SolMiester
Diamond Member
- Dec 19, 2004
- 5,330
- 17
- 76
It looks like you'll be holding that post in your pocket for a while.
Fermi isn't exactly right around the corner, as in here in a week or two. We don't know clock speeds, but do know what Fermi has under the hood to some degree. I wouldn't count on it really whipping up the 5870 in gaming, but as you say, we just don't know until it's out and we see what it can do... anything at this point is little more than guesses and speculation (by fanboys of either side).
Last edited:
at80eighty
Senior member
- Jun 28, 2004
- 458
- 5
- 81
As far as I can tell, based on the AMD website, this new Tesla beast will kill ATI in the HPC market, since it has more than DOUBLE the power of AMD's best offering.
AMD offer a product with ~250GFlops for $1200.
NV are going to offer a product with maybe 640GFlops for $2500.
What people seem to be missing is that Fermi GPGPU is planned to be released in Q2 2010 (so sources say), and they are comparing it to the HD5870.
Fermi GPGPU will compete with ATI Firestream offerings.
There is no HD5870 based Firestream. Fermi will, as things stand, be competing with HD4870 based Firestream, and soundly beat it.
Until AMD release an HD5870 based Firestream, NV are neither late to market, nor priced absurdly, nor being soundly beaten for 6 months, in this market.
AMD offer a product with ~250GFlops for $1200.
NV are going to offer a product with maybe 640GFlops for $2500.
What people seem to be missing is that Fermi GPGPU is planned to be released in Q2 2010 (so sources say), and they are comparing it to the HD5870.
Fermi GPGPU will compete with ATI Firestream offerings.
There is no HD5870 based Firestream. Fermi will, as things stand, be competing with HD4870 based Firestream, and soundly beat it.
Until AMD release an HD5870 based Firestream, NV are neither late to market, nor priced absurdly, nor being soundly beaten for 6 months, in this market.
My argument was that ATI has 6 months to release a Firestream based on the HD5870, which would compete with the new Tesla. It is such a long time away, that we can't be sure those MSRP figures will hold up. At that point in time they may actually have competition that will force them to use a lower price point. My whole point was that it is too far from launch to take those values seriously, as a whole lot of things could happen in the mean time.
I can't see AMD not using the Cypress GPU in the Firestream parts. I don't think they're just abandoning that line after the RV770 based parts, or not updating them. Nvidia plans on the GeForce line in Q1, than the Tesla parts in Q2. So it wouldn't be too much of a stretch to imagine it could take AMD a quarter or two (or more) to introduce their Cypress based parts.
Wreckage
Banned
- Jul 1, 2005
- 5,529
- 0
- 0
AMD also has nothing that compares to CUDA and the 100's of applications that use it. They have pretty much abandoned Brook+ and OpenCL is barely out of beta.
http://www.nvidia.com/object/cuda_home.html#
GaiaHunter
Diamond Member
- Jul 13, 2008
- 3,700
- 406
- 126
To be honest to MrK6, I think he is talking more about the Fermi GeForce derivatives, that will compete with Cypress, rather than Tesla derivatives.
I don't think Tesla are supposed to compete with 58x0 and vice-versa.
I don't think Tesla are supposed to compete with 58x0 and vice-versa.
SolMiester
Diamond Member
- Dec 19, 2004
- 5,330
- 17
- 76
MrK6
Diamond Member
- Aug 9, 2004
- 4,458
- 4
- 81
It's poor argument/debating skills, I pointed it out. Is "stupid" the problematic word? If that's too abrasive, insert obtuse, the point still stands.
I didn't, reread what I wrote. Honestly, some of you get on your sensitivity helmets too quickly - read what I wrote, not what you think I meant. If Idontcare wants to use logical fallacies and poor arguments to back up his points due to a lack of evidence or logical reasoning, he can waste someone else's time with it.
Both arguments have merit. To couple my response with this point:
I think that's a good way to break down this discussion, since, I'd imagine, the two series will develop much different marketing strategies and scope. So, I'll designate which aspect I'm referring to from now on.
-->
-->
Let's just move back to the technology since you don't understand rhetoric.
Are you through with your pity party so I can continue this discussion? Why does "dropping the ball" have to incur a stiff penalty (if any)? Compared to their previous pace and accomplishments, NVIDIA is pretty lackluster right now.
What part of "by all reports" didn't you understand? I know you're getting all giddy defending your buddy, but try to reel yourself back into reality when you converse with me. NVIDIA is marketing a GPGPU that has theoretical double precision performance that's only 14% higher than that of AMD's part, but will be released six months later. My reigning point is "that's it?"
Good info, thanks. Could one assume that a 5870 based Firestream is not too far down the road, given the slide posted further up? They could always compare theoretical computational capacity and never do anything with it, but that seems kind of ridiculous. To summarize a few points, comparing the reported double precision computational power of Tesla Fermi is only 14% higher than that of AMD's 5870. I don't think that's that impressive considering how much NVIDIA has been touting GPGPU recently. One would think that after taking so much extra time to work on the part, they would have something more to show. Could there be other tweaks and benefits in the architecture that make it a great GPGPU? Sure, but I don't have data on those and can't comment on them. And if some of you are so opposed to reading any comments based on speculation and the limited info available, why do you read these types of the threads in the first place?
GaiaHunter
Diamond Member
- Jul 13, 2008
- 3,700
- 406
- 126
Aren't we all guessing at this time?
Edit: And by the way, last time NVIDIA released a product after ATI was the much maligned FX5800 series...
And actually, since then, NV was always the first to the market.
So while there is no rule saying NV needs to be the first to the market and no one can say how that will affect the success of the product (if it has any affect at all), it hasn't been common.
Last edited:
Few random observations-
With 16 DP issues per CUDA core and 16 cores the only way Charlie's imaginary 768GFLOPS would have ever been accurate is if nV was shooting for 3GHZ clock speed on the shaders. I don't think anyone is going to reasonably assume anyone is going to target 3GHZ on a 3Billion xtor part on a yet to be proven build process using bulk silicon for a mass market part. The only time I have seen the number 768 used for Fermi is in the amount of KBs of L2 cache it will have. This isn't any sort of complex insider information needed, just do the math. It is rather safe to say nV was never shooting for 768GFLOPS DP performance.
On Tesla versus ATi counterparts-
http://techreport.com/articles.x/17618/5
According to AMD they lack full IEEE 754-2008 compliance due to numerical exceptions, that combined with lack of ECC support removes them entirely from direct competition with Fermi. It isn't about performance alone, said performance must come with a certain level of accuracy for it to be able to compete in certain markets. The GT2x0 parts weren't ever a factor in certain segments of HPC due to lacking ECC- it didn't matter how fast or cheap it was compared to the competition.
GPUs have been faster, much faster, then CPUs at raw FPU for a long time now. They are making headway into the HPC and GPGPU market not due to raw performance, but due to increased flexibility. From every bit of documentation we have seen, Fermi has significantly more flexibility then 5xxx parts when considering GPGPU tasks. I don't see this as a slam to ATi in any way as they clearly are placing close to no emphasis on GPGPU functionality currently.
With 16 DP issues per CUDA core and 16 cores the only way Charlie's imaginary 768GFLOPS would have ever been accurate is if nV was shooting for 3GHZ clock speed on the shaders. I don't think anyone is going to reasonably assume anyone is going to target 3GHZ on a 3Billion xtor part on a yet to be proven build process using bulk silicon for a mass market part. The only time I have seen the number 768 used for Fermi is in the amount of KBs of L2 cache it will have. This isn't any sort of complex insider information needed, just do the math. It is rather safe to say nV was never shooting for 768GFLOPS DP performance.
On Tesla versus ATi counterparts-
http://techreport.com/articles.x/17618/5
According to AMD they lack full IEEE 754-2008 compliance due to numerical exceptions, that combined with lack of ECC support removes them entirely from direct competition with Fermi. It isn't about performance alone, said performance must come with a certain level of accuracy for it to be able to compete in certain markets. The GT2x0 parts weren't ever a factor in certain segments of HPC due to lacking ECC- it didn't matter how fast or cheap it was compared to the competition.
GPUs have been faster, much faster, then CPUs at raw FPU for a long time now. They are making headway into the HPC and GPGPU market not due to raw performance, but due to increased flexibility. From every bit of documentation we have seen, Fermi has significantly more flexibility then 5xxx parts when considering GPGPU tasks. I don't see this as a slam to ATi in any way as they clearly are placing close to no emphasis on GPGPU functionality currently.
Schmide
Diamond Member
- Mar 7, 2002
- 5,745
- 1,036
- 126
Is Fermi going to provide per-element exceptions? If they do where is this interrupt going to occur?
The whole context of this quote is to note they are generally compliant but lack this one type of operation that is not generally conducive to this type of architecture.
Yes.
Inside the FPU, hardware level support(which exceeds any CPU that I am aware of atm). They also support try/catch although as with the previous revisions of the spec that is still an optional feature.
GaiaHunter
Diamond Member
- Jul 13, 2008
- 3,700
- 406
- 126
Actually AMD estimated 768GFLOPS assuming 1.5GHz for shader clock.
You can see it in this post http://forums.anandtech.com/showpost.php?p=28922008&postcount=19 , in this very thread.
GaiaHunter
Diamond Member
- Jul 13, 2008
- 3,700
- 406
- 126
Hmmm, they must have had a PR guy handle that, one that failed basic math to a shocking degree
These dudes estimate the same http://techreport.com/articles.x/17670: .
I'm not privy to the maths concerning this estimate and I'm in lazy mood, so could you point the error those estimates incur on that leads them to similar numbers?
Thank you.
Sure,
Fermi can handle 16 DP calcs per SM per clock
There are 16 SM clusters in Fermi
16x16=256
So, Fermi can handle 256 DP calcs per clock
768,000,000,000/256=3,000,000,000
So, in order for Fermi to hit the 768GFLOPS number it would need to be clocked at 3GHZ, there is no way that was ever reasonable to expect from anyone.
The only comment that nV had made that could possibly have been remotely misunderstood is that Fermi would have ~8x the DP performance of the GT200, as the chart in the link you provided shows, the 280 had DP performance of 78GFLOPs, 78x8= 624(which given that is the high end and inside their high range, falls right into place). There are no numbers nV has ever hinted at that put the number in the 768 range, I have no clue where people got that assumption from outside of failure of very basic math
Last edited:
Elfear
Diamond Member
- May 30, 2004
- 7,165
- 824
- 126
GaiaHunter
Diamond Member
- Jul 13, 2008
- 3,700
- 406
- 126
http://www.nvidia.com/object/fermi_architecture.html
I guess there is the discrepancy - they state Fermi is 512 CUDA cores and that would put it at 1.5 GHz.
Where do you get 16 CUDA cores from? What am I missing?
lopri
Elite Member
- Jul 27, 2002
- 13,314
- 690
- 126
So using that math, we can expect Fermi's 'CUDA core' frequency will be somewhere around 2.0 GHz ~ 2.5 GHz? Out of curiosity.
Sorry, guess they are calling each shader unit a CUDA core, each cluster of cores a SM. So it's 16 DP per SM, 16 SMs. There are a total of 512 of what they are calling CUDA cores. 32 CUDA cores per SM, but each SM can only issue 16DP ops per clock and they can't dual issue a DP and SFU, so my math is still entirely accurate(although the names a bit off
- edited my previous post to straighten out the names ).According to the numbers nVidia has released so far, yes.
Last edited:
GaiaHunter
Diamond Member
- Jul 13, 2008
- 3,700
- 406
- 126
I've been reading the white paper: Each SM, that features 32 CUDA cores, can do up to 16 DP fused multiply-add operations per clock.
So, 512 CUDA cores, 16 SM.
520GFLops is the minimum range they have for the tesla.
So that puts the shader core a bit over 2GHz.
That does sound more reasonable.
So, 512 CUDA cores, 16 SM.
520GFLops is the minimum range they have for the tesla.
So that puts the shader core a bit over 2GHz.
That does sound more reasonable.
GaiaHunter
Diamond Member
- Jul 13, 2008
- 3,700
- 406
- 126
I've been drawb to attention that a Fused add multiply counts as 2 Flops per cycle.
So in fact, its 16*16*2, and that would put the shader core clock at 1.2-1.3GHz.
Is this information valid?
Thanks.
So in fact, its 16*16*2, and that would put the shader core clock at 1.2-1.3GHz.
Is this information valid?
Thanks.
TRENDING THREADS
-
Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)- Started by DisEnchantment
- Replies: 25K
-
Discussion Intel Meteor, Arrow, Lunar & Panther Lakes + WCL Discussion Threads
- Started by Tigerick
- Replies: 23K
-
Discussion Intel current and future Lakes & Rapids thread- Started by TheF34RChannel
- Replies: 23K
-
-
Facebook
Twitter