AMD Radeon HD 7730 Review: A Harbinger Of The Kaveri APU?

[megamanxtreme]

Just using the toms graphics hierarchy chart, it looks like the regular 7750 is 6 tiers above your GT 440 with GDDR 5. Given that you have a GDDR3 version of the 7750, that puts it probably no lower than 3 tiers under the regular 7750, which is still 3 tiers higher than your GT 440m which should be a pretty solid upgrade honestly. The chart isn't perfect but its a good general reference for where things stack up, and honestly that 7750 should give you a nice boost.

Thanks for the information, it is a head-scratcher when they don't include benchmarks for those cards that we might have lying around and we have to guess-ti-mate.

 

[mohit9206]

I love my 7750 1gb gddr5 and its great for upto 900p gaming at high settings.Good to see more cards coming out giving customers more to choose from but also confusing them.I think its time for ddr3 cards to die in retail cards though oem's may continue to use them if they want.

 

[ta152h]

Talk about Kaveri using GDDR5 is plain uninformed.

It's a terrible memory for a CPU. Latency is too high. It would kill the processor. Not to mention power use.

PS4 gets away with it because it's a game machine. Most people buying a Kaveri APU aren't particularly interested in using it as a game machine, or they'd buy something with a discrete card. So, it's a CPU first, and that means it needs lower latency memory to perform, not higher.

GDDR is great for graphics implementations, but not a good memory for CPUs. If AMD really wanted to make an APU beast, which is highly doubtful, they would be better off going with DDR3 and going with a wider memory channel. It comes with a cost, so it's doubtful, but at least you wouldn't kill your processor with horrible memory latency.

 

[InvalidError]

Is DDR3 significantly cheaper to produce?

Not really.

But the much lower production volumes do mean much fewer units to spread engineering and tooling costs on. The tighter timing margins on GDDR5 signals and much higher power also make GDDR for large memory configurations or designs on a tight power budget, pretty much guaranteeing that GDDR will remain a niche-market memory type.

 

[Onus]

This card being able to flash to a HD7750 would be a killer deal. Mainstream users would scarf it down at $60 for its stock capabilities vs. a typical IGP, and because it won't require a new PSU. Poor gamers would buy it and flash it, and be able to play their games without suffering too badly.

Are you listening, AMD?

 

[InvalidError]

This card being able to flash to a HD7750 would be a killer deal.

But that would also kill AMD's margins - as if they weren't dangerously thin already.

I can't imagine AMD under-cutting themselves at such low price points doing any good to their bottom line.

 

[Onus]

Yeah, that's possible. It's not like as-is it doesn't render the GT640 irrelevant.

 

[cleeve]

Also, how is this Cape Verde GPU a "Harbinger Of The Kaveri APU"? It is a trimmed down 7750 and since a 7750 can provide no real insight into the performance of upcoming Kaveri APUs then how does this entry level card provide any better insight?

Kaveri leverages the GCN architecture, and so a cut-down 7750 most definitely has the potential to portent what Kaveri will do...

 

[cleeve]

On page 10, does the reference HD 7750 really have a load temp of 97C?
(73C + 24C ambient = 93C)

Anyway, the GDDR5 HD 7750 looks like a viable game-enabler for PC gamers on a budget. Also, it's a nice refresh for that price point's options. I do hope it sells for around $60 or less. :-D

I think you might misunderstand. The temperatures are absolute, but we simply started the beginning of the chart at the ambient temp (notice it doesn't start at zero, the chart starts at 24 degree ambient).

 

[cleeve]

...f you want to mention kaveri in the article, then please remember that high end kaveri will ship with 512 shaders and not 384...

There is no reliable source I'm aware of that suggests Kaveri will be 512 shaders.

That's speculation, and the same kind of misinformation has unsuccessfully predicted the launch of every single one of AMD's APUs so far.

 

[cleeve]

I fail to understand why you tested all of these cards at 1920X1080 when they will most likely not be used at that resolution. Why not test them at the resolution that people who buy these would actually be using? Like 720p? I think that it would make for a better review and be closer to real world use.

1080p is the most common gaming monitor resolution, and the GDDR5 7730 can handle it.

 

[InvalidError]

I fail to understand why you tested all of these cards at 1920X1080 when they will most likely not be used at that resolution.

Why wouldn't they get run at that resolution? There are dozens of 1080p screen models in the $100-150 range and it seems quite reasonable for the average would-be PC gamer to try their native resolution first. Without the fancy bling enabled, even old-ish GPUs can handle 1080p reasonably well without giving up that much detail.

I'm still using my old HD5770 and most of the details I disable are things like clutter and shadows that I would disable even if I had high-end graphics to remove unnecessary distractions/obstructions anyway.

 

[spentshells]

I fail to understand why you tested all of these cards at 1920X1080 when they will most likely not be used at that resolution. Why not test them at the resolution that people who buy these would actually be using? Like 720p? I think that it would make for a better review and be closer to real world use.

1080p is the most common gaming monitor resolution, and the GDDR5 7730 can handle it.

Just noticed your avatar, pretty sweet.

 

[Wisecracker]

...f you want to mention kaveri in the article, then please remember that high end kaveri will ship with 512 shaders and not 384...

There is no reliable source I'm aware of that suggests Kaveri will be 512 shaders.

That's speculation, and the same kind of misinformation has unsuccessfully predicted the launch of every single one of AMD's APUs so far.

Far be it for me to tamp down expectations, but pretty much this.

The Kaveri APU 'SIMD Engine' is likely structured similar to the Mobility Radeon HD 86XX 'Mars' discreet chips with 384 cores ... with all the fancy power-gated zero-core and power-tune logic. It's performance envelop should fall easily between the GT640 and Radeon HD 7750

 

[cleeve]

Just noticed your avatar, pretty sweet.

Heheh, thanks man.

You're one of the three people who get it.

 

[Evolution2001]

The 7730 with 2GB DDR3 is up at NewEgg for $75 (+$5.39 shipping)

http://www.newegg.com/Product/Product.aspx?Item=N82E16814202038

 

[Onus]

That's way overpriced, for a DDR3 version. It needs to lose at least $10. That would be an ok price for GDDR5.

 

[rp20]

I seriously fail to understand what is happening here. How is the power consumption between the GDDR5 6670 and the 7730 the same and yet you only have a negligible increase in performance. The average game performance shows a relative performance from the 6670 ddr3 set as a baseline that the 6670 GDDR5 is 41% faster and the 7730 GDDR5 is 52% and if you do the math you get 52/41=1.08. That is odd... you only have a 8% improvement when you have a die shrink from 40nm to 28nm and on top of that you have a new architecture. What am i missing guys?

 

[InvalidError]

That is odd... you only have a 8% improvement when you have a die shrink from 40nm to 28nm and on top of that you have a new architecture. What am i missing guys?

Both have the same 900MHz reference design clock so neither has an advantage over the other there which leaves everything down to architecture. The 6670 has 480 VLIW cores while the 7730 has only 384 GCN shaders so most of the architectural gains are offset by losing 20% of the shader count.

Not that much of a surprise. What does surprise me is that 7730-GDDR5 uses a fair bit more power than 6670-GDDR5 despite the die shrink and lower shader count. This implies individual transistors in GCN are much busier than VLIW's - enough to eat power savings from both the die shrink, axed shaders and some more beyond that.

 

[rp20]

That is odd... you only have a 8% improvement when you have a die shrink from 40nm to 28nm and on top of that you have a new architecture. What am i missing guys?

Both have the same 900MHz reference design clock so neither has an advantage over the other there which leaves everything down to architecture. The 6670 has 480 VLIW cores while the 7730 has only 384 GCN shaders so most of the architectural gains are offset by losing 20% of the shader count.

Not that much of a surprise. What does surprise me is that 7730-GDDR5 uses a fair bit more power than 6670-GDDR5 despite the die shrink and lower shader count. This implies individual transistors in GCN are much busier than VLIW's - enough to eat power savings from both the die shrink, axed shaders and some more beyond that.

That is the thing, where is the power savings? I cannot wrap my head around the fact that you have about a 8% increase in perf and 5% increase in peak power consumption. Where is the advantage of the die shrink? Either people have been bullshitting this whole time when they tout the increased efficiency of the GCN or there is an incomplete story.

 

[Onus]

The increased efficiency is right there; you use 5% more power, but get 8% more performance, and you're using 384 GCN shaders vs. 480 VLIW cores.

 

[InvalidError]

Where is the advantage of the die shrink? Either people have been bullshitting this whole time when they tout the increased efficiency of the GCN or there is an incomplete story.

Advantage of the die shrink: smaller die.
Advantage of GCN: slightly better performance and power-efficiency using a smaller transistor budget.
Combined advantage: lower manufacturing cost (more dies per wafer) for the same performance level - which AMD desperately needs if they are to keep the low-end GPU market alive with their APUs eroding that very same market.

Personally, I'm just surprised the sub-$100 segments still exist.

 

[rp20]

The die size is actually greater. http://hexus.net/tech/reviews/graphics/58677-sapphire-radeon-hd-7730-1gb-gddr5/

That just means that these are most likely 7750 with defects. Which is not a bad thing.

On whether the efficiency gain is acceptable, I always had the assumption that you would get 30% savings in power per die shrink yet in this case it is negligible.

 

[sarinaide]

If it can run dual graphics this could be a mother of a chip at that price tag.

 

[Maygentria]

No surprises here then, but you really should go for a 7770 if you want decent frame rates in most games. A little money goes a long way in low-level graphics.