The Serpent Rider wrote on 2024-08-07, 02:18:

I won't place my bet on any FirePro cards, because just as Quadro cards, they could be locked out in terms of overclocking and tinkering with power limit on GCN 1.0 is not good.

Still, it would be interesting to see these cards all benchmarked together and power consumption measured. When the GPU shortage was at it's worst a lot of videos were popping up comparing cards like these but they were all generally focused on modern-ish gaming performance.

Also, regarding TechPowerup saying the K620 is 45 watts, that is directly from Nvidia's datasheet: https://www.nvidia.com/content/dam/en-zz/Solu … 20_US_NV_HR.pdf

As usual, TDP doesn't seem to really mean much in real world usage.

That said, the HD7750's TDP is 55W, so I'd be curious to know where it actually sits in a retro-gaming sense.

Nvidia is listing maximum power limit (VRM), which is not TDP. Also Radeon 7750 apparently consumes only 40W, while 55W is a hard power limit.

The HD 7770 is more level with the K620 performance wise, in general within 5% but different radeon or geforce favoring code could have either leading or trailing by 10+%
Then the HD 7750 is about 20% behind the 7770... the 650 is about level with the 7750 but sucks twice as much at idle.

However from these generations and about 2 more, cards went relatively weak in DX9 for some reason, but still may be faster overall, but there could be some sneaky ringers in a previous gen that do a bit better on DX9 maybe 5770, but in general the overall speed gain tends to make the inefficiency moot. Also they're much better on DX10/11 ofc.

Why not GT1030?

BitWrangler wrote on 2024-08-07, 03:09:

The HD 7770 is more level with the K620 performance wise, in general within 5% but different radeon or geforce favoring code could have either leading or trailing by 10+%

K620 has GDDR3 and measly 20W power consumption, while HD 7770 has GDDR5 and power consumption well over 50W. Both cards are 28nm. K620 can't be on par with 7770 in any shape or form. The numbers simply don't match.

The Serpent Rider wrote on 2024-08-07, 04:06:

BitWrangler wrote on 2024-08-07, 03:09:

The HD 7770 is more level with the K620 performance wise, in general within 5% but different radeon or geforce favoring code could have either leading or trailing by 10+%

K620 has GDDR3 and measly 20W power consumption, while HD 7770 has GDDR5 and power consumption well over 50W. Both cards are 28nm. K620 can't be on par with 7770 in any shape or form. The numbers simply don't match.

I think you're a bit too hung up on specs. As has been the case since basically the Nvidia G80 series (prior to that things were much more straight forward), numbers from one generation to another are not directly comparable. The farther apart the technologies get (age wise), the less they can be compared, and comparing specs of different brands is even less reliable. Despite TSMC 28nm and GloFo's 28nm having the same number in them, Nvidia's Maxwell was significantly more advanced than anything else before it. That's why a K620 can pull 20W using DDR3 and beat more power hungry cards with higher specs.

Look, someone did the work for us here:
https://www.youtube.com/watch?v=ES8Cbe4TIgM

The 7770 wins handily in Minecraft, and has a decent lead in PUBG minimum FPS, but in the other titles they are basically tied, with the K620 usually having a small lead. This is a much more nuanced thing than just saying X number is so much bigger that nothing else matters. That just isn't the case, and it is highly dependent on games and settings. Of course some games will perform better on a more powerful card vs a K620, but that doesn't halve all of the other benchmark results by proxy.

Anyway, none of this should be surprising to anyone who has been involved in this stuff for the past 15 years. What were the competing higher end cards of the time? How about a GTX 970 (Maxwell) vs an R9 290X (GCN 2.0). They were neck and neck most of the time, and probably still are in newer games, with either card having a 10%-15% or so lead in some situations.

Their specs? (taking into account the whole 3.5GB thing for the 970, we're just going to ignore that last 1/8th of the memory lanes since accessing it completely killed performance)

GTX 970: 196GB/s - 65.9GP/s - 122.5 GT/s - 3920 GFLOPS - 145 Watts
R9 290X: 320GB/s - 64GP/s - 176 GT/s - 5632 GFLOPS - 250 Watts

The K620 vs 7770 is just this comparison scaled way down. The two cards are very similar (or identical) in their pixel fill rate, but the GCN 1.0 and 2.0 cards needed somewhere in the realm of 60% higher bandwidth, 50% higher texel fill rate and 50% higher FLOPS to be able to trade blows with their Maxwell equivalents in actual games. Also, there's no point in getting into the semantics of what type of power draw, TDP, etc. each company uses (which is also not standardized), but just going by the advertised numbers the 290X was rated for a whopping 70% more "watts" than the 970. The power numbers aren't surprising considering how much higher all the other specs are.

All this is well documented, and in a discussion about the fastest XP-compatible GPU that can be jammed into a tiny computer and run efficiently, disregarding these low end Maxwell cards because the numbers aren't high enough just seems odd. The K620 and, as you have found, the K1200 are nearly the fastest you can get for this purpose and are by far the best bang for the buck. A low profile 750 Ti (Maxwell) or the unobtainium low profile GTX 950 (Maxwell again) would beat them for a much higher price, but I don't believe anything from AMD was powerful and efficient enough to be a contender until years after they stopped supporting XP.

EDIT: Also, we might as well throw Kepler into the comparison above with the 780 Ti. It generally trades blows with the 290X and 970 (again, other than specific cases where one of the other dominates), but the specs were actually higher than the 290X (aside from pixel fill rate, which was much lower, and wattage which was rated a bit lower). 336GB/s - 42GP/s - 210GT/s - 5045 GFLOPS - 230 Watts

... and Kepler was significantly more efficient (in specs and in power) than Fermi, especially in comparison to the 400 series. And Fermi was significantly more efficient than Tesla, etc.

You just can't use these numbers to determine performance between generations.

Look, someone did the work for us here:

That's another pointless comparison of apples and oranges. Stock 7770 is severely limited in modern games with 1Gb of VRAM and still decimates K620 in scenarios where 1Gb is sufficient (GTA5 low vs high 🤣). Minecraft is an OpenGL game, which was notoriously bad on AMD cards before OpenGL driver rewrite.

I think you're a bit too hung up on specs.

I don't. Nvidia just can't bullshit their way through more than twice the difference in memory bandwidth. That's why K620 struggles to beat even GTS 450, where memory size isn't an issue, i.e. old DirectX 9 games.

The Serpent Rider wrote on 2024-08-07, 06:10:

That's another pointless comparison of apples and oranges. Stock 7770 is severely limited in modern games with 1Gb of VRAM and s […]
Show full quote

Look, someone did the work for us here:

That's another pointless comparison of apples and oranges. Stock 7770 is severely limited in modern games with 1Gb of VRAM and still decimates K620 in scenarios where 1Gb is sufficient (GTA5 low vs high 🤣). Minecraft is an OpenGL game, which was notoriously bad on AMD cards before OpenGL driver rewrite.

I think you're a bit too hung up on specs.

I don't. Nvidia just can't bullshit their way through more than twice the difference in memory bandwidth. That's why K620 struggles to beat even GTS 450, where memory size isn't an issue.

🤷

I am open minded about this stuff and I have no stock in it either way. If you have some benchmarks to show that the K620 is not an excellent bang for the buck card for low-profile XP gaming, then post them so we can add to the very limited data available.

Also, this is for you, it kinda seems like you need it. 🤗

If you have some benchmarks to show that the K620 is not an excellent bang for the buck card for low-profile XP gaming

Eh, depends. R7 250E cards are quite close in price, but I don't have them. Yet. And if you can find K1200 cheap, it's a steal for Windows XP, no questions asked, even if two slot thick low-profile cards are considered.

The Serpent Rider wrote on 2024-08-07, 11:24:

If you have some benchmarks to show that the K620 is not an excellent bang for the buck card for low-profile XP gaming

Eh, depends. R7 250E cards are quite close in price, but I don't have them. Yet. And if you can find K1200 cheap, it's a steal for Windows XP, no questions asked, even if two slot thick low-profile cards are considered.

The R7 250E is identical to the HD 7750 (except for the E?), so I'm sure it will trade blows with the K620 in some games but most likely will be slower overall... just going by what I have seen in benchmarks online. Also, power consumption of the K620 will be much better.

Just doing a quick search on ebay, I can't even find any cards labeled as R7 250E, so I tend to think this is one of those SKUs that TechPowerUp has kind of made up out of thin air because a different die revision was found (Cape Verde Pro-E). Unless there is some specific benefit to the Cape Verde Pro-E core (which I find no mention of outside of TPU), it seems that any low profile Cape Verde Pro card like this would be suitable for this comparison.

But then, that is one of the main issues... finding low profile cards from these lineups is getting tougher and tougher as more people cram them into old workstations to play Minecraft. The K620 and K1200 are always low profile and just need appropriate brackets.

I see another interesting card mentioned on TPU and that is the R7 250XE, which they say is a downclocked Cape Verde XT (7770). The picture shows a low profile card that looks like any random AMD card from the HD6450 on up, so I have my doubts that this has any chance of being a repackaged 80W GDDR5 mid-range gaming card. Again, cannot find this anywhere online. TPU's database is "neat" but there are a lot of inconsistencies and devices listed with basically no trace of them elsewhere.

Palit GeForce GTS 450 low-profile (stock):
-------------------------------------------------------
FEAR: high settings without soft shadows, 1600x1200 4x AA 16x AF

Min 45
Average 88
Max 195

Unigine Heaven 1280x720 Ultra settings, Normal tessellation, 2x AA

Average 24.4
Min 15.3
Max 51.2

Score 615

3DMark 2005: 27869

PNY Quadro K620
-------------------------------------------------------
FEAR: high settings without soft shadows, 1600x1200 4x AA 16x AF

Min 56
Average 96
Max 190

Unigine Heaven 1280x720 Ultra settings, Normal tessellation, 2x AA

Average 33.2
Min 17.0
Max 67.3

Score 837

3DMark 2005: 29595

The Serpent Rider wrote on 2024-08-07, 18:47:

Palit GeForce GTS 450 low-profile (stock): ------------------------------------------------------- FEAR: high settings without s […]
Show full quote

Palit GeForce GTS 450 low-profile (stock):
-------------------------------------------------------
FEAR: high settings without soft shadows, 1600x1200 4x AA 16x AF

Min 45
Average 88
Max 195

Unigine Heaven 1280x720 Ultra settings, Normal tessellation, 2x AA

Average 24.4
Min 15.3
Max 51.2

Score 615

3DMark 2005: 27869

PNY Quadro K620
-------------------------------------------------------
FEAR: high settings without soft shadows, 1600x1200 4x AA 16x AF

Min 56
Average 96
Max 190

Unigine Heaven 1280x720 Ultra settings, Normal tessellation, 2x AA

Average 33.2
Min 17.0
Max 67.3

Score 837

3DMark 2005: 29595

Nice! That's actually very solid for such a low power card. On average 10% faster in FEAR at 1600x1200 with 4xAA and 16xAF isn't bad at all considering the K620 has half the bandwidth of the GTS450. And ~36% faster in Heaven is a good result as well.

What are the rest of the system specs?

Also, not sure if it has been mentioned in the thread yet, but the Firepro W4300 may be another option. Apparently with some driver tinkering they can be made to work in XP as a Radeon 200 series (R7 260). I'm sure the K1200 is faster and more efficient, but this would be yet another option.

Cache is probably the reason that the K620 DGAF that 7750/7770 has faster RAM, 2MB L2 vs 256kB, L1 is kinda closer due to pooling it different, 64K per SMM on K620 but 16k per CU with less shaders per CU on AMD I think, so comes out like about ~24kB equivalent on the K620

I couldn't get the Firepro W2100 to work on XP as I mentioned before, the W4300 and the K1200 look interesting, but their price is not correlative to their performance I believe.
Still, hopefully someone will stop by who has one of the two.
As a note, I have tested with Windows 7 32 bit and the performance of the K620 is superior to XP, at least in my case, with the same tests.
I did not take notes on this because I moved on to other computers.

Ozzuneoj wrote on 2024-08-07, 17:49:

The R7 250E is identical to the HD 7750 (except for the E?), so I'm sure it will trade blows with the K620 in some games but most likely will be slower overall... just going by what I have seen in benchmarks online. Also, power consumption of the K620 will be much better.

Some Radeon HD 7750 benchmarks and tests here:
AMD Catalyst drivers for Windows XP - newer is NOT better
3DMark 2005 for driver 13.1: 28025 Marks (GT1=115,8 fps / GT2=88,1 fps / GT3=138,1 fps)
"iCafe" 2015 Windows XP driver - 3Dmark2005 result was a pretty much the same as with the v13.1 driver

I did my tests on a mildly overclocked X5670, boosting around 3600-3750MHz, so Radeon 7750 will most likely get more than 30k on such system.

I have not read this whole thread.

The smallest computer I have that's appropriate is a small possibly mini-ITX form factor AMD Geode (is that right, can't remember) development box. It's quite cute and was dirt cheap. Not for sale, sorry.

You may want to consider a thin client. Or a single board computer. What is your minimum processor requirement?

(Sorry you don't want a TC.)

I have a laptop from 2019 that gas legacy support in the firmware. I'm going to attempt to boot it with a dos cd. I know it won't accept a Win2K install. Then I can practice assembler as I'm brushing up. Hopefully. O/w I'm using the Geode or some other damned thing around here.

I also have an HP DC7600 SFF that came with XP. It's not tiny, but you can't have everything. Something like that is probably your best bet.

There's a guy on eBay who is selling a dual server mobo cheap. About 35$ total. No io plate or driver's or anything though. It's a Supermicro x7 blah blah. Cpus are cheap (1 or 2). Heatsinks are not necessarily cheap (dual socket 771). Those cpus and ram (which is also cheap) runs mucho calienye hot though. Very hot. When I had a big EATX version of the same board (Intel), I could NOT keep it cool. Had huge passive heatsinks. The problem was the ram though. Leaving the side.open and pointing a desk fan at it worked well though.

The Supermicro board is also 12" x 10". Not totally standard atx ff. Keep that in mind. I also don't know what driver support there is on the net. 2003 maybe. Not Server 2008. So that may be a deal breaker.

Turns out there's a boatload of driver's, even for XP.

https://www.supermicro.com/en/support/resourc … ter/MBD-X7DVL-3

Ordered cheap Firepro W4100. Lets see how it fares up. It's clocked 630 MHz by default though, so I would nee to overclock.

The Serpent Rider wrote on 2024-08-08, 17:16:

Ordered cheap Firepro W4100. Lets see how it fares up. It's clocked 630 MHz by default though, so I would to overclock.

I am eager to know what you get with the XP drivers for the W4100.