NV's late, but do ppl really think it'll be slow, or just not as profitable as it could be given its timing? If it's the latter, can't we just let NV worry about that--or at least keep that in another thread? The idea that it'll be NV30 redux seems like a huge reach. I haven't been following things too closely, but I thought GT200's main flaw was size, not performance.
NVIDIA Fermi: Architecture discussion
- Thread starter Rys
- Start date
Pete hit the nail in the head. It's clear by now (and has been for a while) that the product is late, very late. But where do people get their idea that performance will be bad? Until someone gets their hands on real hardware and runs some tests all we have is some (biased) speculation.
The emphasis on HPC does not mean that Fermi based cards will not be good for gamers or that Nvidia is moving away from the gaming market. That would ultimately stupid considering the volume.
So I hope they stop wasting time with stupid comics about other companies and work hard to deliver a solid GPU before Q2.
The emphasis on HPC does not mean that Fermi based cards will not be good for gamers or that Nvidia is moving away from the gaming market. That would ultimately stupid considering the volume.
So I hope they stop wasting time with stupid comics about other companies and work hard to deliver a solid GPU before Q2.
chavvdarrr
Veteran
offtopic
I do believe R600 was ATi's nv30 - it was late, slow and Ati needed major arhitecture changes to get competative (ring bus, 512b memory interface - where are these?)
/offtopic
And I'm pretty sure G100 will be fast enough, but denying that Nv missed its target with time to market is weird. Not to mention I really hate their constant countless renaming of chips. I won't even bother recommending NV card until they start using fair naming scheme again.
I do believe R600 was ATi's nv30 - it was late, slow and Ati needed major arhitecture changes to get competative (ring bus, 512b memory interface - where are these?)
/offtopic
And I'm pretty sure G100 will be fast enough, but denying that Nv missed its target with time to market is weird. Not to mention I really hate their constant countless renaming of chips. I won't even bother recommending NV card until they start using fair naming scheme again.
chavvdarrr: Problem of R600 wasn't its architectual concept (like NV30), but cracked ROPs and unbalanced performance of some parts (plenty of shading power, mediocre INT8 filtering rate and underperforming ROPs - too few Z-ops units). Due to the broken ROPs (HW resolve) the chip wasn't able to use more than ~half of its bandwidth, so advantage of 512bit memory bus wasn't utilized. /my opinion.
GF100 will very likely stand against refreshed RV870, so relative performance compared to current products is more or less irrelevant...
GF100 will very likely stand against refreshed RV870, so relative performance compared to current products is more or less irrelevant...
PSU-failure
Newcomer
In fact, way more than "only one person" are inclined to think Fermi won't be a performance monster...So far we know knowthing about the performance. The uncanny impression comes only from one person.
Anyone who had been fooled by AMD before the launch of R600 and NV before the launch of NV30 simply can't believe their promises without proofs.
And they did not beat them. They took the performance leadership with the 4870x2 from nVidia for 4 months but that was all. Even nVidia's g92 was a very tough chip for the GT200 cards - compare 9800gtx+ to GTX260-192.
I never wrote they were not a competition to nVidia only they needed three generation to beat them - that happend with nv30 -> nv35 -> nv40 -> g70, too.
Thinking it's not the same like knowing.
Last edited by a moderator:
PSU-failure
Newcomer
Sure, but if NV PR know their job, they know promises don't have enough effect when previous ones were lies, so not showing anything is just a confession.Thinking it's not the same like knowing.
Add that to very bad PR BS around HD5800 launch and the clear HPC focus announcement and it tends to confirm that they simply have a flawed product coming.
By the way, you sure do realize HPC market has a severely longer lifecycle, right? So this could be an emergency exit if they lack competitiveness on gamer 3D market... but only if Fermi is GREAT for this market, else it'll be a disaster as AMD on the other hand shortened its GPU lifecycle, thus we can expect strong competition before 2011. I don't know what R900 arch will look like, but expect between 50 to 100 percent performance improvement, leaving Fermi way behind, on every front except perhaps CAD where NV now has some good reputation.
And what are the odds that nv won't have a new chip by 2011 to catch up with r900?
NV survived nv30. ATI survived R600. nv will surely survive fermi, even if it is craptacular. Companies have been known to recover from 1 bad chip. 2 consecutive bad chips however, is another different story.
GT200 had bad efficiency, no doubt. But it wasn't bad enough to fall in the above-mentioned category.
I don't see this. With the small amount of information they gave us, there will be easily faster than GTX285. Look at this: Fermi has 128 TMUs instead of 80, 512 Cores instead of 240, 48 ROPs instead of 32, 384bit+GDDR5 instead of 512bit+GDDR3. And there is the other stuff, which is new over GT200.
nv30 and r600 had a lot more similarities http://www.dict.cc/englisch-deutsch/similarities.html to the older generation than Fermi to GT200.
I don't believe that we will see r900 this year because there is no new process node before q4. And the delay of Fermi means not that this will effect the next generation. Look at nv35/nv40 or r580 or r670. They all came in the expected timeframe.
Ninjaprime
Regular
As I recall, in the RV770 generation, ATI won ~13% of NV's marketshare.
PSU-failure
Newcomer
To hold their marketshare, they'll need a flagship (enthusiast market) and competitive mainstream/value parts (OEM market).
The flagship needs to be faster than its competitor, HD5870 or HD5970 depending on its power draw.
The mainstream/value parts need to be power and cost efficient while not lagging too much behind AMD's in terms of performance, which is already almost guaranteed to not be the case (GT218 vs Cedar, GT216 & GT215 vs Redwood).
NV won't die instantly because of a bad product, but that will hurt them.
PSU-failure
Newcomer
They already have issues with their 40nm design and it's quite unlikely they could do something simpler while still having the exact same fonctionality (needed for the HPC market), so they simply can't design a more complex chip on 40nm and if they switch to 28nm, they'll face issues too, which will imply a smaller die and less than twice the complexity.
Let's face it : they need Fermi to be an HPC monster to really enter that market, so they can't count on a new architecture there. At the same time, if it's not competitive as a GPU it'll need a complete rethink. Mix this and you end up with Fermi being stillborn if not 40-50% faster than Cypress, be it on gaming or HPC workloads (and, as has been pointed out, for games it will probably be hindered by its rasterizing capabilities, system performance...)
Yeah and we know that nVidia is a company which never ever rethink their strategies.
Even GF100 is not faster than Cypress not everbody is switching right now to a new gpu card. nVidia put a GTX260-192 against 4870 und sold a lot of them. Don't underestimate their features like PhysX and 3D Vision.
PSU-failure
Newcomer
Sorry, you forgot to mention CUDA...Yeah and we know that nVidia is a company which never ever rethink their strategies.
Even GF100 is not faster than Cypress not everbody is switching right now to a new gpu card. nVidia put a GTX260-192 against 4870 und sold a lot of them. Don't underestimate their features like PhysX and 3D Vision.
Oh, btw there's a reason for only GT21x and Fermi supporting 3D Blu-Ray except marketing. it's not like if any recent GPU could render two HD streams at the same time.
PhysX is dead and gone now, CUDA can't compete with DC when it comes to games and 3DVision has always been a joke that will soon have an answer in the red camp (which will be a joke too since nothing is needed to do stereo except the avoidance of some tricks in games rendering).
you know, only 448 active Cores and therefore only 112 active TMUs.
If the rumored base-speed of 600MHz is correct than Fermi has a ~30% higher texturing fillrate and ~39% higher pixel fillrate than a GTX285. The Bandwidth advantage is not clear yet, but likely ~50%. Therefore I would expect that Fermi is roughly 30-60% faster than a GTX285, or maybe 10 - 20% faster than a 5870. As seen with the AMD GPUs of the last generations FLOPS alone do not count. IMHO; the performance of a GPU is still largely related to the TMU's, ROP's and bandwidth of the GPU.
