AMD: R8xx Speculation

[itsmydamnation]

slide 2,


Highest performance and most feature rich products for at least the next 6 months.

date on slide is dec 2009, so AMD expect G100 in MAY

 

[SimBy]

slide 2,


Highest performance and most feature rich products for at least the next 6 months.

date on slide is dec 2009, so AMD expect G100 in MAY

Or that AMD refresh is coming in 6 months

 

[neliz]

slide 2,


Highest performance and most feature rich products for at least the next 6 months.

date on slide is dec 2009, so AMD expect G100 in MAY

Remember this Post by Ailuros?
By that time (2 Months after the first tape-out) AMD were aiming for somewhere between February or May for Fermi. Mostly because NV was telling it's partners February by then.

But I do wonder why all NV's product and family briefs are date in October, I guess they hoped to be finished by then.

 

[fellix]

Now, this is 5 new cents for me!

So, Cypress is eight channel config (8*32-bit), not four. The L2 partitioning is also indicated as eight banks of 64KB.

A die-shot would have revealed this in plain sight, if it was available.

 

[Psycho]

Highest performance and most feature rich products for at least the next 6 months.
date on slide is dec 2009, so AMD expect G100 in MAY

Highest performance is easy as 2.7 tflops (peak..) is more than ~1.5 tflops and features can be counted in many ways I guess..

 

[Silent_Buddha]

Need more info on Rv810/830 (equivalents, I hate trying to remember tree names).

Anyone have rumors/info on when NDA breaks for those? I'm jonesing for a new video card for my HTPC.

Regards,
SB

 

[neliz]

Need more info on Rv810/830 (equivalents, I hate trying to remember tree names).

Anyone have rumors/info on when NDA breaks for those? I'm jonesing for a new video card for my HTPC.

Regards,
SB

Hint: There's a big trade show this month..

 

[hoom]

Has anyone worked out how ATI has implemented the FMA so it can also do MAD with separate rounding?

Would be interesting to know if they've gone for bridged FMA as AMD seems likely to be doing for Bulldozer (with significant area penalty vs straight FMA) or if they are doing something else?

 

[Jawed]

Has anyone worked out how ATI has implemented the FMA so it can also do MAD with separate rounding?

Would be interesting to know if they've gone for bridged FMA as AMD seems likely to be doing for Bulldozer (with significant area penalty vs straight FMA) or if they are doing something else?

There's a wodge of additional circuitry in the 4 main lanes to do with dot-product. And there's also the double-precision stuff (not sure about that, though - we're not sure yet if the non-DP GPUs support FMA yet...).

See page 119:

http://developer.amd.com/gpu/ATIStr...een-Family_ISA_Instructions_and_Microcode.pdf

the INTERP_XY instruction. Note the use of the dot-product stuff. Also page 12:

http://sa09.idav.ucdavis.edu/docs/SA09_AMD_IHV.pdf

shows other capabilities using the dot-product adders. You can see that the two parallel dependent MULADDs consist of:

X: A * B + C
Y: D * E

Z: G * H + I
W: J * K

So A = X: + Y: and F = Z: + W:

This isn't an answer, just stuff.

Jawed

 

[MfA]

Their hardware is significantly more complex to start with to be able to support SP/DP with reused multipliers/adders, the overhead of the bridge unit is likely not as big a deal for them. (For desktop processors the area it takes is not really an issue either, but for different reasons ... a 40% increase in something that occupies an infinitesimal part of the die in the first place is irrelevant.)

PS. I think we can safely say they have the most complex piece of multifunctional floating point circuitry in existence at the moment.

 

[Gipsel]

128 KB R/W cache for global atomics ... that's quite a bit. Is this a general R/W cache for all UAV load/stores?

Actually, on page 18 it reads "Global Memory R/W Cache with Atomics". So yes, from that it should be a general cache for all UAV accesses.

 

[Tchock]

P7260 for Mobility 5870 + i7 720QM + 8GB

http://www.engadget.com/2010/01/05/live-from-asuss-ces-press-event/

 

[neliz]

http://guru3d.com/news.html#10592

First Exposure of R5870 Lightning in the World
MSI Lightning series have smashed lots of world records after launch. To fulfill the demand of extreme overclockers and hardcore games, now MSI evolves Lightning series into ATI series and introduces R5870 Lightning, built to break world records. To complete ultimate overclocking, R5870 Lightning is designed with 10-layer wider PCB and 13 phase PWM for GPU to greatly enhance overclocking capability and performance.

Moreover, the thermal design of R5870 Lightning uses Twin Frozr II thermal design with SuperPipe technology. Twin Frozr II is equipped with two 8cm large size fans, 5 heatpipes, and high density fins to effectively dissipate heat from GPU, memory, and power module area. With Twin Frozr II thermal design, cooling efficiency and overclocking stability will be reinforced. Applied with “Military Class” components, R5870 Lightning is the best choice for utmost overclocking.

Now that's something I want to see pics off, the high leakage Cypress GPU's on something that begs for power.

 

[mczak]

P7260 for Mobility 5870 + i7 720QM + 8GB

http://www.engadget.com/2010/01/05/live-from-asuss-ces-press-event/

That's almost dead on HD 5750 performance isn't it. So since it has 800 cores clocks (gpu or mem or both) are probably slightly below that of HD5750.
For a part with 5870 in the name that's not really all that much, but it should be a bit faster than any other current "mobile" (single) graphic chip.

 

[AnarchX]

P7260 for Mobility 5870 + i7 720QM + 8GB

http://www.engadget.com/2010/01/05/live-from-asuss-ces-press-event/

@ 1366x768: http://images.anandtech.com/galleries/539/asus-rog-g73jh-3dmark.jpg

So at standard 1280x1024 it is even lower. Not really something that should be named HD 5870.
Cypress @ 500-550MHz would be a nicer solution: http://www.computerbase.de/artikel/..._hd_5000_weitere_details/13/#abschnitt_part_2

 

[Pressure]

That's almost dead on HD 5750 performance isn't it. So since it has 800 cores clocks (gpu or mem or both) are probably slightly below that of HD5750.
For a part with 5870 in the name that's not really all that much, but it should be a bit faster than any other current "mobile" (single) graphic chip.

You do realize that the Mobility Radeon HD 5870 is just that, a Juniper based card, and not Cypress?

 

[neliz]

One Dealer just canceled Sapphire HD5870 Vapor-X cards mentioning they have been EOL-ed. Anyone else got a hint of this?

edit: Figured it out. Some distri's cancelled orders for the part since availability will be non-existent until mid-march.

 

[trinibwoy]

Availability of what? Chips or Vapor-X coolers?

 

[neliz]

Availability of what? Chips or Vapor-X coolers?

A combination maybe? most likely the chips though since multiple distri's now say mid-march for multiple products. The Vapor-X got that date, MSI OC models also have that date for availability (March 27nd.)

 

[mczak]

You do realize that the Mobility Radeon HD 5870 is just that, a Juniper based card, and not Cypress?

Sure. But that it's even slower than the stripped, downclocked desktop juniper part (I didn't realize the vantage score is cheated) is slightly disappointing given its name.
But what do I complain, nvidia sold g92 based GTX280M and friends (which were also slower than desktop GTX 9800) for ages.