Will R580 have 48 shader 'pipes?'
- Thread starter ERK
- Start date
Supreet Virdi
Newcomer
Ok Chal thanks, well I can hardly wait 
Supreet Virdi
Newcomer
I am not sure.
davepermen
Regular
Chalnoth said:
the newest itanium is over 1 billion.. okay, right.. you mean gpu-chips
Yeah but doesn't that particular Itanium have like 24+MB of L3 SRAM on die? This is not really compareable to logic transistors IMO, since from what I understand, SRAM is
- easier to manufacture (structurally regular)
- denser
- consumes much less power
- easier to make defect tolerant
than logic transitors.
- easier to manufacture (structurally regular)
- denser
- consumes much less power
- easier to make defect tolerant
than logic transitors.
Chalnoth said:
You were in a higher range than that just for R580 a few days ago.
Are you expecting ATI to save transistors going unified? Well, maybe they might at that, but not net I suspect (i.e. I'm guessing they'd still add more units or "stuff" going from R580 to R600).
No, I'm suggesting this based upon a few things:
1. Die shrinks are slowing. And as we get smaller, it gets harder and harder to build the processors on the smaller dies. As such, as we move forward we're more likely to see IHV's not increase transistor counts dramatically on their first products for a particular process for reasons of yield.
2. The 80nm shrink should be very closely in-line with when the first DX10 parts should be shipping.
3. We already have the majority of the features that DX10 parts will need, so increases in transistor count on the order of what we have seen in the past are less likely.
In sum, I do expect that the first DX10 parts will be high-end parts targetted to sit at the top end of the scale, with configurations very similar to the parts to be released this coming Spring. That is to say, from ATI I would expect a Xenos-like setup with 48 shader pipelines and 16 texture pipelines, and from nVidia we'll probably get something almost exactly like the upcoming 90nm version of the G70 except with added DX10 features.
So, ATI will save some transistors by getting rid of the vertex pipelines, though both IHV's will have to add some more transistors to support DX10-level functionality. I don't expect to see either IHV really press the 80nm process to larger die sizes until the refresh of these parts.
1. Die shrinks are slowing. And as we get smaller, it gets harder and harder to build the processors on the smaller dies. As such, as we move forward we're more likely to see IHV's not increase transistor counts dramatically on their first products for a particular process for reasons of yield.
2. The 80nm shrink should be very closely in-line with when the first DX10 parts should be shipping.
3. We already have the majority of the features that DX10 parts will need, so increases in transistor count on the order of what we have seen in the past are less likely.
In sum, I do expect that the first DX10 parts will be high-end parts targetted to sit at the top end of the scale, with configurations very similar to the parts to be released this coming Spring. That is to say, from ATI I would expect a Xenos-like setup with 48 shader pipelines and 16 texture pipelines, and from nVidia we'll probably get something almost exactly like the upcoming 90nm version of the G70 except with added DX10 features.
So, ATI will save some transistors by getting rid of the vertex pipelines, though both IHV's will have to add some more transistors to support DX10-level functionality. I don't expect to see either IHV really press the 80nm process to larger die sizes until the refresh of these parts.
With the only other difference that Xenos with those unit amounts is lower than ~280 M transistors (add an estimated transistor count for logic and ROPs in the daughter die and leave the transistor count for eDRAM out of the equation). Considering the transistor count I expect for R580, I'd say that R600 could with ease be something like a 6* 16-way SIMD core.
As for NV's sollution it's can't be by far not just a G7x + a Geometry Shader, but a whole damn lot more than that. Not as much related to the amount of units, but how they operate in general.
It also doesn't have that whatever-it-is in the middle-left of the R520 die. The amount of die that thing takes makes most sense to me as a forward-looking element that is expecting to get relatively smaller (as a percentage of the die) right along --and that means way more transistors.
But NV? Heckifino. If my sig turns out right, that could throw a monkey in the wrench in a lot of ways.
But NV? Heckifino. If my sig turns out right, that could throw a monkey in the wrench in a lot of ways.
