NVIDIA Fermi: Architecture discussion
- Thread starter Rys
- Start date
Nah! As others have mentioned already, the 2 GB GTX 380 isn't possible, given what we know about Fermi. If it said 1.5 GBs, then it could be a worthy leaked benchmark
I think it's been mentioned more than once though, that NVIDIA should finally talk about Fermi's "graphics bits" this month, so it's wait and see and maybe some of those "worthy" leaked benchmarks appear too.
jimmyjames123
Regular
Something doesn't quite add up. While the NVIDIA slides look professional, there is no "NVIDIA confidential - NDA xxxx" on them, and the benchmark data matches perfectly with the Tom's Hardware Guide review data. Since when did NVIDIA slides use Tom's data?
Edit: The GTX 295 numbers for Resident Evil and Crysis in the nvnews post look significantly different than what was in Tom's HD 5970 review. All the other data in the nvnews post matches almost spot on with Tom's HD 5970 review.
Hell, I'll just wait for more leaked info later in the month or early next year
One thing to note: the NVIDIA slides seem to only show scenarios where performance improvement from HD 5870 to HD 5970 is up to ~ 50%. I'm sure there are scenarios where the performance difference between these two cards is significantly more (or less too I suppose).
Edit: The GTX 295 numbers for Resident Evil and Crysis in the nvnews post look significantly different than what was in Tom's HD 5970 review. All the other data in the nvnews post matches almost spot on with Tom's HD 5970 review.
Hell, I'll just wait for more leaked info later in the month or early next year
One thing to note: the NVIDIA slides seem to only show scenarios where performance improvement from HD 5870 to HD 5970 is up to ~ 50%. I'm sure there are scenarios where the performance difference between these two cards is significantly more (or less too I suppose).
Last edited by a moderator:
Something doesn't quite add up. While the NVIDIA slides look professional, there is no "NVIDIA confidential - NDA xxxx" on them, and the benchmark data matches perfectly with the Tom's Hardware Guide review data. Since when did NVIDIA slides use Tom's data?
Hell, I'll just wait for more leaked info later in the month or early next year
Forget about those. CJ is usually on point with leaked benchmarks, so maybe he will have something for us soon
jimmyjames123 said:
It depends on the load really. In some loads, the lead can be significant. In others, not so much. The 8800 GTX was quite a bit faster than the HD 2900 XT, but in Call of Juarez the XT actually came close or beat the 8800 GTX in some levels and/or settings.
The GTX 280 was also faster than the HD 4870 (although not as much as the 8800 GTX was compared to the HD 2900 XT), yet in games like GRID, the HD 4870 dominated the GTX 280.
Jawed and others, have quite a bit of data on that subject, where the specificity of the load itself (like number of shaders, etc), helps determine the architectural advantages/disadvantages one has over the other. I don't have the specific links for those threads, but you can look them up in this very same thread I believe.
Slightly over 700MHz afaik with a ~2.2x ALU:TMU ratio under normal voltage rates. If they hypothetically would raise the voltage slightly a higher hot clock could be possible. It's senseless at this point to speculate on final frequencies until they get A3 done. IMHO they will then decide what the best level of frequencies/voltage will be considering that they want to keep power consumption roughly on GTX285 levels.
jimmyjames123
Regular
All very good points Silus. The one big advantage for NVIDIA is that, as long as they can get close to HD 5970 performance with something like GTX 360/380, then it's a win for them, due to the latter not needing SLI/Crossfire profiles and not having the issues of a multi-gpu setup. If the GTX 380 wins the performance crown outright, then it's a huge win (at least for NVIDIA and it's profit margins).
why thats what thier targets are. The benchmarks might be fake but someone knows what nV was going for.
I don't think that will happen on a regular basis - maybe in games where CF is scaling more than just a bit sub-par, but I'm quite sure, that this won't be the rule to the exception.
jimmyjames123
Regular
I don't think that will happen on a regular basis - maybe in games where CF is scaling more than just a bit sub-par, but I'm quite sure, that this won't be the rule to the exception.
I agree, and that is probably why NVIDIA [allegedly] has a dual gpu Geforce card in the works.
For A2 ? All I read about 650-700 Mhz GPU clock was related with the target clocks for the final chip (which seems to be A3). If A2 was already there, then they are trying to get to even higher frequencies, but probably on the Stream Processors.
I wouldn't put much faith on a single GTX 380 beating a HD 5970 every time, but if it gets close enough, it's obviously good for NVIDIA, since they can keep the price of the GTX 380 very close to the price of the HD 5970 and the HD is much more expensive to produce, which puts pressure on AMD and thus creates a price war, where we all win
I see. Where was it reported/rumored ?
I don't think that there's any question that a GTX 395 will exist. The question is really how close the GTX 380 can get to the HD 5970, so that the GTX 395 doesn't really require two full Fermi chips and maybe just two GTX 360 (with 384 SPs for example) to beat the HD 5970.
Cookie Monster
Newcomer
If the GTX360 had 384SP, isnt that too much of a gap from the GTX380 (512SP)? Wouldn't 448 be the logical number for the next model down from their flagship card?
Yes, it's all up in the air. No one knows for sure yet.
Bouncing Zabaglione Bros.
Legend
I don't believe the war will be won over the top end battle, but in the performance mid-range battle. Just matching 5970 with Fermi whilst being larger/more expensive to produce for the top end, low-volume SKU is not a viable approach to the whole market. It's especially the case if AMD have a refresh product ready to beat Fermi over the head.
I think that's why Nvidia is looking to at HPC as a new market to skirt that fight, and is gambling on looking further down the line to a future of more programmable GPUs.
Hmm nope.
If Fermi matches the 5970:
1) Fermi will be a less expensive card to produce then the 5970.
2) Cut down versions of Fermi, will be much more cost effective for midrange cards.
3) we won't see huge performance changes from updated AMD cards (higher clocks, a refresh of products they have now)
4) HPC is a whole new area nV is looking at, the potential is higher then the gaming market, much higher, but they aren't leaving gaming to chase something that is just now starting up, key word potential, no company in their right mind would leave thier bread and butter to go after potentials, they stick with thier guns and go for the potential.
