Radeon HD 5000 series

 

[GZ007]

I have partners that want to ship ASAP (and product is flowing into the channel right now), I've just come off the back of Chinese new year and I have two major tradeshows / confernces in the next couple of weeks, one of which a large portion of the European journalists will be busy with. I don't see how this constitutes your comment, especially off the back of 10 products that have been launched in the the last 6 months.

Will there be custom pcb cards too ? The 5870 pcb chosen by almost all AIB partners doesnt seems to fit the card to much.
In the pictures from the different cards this XFX looks quite strange with all the 5870 PCB cards. Without refference design why did they choose the 5870 PCB instead the 5850 PCB anyway ?

 

[Dave Baumann]

Will there be custom pcb cards too ? The 5870 pcb chosen by almost all AIB partners doesnt seems to fit the card to much.

If you look at the PCB's most of the partners are using they are already smaller than the previous reference HD 5870's (and this is smallor, or about the same size than GTX 260 that is playing in a much lower space that HD 5830 is). PCB sizes are generally driven by the power and thermal requirements and as pointed out the HD 5830 has a higher TDP than HD 5850.

 

[Alexko]

Dave, could you give us a few hints as to why the 5830 is often slower than the 4890? I still haven't seen anything convincing.

 

[Kaotik]

Will there be custom pcb cards too ? The 5870 pcb chosen by almost all AIB partners doesnt seems to fit the card to much.
In the pictures from the different cards this XFX looks quite strange with all the 5870 PCB cards. Without refference design why did they choose the 5870 PCB instead the 5850 PCB anyway ?

Well none are using the reference design (5870), and that XFX is the only one I've seen with 5750 PCB, they've been custom PCBs from what I can tell