This is the second time it's been repeated, but I don't see how you can categorize this as a gt200 with new front-end unit. Before we even knew anything about the frontend/fixed function units back when the Tesla/Fermi compute-sections were revealed, it was obvious that a lot had changed. so much so that back then, people were speculating that NVidia had not spent any effort on graphics and were getting out of the gfx business. Different DP arch, lots of changes to cache architecture, support for exceptions, concurrent kernels, etc etc. Add in the completely new front-end units and I just don't see how your conclusion can be reached.
NVIDIA GF100 & Friends speculation
- Thread starter Arty
- Start date
This is the second time it's been repeated, but I don't see how you can categorize this as a gt200 with new front-end unit. Before we even knew anything about the frontend/fixed function units back when the Tesla/Fermi compute-sections were revealed, it was obvious that a lot had changed. so much so that back then, people were speculating that NVidia had not spent any effort on graphics and were getting out of the gfx business. Different DP arch, lots of changes to cache architecture, support for exceptions, concurrent kernels, etc etc. Add in the completely new front-end units and I just don't see how your conclusion can be reached.
To do that, they must have reapportioned a fair bit of area from other units, without reducing their throughput, right?
Honestly, is AMD sooo enamored of nv's renaming shenanigans that they have decided to rebrand their drivers?
2650*1600, 8xAA -- And the contestant, that's running out of memory is... guess who!
Almost all the English language leaks so far seem to call it Catalytic.aaronspink
Veteran
Lots and lots of CLAIMS of new and different things. I'm not sure that any of these really are new and different things instead of just minor packaging differences. Lets just say that since the whole G80 128 "cores" BS, I take claims made by Nvidia PR at less than face value. The only difference that you've listed that MIGHT be real is the changed (not completely new but CHANGED) front end unit and DP SIMDs (which isn't much of an architectural big deal). So I remain at the evo side of the road.
aaronspink
Veteran
Or the additional 480 ALUs they added didn't take up that much area do to their minimal control via a x160 design. But who am I kidding, it must be some super secret magic sauce, right? Its not like they added another ~70 mm^2 of area to the design with the primary change being the number of ALUs.
That may be. I'd think HDMI would be more usable than DP though, given the market penetration of displays which support each connector.
The TU's were increased in the same ratio. The ROPs stayed put.
Of course, all you know about G80 is that it also has lots of CLAIMS about being a new architecture as compared to G70. Similarly, the Core architecture is just a set of CLAIMS about it having a few minor different things from the Pentium 4. Right?aaronspink said:
Where do you draw the line as to what is and isn't a new architecture? New ISA? Different block diagrams? Different gross process technology (eg: CMOS vs TTL)?
aaronspink
Veteran
DavidGraham
Veteran
Seriously , I would love to hear your take on the current designs from both ATi and Nvidia , it seems that you think everything accomplished today is just a piece of propaganda .. does that apply to ATi too ? I mean the whole "3200" "cores" thing ?
and I see you have a problem with G80 having a 128 cores definition .. do you think the number of true cores is less than advertised ? then why ? and what is the real number ?
aaronspink
Veteran
Last new architecture for Nvidia? G80. ATI? IIRC, R600. Fundamentally everything since have just been upsized designs with overall minor architectural changes here and there.
Likewise, last new architectures for Intel/AMD are Atom and K7 respectively.
as for G80, IIRC, the design contained 8 cores each with a 16 wide SIMD or its 16x8. Can't remember off the top of my head.
aaronspink
Veteran
Yes irony, given the vast number of Intel products that are the same architecture you manage to pick one of the only combos that have nothing in common architecturally.
Don't share anything at the pipeline level. Check
Don't share anything at the hazard level. Check
based on completely different OOE philosophies. Check.
etc.
All of which are observable at the performance/programming level? check
G80/gt200, no so much. G100, unlikely to prove itself much different than G80/gt200 since its still fundamentally the same architecture.
Last edited by a moderator:
aaronspink
Veteran
I think it's just PHK's sick joke...Honestly, is AMD sooo enamored of nv's renaming shenanigans that they have decided to rebrand their drivers?
Do I have to point you at the Fermi Compute whitepaper again? *sigh* Even in the SMs alone, the list of differences from GT200 is pretty long (let alone differences from G80).
1- Denormal support for fp32.
2- Integer ALU with 32-bit multiply, 64-bit operation support, and new instructions (conversions, bit field extract/insert, pop count).
3- Support for all C++ functionality (indirect branching, C++ exceptions, etc).
4- Predication.
5- MADs replaced by FMAs.
6- IEEE 754-2008 instead of 1985.
7- More SPs per SM.
8- Cache hierarchy
9- Configurable shared memory / cache size, instead of a dedicated RAM.
10- Unified address space.
11- 40-bit pointers instead of 32-bit pointers.
12- Two warp schedulers.
13- Out of order thread block execution & completion.
14- ECC in the register file, L1 cache, shared memory, L2 cache, and DRAM.
From #1 to #6, there is little to no reuse of the logic inside the SPs from GT200.
From #7 to #14, there is little to no reuse of the SM logic outside the SPs or in the memory subsystem.
And that's only the list for the SM, only for Compute, and only what's been publicly announced. If you start adding the differences in Graphics, you end up with a very, very long list of differences, and very little in common other than at some hand-wavy high level (duuuh, dem chips do math!)
Of course, this is unlikely to convince you because you haven't set forth any metric by which one could determine whether two architectures are the same or different.
1- Denormal support for fp32.
2- Integer ALU with 32-bit multiply, 64-bit operation support, and new instructions (conversions, bit field extract/insert, pop count).
3- Support for all C++ functionality (indirect branching, C++ exceptions, etc).
4- Predication.
5- MADs replaced by FMAs.
6- IEEE 754-2008 instead of 1985.
7- More SPs per SM.
8- Cache hierarchy
9- Configurable shared memory / cache size, instead of a dedicated RAM.
10- Unified address space.
11- 40-bit pointers instead of 32-bit pointers.
12- Two warp schedulers.
13- Out of order thread block execution & completion.
14- ECC in the register file, L1 cache, shared memory, L2 cache, and DRAM.
From #1 to #6, there is little to no reuse of the logic inside the SPs from GT200.
From #7 to #14, there is little to no reuse of the SM logic outside the SPs or in the memory subsystem.
And that's only the list for the SM, only for Compute, and only what's been publicly announced. If you start adding the differences in Graphics, you end up with a very, very long list of differences, and very little in common other than at some hand-wavy high level (duuuh, dem chips do math!)
Of course, this is unlikely to convince you because you haven't set forth any metric by which one could determine whether two architectures are the same or different.
So you believe those benches are real?
yep,I think Auska gave those benchmark numbers to PHK,some results are exactly the same.
PSU-failure
Newcomer
Although there's no doubt it's a sensible step-up from G80, many of the changes you list are just prerequisites for SM5 compatibility or whithout any relation to the architecture itself.
In fact:
- points 1 and 5 are part of point 6, and as point 6 is an SM5 prerequisite it's part of the differences between R700 and Evergreen too
- points 10 and 11 are part of point 3 (and btw adresses are locally way less than 40-bit, that's just a schematic view with almost no hardware requirement... only exceptions handling is new)
The real "issue" is whether or not all this have an effect on D3D performance, and it seems to not be the case as the only scenarii where GF100 takes the lead are those where Cypress struggles either because of highly incoherent shaders dispatch (Heaven's tessellation which is the only 99% certain real measurement) or games that always favored nVidia GPUs (FC2 with AA... strange considering how many games perform way better with AA and particularly 8x with AMD GPUs but not this one, don't you think?)
Now, since FC2 numbers come from a source which also gave some strange STALKER CoP numbers, there's no reason to draw conclusions already, we'll know the truth in 2 weeks now after all.
Similar threads
