nvidia downgrades tesla, architecture broken

[flippin_waffles]

Nv isn't finished ordering its wares to the chopping block it seems.

http://www.semiaccurate.com/2010/05/...es-tesla-again

[NaMcO]

Yay, someone took semiaccurate down. Please be forever.

[Rebel44]

Yay, someone took semiaccurate down. Please be forever.

Its up.

[0xdeadbeef]

Great. More Nvidia hate from Charlie. Thanks.

[annihilat0r]

How did you come to the conclusion "the architecture is broken"? Because Charlie said so?

[god_43]

nice article.

[Eastcoasthandle]

nice article.

Indeed, I nearly forgot the timeline of specs back then.

[saaya]

wow, only 1.15ghz shader clock at 250W? wtf?
so those tesla cards are not good enough for gtx460s even, which they sell for 350$, and then they sell them as tesla cards for 2000$+ at lower clocks and higher tdp? HA!
and i didnt know nvidia initially aimed for gddr5 clocks of 2500... and now they ended up with 1500... OUCH... nv def seems to have some big probs with their memory controllers... they dont seem to work very efficiently, the 470 seems to be very mem bw limited already, and the 480 gets mem bw limited at high clocks, and thats with 384bits of gddr5 bw...

if they can solve their mem bw efficiency AND clocks, and get 512core cards with 750/1500... then fermi might be good enough to take on atis refresh...
but nv sure has a lot of work to do...

fix memory controller
improve sp yields
improve clock yields
improve functional chip yields
reduce tdp
... pheew...

2,3 and 4 seems to be what nvidia is aiming for with gf104, i wonder if 1 is a priority as well...
5 is the least of their problems actually i think

imo whats really surprising is how fermi is STILL able to compete despite all those issues..

[Coldon]

you do realize that tesla cards arent gaming cards... you should be comparing them to ATI firePro offerings, for all you know the fermi architecture might destroy the firepros in double precision which at the end of the day is what a lot of the professional cards are going to be used for.

lets see the ATI firepro v8800 is a 5870 with 2gb ram... compare that to the tesla card with 3gb or 6gb. Simulation wise more memory is a damn good thingas it means your datasets can be larger. Memory bandwidth is the same.

So yeh until someones runs a few GPGPU benchmarks on both cards and compares them: apples to apples - charlie can STFU. I'm not saying the card is better than the v8800 but you cant purely judge a product on specs... The v8800 might be faster on paper but its double precision operations might be crippling slow. For example everyone said fermi would be absolutely useless at tesselation and lo and behold its better at it than ATI's offerings :/

[BoredByLife]

Yay, someone took semiaccurate down. Please be forever.

SA is quite accurate and a good source of info if you filter the nvidia hate

[saaya]

you do realize that tesla cards arent gaming cards... you should be comparing them to ATI firePro offerings, for all you know the fermi architecture might destroy the firepros in double precision which at the end of the day is what a lot of the professional cards are going to be used for.

lets see the ATI firepro v8800 is a 5870 with 2gb ram... compare that to the tesla card with 3gb or 6gb. Simulation wise more memory is a damn good thingas it means your datasets can be larger. Memory bandwidth is the same.

So yeh until someones runs a few GPGPU benchmarks on both cards and compares them: apples to apples - charlie can STFU. I'm not saying the card is better than the v8800 but you cant purely judge a product on specs... The v8800 might be faster on paper but its double precision operations might be crippling slow. For example everyone said fermi would be absolutely useless at tesselation and lo and behold its better at it than ATI's offerings :/

o rly? :P
and why should i compare it to an ati professional card?
cause thats how nvidia positions it?

the point is that yields must be pretty bad if nvidia is selling tesla cards with so low speeds and such a high tdp... tesla is the card they make the most money with, so why are they selling the worst or second worst after the gtx460, bin to the gpgpu customers? perf per watt matters a lot there, other than in the desktop segment... so this is weird... why they sell their worst bin there? well probably because they CAN...
it still says a lot, if they create a bin that low, it means that they have quite some chips that fall into this bin... so it gives an interesting view on what their yields are probably like...

and about specs...
radeon 5870:
SP: 2.72 Tflops
DP: 544 Gflops

so the 5870 has a higher theoretical DP rate than a tesla card that costs 5x as much... sure, not as much ECC as the gf100, but for this cost you could let 2 or 3 cards do the same task and then compare results lol...

[grimREEFER]

from a gross profit standpoint, tesla is the least profitable of their cards.

but depending on the type of work that is being done, there is a good chance that it might be your only choice if you want to use gpu computing.

http://www.appro.com/product/1426G4server_overview.asp

they have a 1u server with 4 fermi tesla's and 2 6 core xeons. it's crazy.

[Coldon]

o rly? :P
and why should i compare it to an ati professional card?
cause thats how nvidia positions it?

Cause it is a professional card, what you're saying is that i should compare it to the mainstream cards? So if we were talking cars, its like comparing a hmmwv to a land rover, and then saying the land rover has a much faster top speed and better fuel efficiency so the hmmwv is useless... Another example is lets say nvidia positions a gtx 285 in the high end and then you say you can compare that to a ati 4350, because obviously where nvidia positions its cards doesnt matter.

the point is that yields must be pretty bad if nvidia is selling tesla cards with so low speeds and such a high tdp... tesla is the card they make the most money with, so why are they selling the worst or second worst after the gtx460, bin to the gpgpu customers? perf per watt matters a lot there, other than in the desktop segment... so this is weird... why they sell their worst bin there? well probably because they CAN...
it still says a lot, if they create a bin that low, it means that they have quite some chips that fall into this bin... so it gives an interesting view on what their yields are probably like...

and about specs...
radeon 5870:
SP: 2.72 Tflops
DP: 544 Gflops

so the 5870 has a higher theoretical DP rate than a tesla card that costs 5x as much... sure, not as much ECC as the gf100, but for this cost you could let 2 or 3 cards do the same task and then compare results lol...

have you ever done any sort of GPGPU programming? ECC is extremely important, doing any sort of numerical simulation on previous gen cards was a nightmare, for example doing a PSO implementation on a GPU just wasnt feasible due to the atrocious FP precision (at least not the g80/g92 series) and performance went down the drain once you try DP on those cards.

customers in that industry are more concerned with accuracy and memory bandwidth than speed, just look at ATIs product page, you cannot find the core or memory speeds listed on any of spec sheets. Why not? Cause their target market doesnt care about that!

okay so if you are saying that the tesla cards are worse than a "gtx460" and positioning doesnt matter, why arent companies buying 5870s instead of v8800s? hm... Why arent gamers buying v8800s instead of 5870s?

They are different cards, different markets, different requirements!

Once again you bring specs into play, why are fermi based cards at 850mhz faster than 5870s at 1000mhz? and shouldn't 1600 cores be ALOT faster than 448?! OH noez, the specs liez to uz!!! Paper specs dont necessarily show the whoel picture.

[hyperdoggy]

In theory tesla is the better card for gpu computing than anything else as it was designed for it from the ground up. staying on topic however, SA got it right about the half the performance per watt bit. Not a huge deal in low-mid range workstation but for the high-end market(which is also the largest pie for tesla) this is a big issue.

[Sn0wm@n]

you do realize that tesla cards arent gaming cards... you should be comparing them to ATI firePro offerings, for all you know the fermi architecture might destroy the firepros in double precision which at the end of the day is what a lot of the professional cards are going to be used for.

lets see the ATI firepro v8800 is a 5870 with 2gb ram... compare that to the tesla card with 3gb or 6gb. Simulation wise more memory is a damn good thingas it means your datasets can be larger. Memory bandwidth is the same.

So yeh until someones runs a few GPGPU benchmarks on both cards and compares them: apples to apples - charlie can STFU. I'm not saying the card is better than the v8800 but you cant purely judge a product on specs... The v8800 might be faster on paper but its double precision operations might be crippling slow. For example everyone said fermi would be absolutely useless at tesselation and lo and behold its better at it than ATI's offerings :/

even if fermi looks good on paper we saw that they didnt do too well against existing ati card .... now compare both of the same professional card and i doubt that the difference can be this much... + since those card are high proffit margin card ati could lower their price a bit... since they seems to get great yield out of those... and blam... the fireprogl seems more appropriate

[Kuroimaho]

The clocks were lower on teslas than the desktop cards fore the last gen as well, even after the die shrink. Compared to the last gen, now it didn't take NV a die shrink to release the Teslas is a step forward.

[ElSel10]

and about specs...
radeon 5870:
SP: 2.72 Tflops
DP: 544 Gflops

so the 5870 has a higher theoretical DP rate than a tesla card that costs 5x as much... sure, not as much ECC as the gf100, but for this cost you could let 2 or 3 cards do the same task and then compare results lol...

You also have to consider that CUDA is strongly ingrained into a lot of GPGPU. It's the gold standard of GPU computing taught at universities as well.

Not only that, I don't know the technical reasons, but the 1.3TFlop SP GTX480 beats the 2.7TFlop SP 5870 in games. So if you translate that to DP, 515GFlop Tesla vs 544TFlop FirePro?

o rly? :P
and why should i compare it to an ati professional card?
cause thats how nvidia positions it?

Yes, that's exactly why. Nobody buys 480s or 5870s for large GPGPU farms. That isn't how it works, and so they isn't how they are compared. You can't dismiss industry standards.

[Mechanical Man]

...

Not only that, I don't know the technical reasons, but the 1.3TFlop SP GTX480 beats the 2.7TFlop SP 5870 in games. So if you translate that to DP, 515GFlop Tesla vs 544TFlop FirePro?

...
.

Chips are made of different arch, So one game code good for other aint so good to otherone.

Basically when calculating something with gpgpu, its lot easier to optimize, so I wouldnt take any conclusions about gfx performance.

[Chumbucket843]

the point is that yields must be pretty bad if nvidia is selling tesla cards with so low speeds and such a high tdp... tesla is the card they make the most money with, so why are they selling the worst or second worst after the gtx460, bin to the gpgpu customers? perf per watt matters a lot there, other than in the desktop segment... so this is weird... why they sell their worst bin there? well probably because they CAN...
it still says a lot, if they create a bin that low, it means that they have quite some chips that fall into this bin... so it gives an interesting view on what their yields are probably like...

and about specs...
radeon 5870:
SP: 2.72 Tflops
DP: 544 Gflops

so the 5870 has a higher theoretical DP rate than a tesla card that costs 5x as much... sure, not as much ECC as the gf100, but for this cost you could let 2 or 3 cards do the same task and then compare results lol...

yields are obviously bad but these are normal clocks for the intended market. if you look at the Chinese supercomputer, milkyway one, with 5120 rv770's the shaders are 575MHz. they are lucky because the problems they are solving arent prone to memory errors which is probably downclocked too. errors scale linearly with number of chips and becomes an issue because you cant afford to have your system failing constantly. that's why ecc is so important. performance per watt is 0 if you cant even finish the task.

not to mention 1GB is painfully small for how fast a gpu is. for a gamer it would be like a 256MB 5970. 6GB will be very convenient for HPC.

[spursindonesia]

Blah, blah, blah, outside all of Charlie's hate spewed in this thread from certain forum members which belong to "Charlatan (Charlie the Satan) fans group", can any of them REFUTE what this article claimed of regarding the downgrading of Tesla Fermi's specs ???

[570091D]

why is everyone surprised again? let's take a minute and remember all of the hardware that has been ENABLED on these chips vs the geforce range. the higher power consumption isn't totally unreasonable considering what this card is packing....

[zanzabar]

Blah, blah, blah, outside all of Charlie's hate spewed in this thread from certain forum members which belong to "Charlatan (Charlie the Satan) fans group", can any of them REFUTE what this article claimed of regarding the downgrading of Tesla Fermi's specs ???

the downgrade was in performance per watt, the gf card got the same downgrade so i would expect it out of the tesla as they use the same die. he isnt saying that its much slower other than the clocks that they had to lower

[stinkas]

Yay, someone took semiaccurate down. Please be forever.

Hope Nv improve soon!
Looks like the fans are reaching breaking point...

[Russian]

This article is like beating the Fermi dead horse. We know it isn't as good as everyone (including nvidia) had hoped, but my god repeating this whole thing all over again is quite annoying.

[Mescalamba]

Architecture is fine, just manufacturing process cant handle that. Nothing new under sun. It will be better and maybe even great if they use it with next manufacturing process. Tough not sure which one it will be.. TSMC said something about 20nm.