Is the 5770 confirmed for 192b or is it 128b? If its the latter, I dont see it being competitive with the 4890 or GTX275s.EDIT: Damn, I'm late.
5750 castrated in terms of SIMDs? If 5770 is 850Mhz and still is in compact form, hello 9600GT redux.
AMD: R8xx Speculation
- Thread starter Shtal
- Start date
So, anyone want to go and try some estimation on how many SIMDs the Cypress chip could have based on the diesize, given the Juniper has 14 SIMDs, 56 TUs and 2x64bit mem controllers? ROP-count could be either 16 or 24
edit:
edit:
128bit due the amount of memchips
Even if it *is* using 128b, you're most likely seeing 850/1300 for the core.
If 5750 ==< 4870 (slightly), 5770 ==< 4890 (memory b/w was ample for the 4870/90s, so engine clock was the one that really mattered 750-850 vs 700-850 with potential for more SIMD groups?).
If it's slightly slower than the 4870/90, but uses MUCH less power (5850 uses ~50W less than 4890 in norm conditions, what about this? 90W?) and can be single-slotted, it would be much more desirable for the potential customers of this chip (G92/RV670 and earlier, OEMs, Mobile-!!!) than say, the extra 10% of performance
BTW AMD prices in terms of performance relative to competitors so we can approximate I guess. I don't think they overprice midrange products even with DX11 and Eyefinity, but I could be wrong...
If 5750 ==< 4870 (slightly), 5770 ==< 4890 (memory b/w was ample for the 4870/90s, so engine clock was the one that really mattered 750-850 vs 700-850 with potential for more SIMD groups?).
If it's slightly slower than the 4870/90, but uses MUCH less power (5850 uses ~50W less than 4890 in norm conditions, what about this? 90W?) and can be single-slotted, it would be much more desirable for the potential customers of this chip (G92/RV670 and earlier, OEMs, Mobile-!!!) than say, the extra 10% of performance
BTW AMD prices in terms of performance relative to competitors so we can approximate I guess. I don't think they overprice midrange products even with DX11 and Eyefinity, but I could be wrong...
For HD5750, which would means 83 Gb/s of BW.
To match the GTX285, the HD5770 has to have a 192bit MC (124 Gb/s BW).
HD5770 seems to be a weapon of mass destruction: small die-size, not power hungry, cheap, Dx11 and probably GTX285 performance class. Oh God, that must hurt.
Silent_Buddha
Legend
Well, 5850 is anywhere from about the same as GTX 285 to 20%+ faster than GTX 285.
If I were to pull a random guess out of my backside. I'd say 5770 should be anywhere from 20-30% slower to about equal to GTX 285.
So basically the GTX 285 will be sandwiched between the 5770 and 5850.
Regards,
SB
If I were to pull a random guess out of my backside. I'd say 5770 should be anywhere from 20-30% slower to about equal to GTX 285.
So basically the GTX 285 will be sandwiched between the 5770 and 5850.
Regards,
SB
Yep.
Man... looking at the math, 192bit woulda been killer w/ 24ROPs. With 4.8ghz GDDR5 it would have the same bandwidth as a 4870, 24ROPs @ 800mhz puts it at a 30% disadvantage to a 5850, 1120SPs @ 800mhz 1.8TFlops = 13.5% disadvantage vs 5850.
Priced @ $199 it would be a decent deal, ~75% of the price of a 5850 and performing @ about 70-85% of the 5850.
I'm getting, with very rough math mind you, ~24-26SIMDs.
Last edited by a moderator:
What does the 128bit bus matter if it still beats 4890 by some margin too?
4870 and 4890 weren't bandwidth bound, so it might not be that much limiting factor for the HD5700 series with faster clocked memory parts even though the buswidth is halved
Well of course it could be 256bit, but I doubt that.
And while 4850 was bandwidth bound, these chips will have more bandwidth with 128bit bus than 4850 had with it's 256bit bus.
Actually, as the owner of a pair of 4850's, I still get better performance out of GPU clock increases than MEM clock increases.
My ram will do almost 1150Mhz (up from the stock 1000-ish) if I leave the GPU clock at stock. OR if I leave the mem stock, the GPU will do right at 700Mhz (up from the stock 625.
Performance in all the games I play (Oblivion, Fallout 3, FC2, HL2, Crysis, World at War, the various GTA's, etc) all perform measurably and notably better with the extra GPU clock.
Realistically, the 4870 has what, twice the memory bandwidth? But what's the actual speed increase from a 4850 to a 4870?
Well, from bandwidth alone... http://www.behardware.com/articles/726-4/radeon-hd-4800-s-gddr5-an-advantage.html
^^ dunno, maybe it's the resolution then? I play at 1680x1050, although I do use AA wherever feasible. I did a considerable amount of benchmarking when I built the rig, and GPU clocks always won out in my config.
Hmm I thought the 4890 did about 60 FPS in Perlin Noise (800SP/850Mhz).
However, the 5870 does 158 FPS (1600SP/850Mhz). If we assume 12 simd for 5750 (and simd width 16), that's still 960SP - at the reported 750Mhz that should be 84 fps (158 / 850 / 1600 * 750 * 960). So this result is very disappointing, if that's really 960SP at 750Mhz - it is even worse (in percentage of alu peak rate) than what rv770 did, and rv870 improved (in this test) quite a bit there. Would be near perfect scaling if the simds only have width 12 (hence 720 simd for the 12 simd part) but noone seems to believe that
. Or can the Perlin Noise test be bandwidth limited?
Hmm I thought the 4890 did about 60 FPS in Perlin Noise (800SP/850Mhz).
However, the 5870 does 158 FPS (1600SP/850Mhz). If we assume 12 simd for 5750 (and simd width 16), that's still 960SP - at the reported 750Mhz that should be 84 fps (158 / 850 / 1600 * 750 * 960). So this result is very disappointing, if that's really 960SP at 750Mhz - it is even worse (in percentage of alu peak rate) than what rv770 did, and rv870 improved (in this test) quite a bit there. Would be near perfect scaling if the simds only have width 12 (hence 720 simd for the 12 simd part) but noone seems to believe that
5750 is supposed to have 10 SIMDs aka 800 SPs enabled, while 5770 should have all 14 SIMDs aka 1120 SPs
4 simds disabled? That's a lot of disabled units. Even in such a config assuming scaling according to peak alu rate of HD 5870 at 750Mhz it should still hit ~70fps, so still quite bad (though at least now would beat rv770 in efficiency wrt to peak alu rate).
I note that the 5750 benches some 0-5% better than my personal HD4770, (or 20% better than a stock HD4770), and still requires an extra 6 pin connector. While AMD seems to have improved idle power draw greatly (halleluja!), circumstantial evidence implies that power draw at work hasn't improved over the already 40nm RV740. If this is borne out by tests, it would be an expected, but still slightly disappointing result.
As an aside, in my experience the HD4770 responds very favourably to higher GDDR5 clocks, both in benchmarks and particularly in real world gaming. I can't see other than that the speculated 1120SPs of the HD5770 would be quite bandwidth limited with a 128bit memory interface, for most titles you'll find on retailer shelves during the life time of the part. With a wider bus however...
As an aside, in my experience the HD4770 responds very favourably to higher GDDR5 clocks, both in benchmarks and particularly in real world gaming. I can't see other than that the speculated 1120SPs of the HD5770 would be quite bandwidth limited with a 128bit memory interface, for most titles you'll find on retailer shelves during the life time of the part. With a wider bus however...
A 14 SIMDs and 128-Bit RV700 with Cypress' transitor density would be around 160mm².
(Cypress: 6.45 million transitors per mm², RV740 826 million transitors, a RV770 SIMD = 10mm²@55nm, RV770 to Cypress density factor 1.75)
So they put in the left 25mm² D3D11, another 64-Bit-MC/8ROPs and additional caches?
(Cypress: 6.45 million transitors per mm², RV740 826 million transitors, a RV770 SIMD = 10mm²@55nm, RV770 to Cypress density factor 1.75)
So they put in the left 25mm² D3D11, another 64-Bit-MC/8ROPs and additional caches?
Similar threads
