The ATI Radeon 5XXX Thread

[Jamesrt2004]

If that's the pic of a single HD5870, I can't imagine the lenght of a hypothetical HD5870x2, maybe 12"?

ahh that cards only very slightly if not the same size as 4870 so i expect it do be the same as the 4870x2 atm...
you can see if you get a pic of a 4870 from google, from a similar angle, line them up under each other and you can see =)

It might but i think it would be worth it, they are going to charge $399 for a chip with 256bit memory interface, doesnt look too good. To have 2GB of ram without more bandwidth might be a bottleneck.

no.. no bottleneck hence why the 4890 2gb cards are fine.. honestly you could up the clock 200mhz on the memory and no difference in performance maybe 1/2 % the performance was all in the core clocks etc..

is it me or is that vent/exhaust for the card a little small?

Nah it's fine the vapour card iirc, uses half slot vent, the difference it makes is neglible so its all good

[Origin_Unknown]

Both, not sure if 2GB will be at available at launch though.

the 2gb would interest me greatly!

[spursindonesia]

Put 5.5 Gbps GDDR5 chip on the HD 5870 variant, clock it at 1.25 GHz, everybody should be a happy camper, especially if there would be 2 GB VRAM capacity equipped variant. HD 5850 can have 4 Gbps chip clocked @800-900 MHz.

[Calmatory]

For radial fans(which can generate alot more pressure than conventional axial fans) the small exhaust will not be a problem as long as the air is well directed towards the exhaust.

[Chumbucket843]

Umm... because a 4770 has less bandwidth than a 4830 and 4850...

it has less shaders so its bottlenecked by calculation.

[Sh1tyMcGee]

oks i got confirmation of specs for the new cards

5870 = 1600SP's, clocks at 850/1200
5850 = 1440SP's clocks at 700/1000

Sounds like the 5870 should perform like 2 4890's in crossfire.

[GoldenTiger]

oks i got confirmation of specs for the new cards

5870 = 1600SP's, clocks at 850/1200
5850 = 1440SP's clocks at 700/1000

Sounds like the 5870 should perform like 2 4890's in crossfire.

http://alienbabeltech.com/main/?p=11135 claims those specs, yeah. It would be a lot faster though due to no crossfire scaling issues if that's the case.

[perkam]

oks i got confirmation of specs for the new cards

5870 = 1600SP's, clocks at 850/1200
5850 = 1440SP's clocks at 700/1000

Sounds like the 5870 should perform like 2 4890's in crossfire.

Ummm...at $399, it is double the price of a 4890 at $199 and thus will HAVE to perform like 2 4890s in crossfire to justify the price

Perkam

[Manicdan]

Ummm...at $399, it is double the price of a 4890 at $199 and thus will HAVE to perform like 2 4890s in crossfire to justify the price

Perkam

kinda has too, but it could be like cpus and charge 5x the price for just 20% more performance

im not impressed with the ram speeds, we had 800mhz GDDR5 18 months ago, and now were only at 1200mhz

[Vit^pr0n]

kinda has too, but it could be like cpus and charge 5x the price for just 20% more performance

im not impressed with the ram speeds, we had 800mhz GDDR5 18 months ago, and now were only at 1200mhz

Yea, a 50% increase is pretty bad imo.

[003]

Ummm...at $399, it is double the price of a 4890 at $199 and thus will HAVE to perform like 2 4890s in crossfire to justify the price

Perkam

Current reports indicate the 5870 will be slightly slower than the 4870X2.

[cegras]

Ummm...at $399, it is double the price of a 4890 at $199 and thus will HAVE to perform like 2 4890s in crossfire to justify the price

Perkam

No, you forgot to include many factors in your analysis like consolidation of PCBs, etc, etc - these savings can be passed on to consumers. Was the GT200 linearly better than the G92 with regards to price/performance?

[labs23]

Current reports indicate the 5870 will be slightly slower than the 4870X2.

Not looking good.
But I doubt that, based on the pricing.

[SimBy]

Current reports indicate the 5870 will be slightly slower than the 4870X2.

Let me think, it's gonna cost more and be slightly slower. Excellent logic bro.

[cegras]

Current reports indicate the 5870 will be slightly slower than the 4870X2.

Any references? Or ...

[Jowy Atreides]

Current reports indicate the 5870 will be slightly slower than the 4870X2.

Current reports need rewritten

[hennyo]

Not looking good.
But I doubt that, based on the pricing.

This most likely means that the drivers this was tested on currently SUCK, hardly proof of how it will work in the future.

[TheBlueChanell]

Any references? Or ...

If the 1600sp, 80 TMU and 20 ROP?? rumour is true I seriously believe that it will be at least 10-20% faster than a 4870x2 and probably even more. This is a single chip card thus you aren't going to have scaling issues or any of the caveats that come along with any multi-chip solution.

It would be as if the 4870x2 had 100% perfect scaling plus some additional horse power.

If the 4770 is any testement to the overclocking capabilities of the 58**'s we are in for quite a treat and I if I were a betting man I'd bet TSMC's 40nm process has matured quite nicely over the past few months.

This is all based on speculation and personal opinion but that is how I feel.

[highoctane]

While there is more muscle in the gpu if its still a 256bit memory bus then bandwidth is pretty much what a single 4870/90 is, half of an x2. So even if it performs on par with an x2 thats a pretty admirable gain.

[zerazax]

One more day for some leaks!

[hurleybird]

if its still a 256bit memory bus then bandwidth is pretty much what a single 4870/90 is, half of an x2.

That's not how it works. For all practical purposes, the effective bandwidth of the X2 is the same as if there were only a single card. In the future a NUMA system might come into play where both memory controllers are in communication and then you could (theoretically, assuming the link between memory controllers is fast enough) have 2x the bandwidth in a practical sense.

[jaredpace]

This is addressing comments about memory bandwidth. The hd4870 has 115gb/sec gddr5 memory, if you take the 4850's limited 64gb/sec memory bandwidth and apply it to the same clocked 800 shaders of the 4870 (750mhz ie an overclocked 4850) you only lose about 5% overall performance compared to the stock hd4870. The 5870 with 153+gb/sec gddr5 gives plenty of bandwidth for 1600 shaders considering 800 shaders at 64gb/sec gddr3 nets you 95% of 800 shaders at 115gb/sec gddr5. The bottleneck must begin somewhere below ~70gb/sec ddr3/5 for 800 shaders in ati's architecture. This would mean that 1600 shaders only needs greater than 140+gb/sec memory bandwidth to operate at it's fullest potential. I also found this graph on TPU where wizzard was disabling SIMD units in rv770 for his 4830 review to see how many FPS he could get with 80 to 800 shaders in increments of 80. Seems like as he enables shaders the fps increases linearly, which is good news for going from 800 to 1600.

http://www.techpowerup.com/articles/other/155

[annihilat0r]

jaredpace: I find that surprising, but how does that explain good performance gains from OC'ing DDR5 speeds in 4870, if it's so much over the bottleneck?

[Final8ty]

Latency.

[hurleybird]

jaredpace: I find that surprising, but how does that explain good performance gains from OC'ing DDR5 speeds in 4870, if it's so much over the bottleneck?

What good performance gains?

This is addressing comments about memory bandwidth. The hd4870 has 115gb/sec gddr5 memory, if you take the 4850's limited 64gb/sec memory bandwidth and apply it to the same clocked 800 shaders of the 4870 (750mhz ie an overclocked 4850) you only lose about 5% overall performance compared to the stock hd4870.

how much of that is due to bandwidth saturation, and how much of that is due to latency? All else being equal, GDDR3 outperforms GDDR5 at equal Hz.


In any case, we'll just need to wait and see. After we have the cards in hand it won't be hard to see if core and memory performance is balanced. Maybe it's imbalanced because AMD is planning on debuting faster models with faster memory. Or maybe we'll finally see some kind of shared memory on X2, and the bus AMD is using between memory controllers would be oversaturated (as well as having less space along the edge of the chip for it to fit on) with a larger memory interface. Maybe they've done some tweaking so they make better use of bandwidth? At the very least the enhancements from RV740's memory controller will be brought over. For all we know RV870 could use tile rendering! One more day.